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");
1954 /* Return subexpression representing possible alloca call, if any. */
1956 inline_forbidden_p (tree fndecl)
1958 location_t saved_loc = input_location;
1959 block_stmt_iterator bsi;
1961 tree ret = NULL_TREE;
1963 FOR_EACH_BB_FN (bb, DECL_STRUCT_FUNCTION (fndecl))
1964 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
1966 ret = walk_tree_without_duplicates (bsi_stmt_ptr (bsi),
1967 inline_forbidden_p_1, fndecl);
1973 input_location = saved_loc;
1977 /* Returns nonzero if FN is a function that does not have any
1978 fundamental inline blocking properties. */
1981 inlinable_function_p (tree fn)
1983 bool inlinable = true;
1987 /* If we've already decided this function shouldn't be inlined,
1988 there's no need to check again. */
1989 if (DECL_UNINLINABLE (fn))
1992 /* We only warn for functions declared `inline' by the user. */
1993 do_warning = (warn_inline
1995 && DECL_DECLARED_INLINE_P (fn)
1996 && !DECL_IN_SYSTEM_HEADER (fn));
1998 always_inline = lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn));
2000 if (flag_really_no_inline
2001 && always_inline == NULL)
2004 warning (OPT_Winline, "function %q+F can never be inlined because it "
2005 "is suppressed using -fno-inline", fn);
2009 /* Don't auto-inline anything that might not be bound within
2010 this unit of translation. */
2011 else if (!DECL_DECLARED_INLINE_P (fn)
2012 && DECL_REPLACEABLE_P (fn))
2015 else if (!function_attribute_inlinable_p (fn))
2018 warning (OPT_Winline, "function %q+F can never be inlined because it "
2019 "uses attributes conflicting with inlining", fn);
2023 /* If we don't have the function body available, we can't inline it.
2024 However, this should not be recorded since we also get here for
2025 forward declared inline functions. Therefore, return at once. */
2026 if (!DECL_SAVED_TREE (fn))
2029 /* If we're not inlining at all, then we cannot inline this function. */
2030 else if (!flag_inline_trees)
2033 /* Only try to inline functions if DECL_INLINE is set. This should be
2034 true for all functions declared `inline', and for all other functions
2035 as well with -finline-functions.
2037 Don't think of disregarding DECL_INLINE when flag_inline_trees == 2;
2038 it's the front-end that must set DECL_INLINE in this case, because
2039 dwarf2out loses if a function that does not have DECL_INLINE set is
2040 inlined anyway. That is why we have both DECL_INLINE and
2041 DECL_DECLARED_INLINE_P. */
2042 /* FIXME: When flag_inline_trees dies, the check for flag_unit_at_a_time
2043 here should be redundant. */
2044 else if (!DECL_INLINE (fn) && !flag_unit_at_a_time)
2047 else if (inline_forbidden_p (fn))
2049 /* See if we should warn about uninlinable functions. Previously,
2050 some of these warnings would be issued while trying to expand
2051 the function inline, but that would cause multiple warnings
2052 about functions that would for example call alloca. But since
2053 this a property of the function, just one warning is enough.
2054 As a bonus we can now give more details about the reason why a
2055 function is not inlinable. */
2057 sorry (inline_forbidden_reason, fn);
2058 else if (do_warning)
2059 warning (OPT_Winline, inline_forbidden_reason, fn);
2064 /* Squirrel away the result so that we don't have to check again. */
2065 DECL_UNINLINABLE (fn) = !inlinable;
2070 /* Estimate the cost of a memory move. Use machine dependent
2071 word size and take possible memcpy call into account. */
2074 estimate_move_cost (tree type)
2078 size = int_size_in_bytes (type);
2080 if (size < 0 || size > MOVE_MAX_PIECES * MOVE_RATIO)
2081 /* Cost of a memcpy call, 3 arguments and the call. */
2084 return ((size + MOVE_MAX_PIECES - 1) / MOVE_MAX_PIECES);
2087 /* Arguments for estimate_num_insns_1. */
2091 /* Used to return the number of insns. */
2094 /* Weights of various constructs. */
2095 eni_weights *weights;
2098 /* Used by estimate_num_insns. Estimate number of instructions seen
2099 by given statement. */
2102 estimate_num_insns_1 (tree *tp, int *walk_subtrees, void *data)
2104 struct eni_data *d = data;
2108 if (IS_TYPE_OR_DECL_P (x))
2113 /* Assume that constants and references counts nothing. These should
2114 be majorized by amount of operations among them we count later
2115 and are common target of CSE and similar optimizations. */
2116 else if (CONSTANT_CLASS_P (x) || REFERENCE_CLASS_P (x))
2119 switch (TREE_CODE (x))
2121 /* Containers have no cost. */
2128 case ALIGN_INDIRECT_REF:
2129 case MISALIGNED_INDIRECT_REF:
2131 case ARRAY_RANGE_REF:
2133 case EXC_PTR_EXPR: /* ??? */
2134 case FILTER_EXPR: /* ??? */
2137 case WITH_CLEANUP_EXPR:
2140 case VIEW_CONVERT_EXPR:
2145 case CASE_LABEL_EXPR:
2148 case EH_FILTER_EXPR:
2149 case STATEMENT_LIST:
2151 case NON_LVALUE_EXPR:
2154 case TRY_CATCH_EXPR:
2155 case TRY_FINALLY_EXPR:
2162 case WITH_SIZE_EXPR:
2166 case OMP_SECTIONS_SWITCH:
2167 case OMP_ATOMIC_STORE:
2170 /* We don't account constants for now. Assume that the cost is amortized
2171 by operations that do use them. We may re-consider this decision once
2172 we are able to optimize the tree before estimating its size and break
2173 out static initializers. */
2174 case IDENTIFIER_NODE:
2184 /* CHANGE_DYNAMIC_TYPE_EXPR explicitly expands to nothing. */
2185 case CHANGE_DYNAMIC_TYPE_EXPR:
2189 /* Try to estimate the cost of assignments. We have three cases to
2191 1) Simple assignments to registers;
2192 2) Stores to things that must live in memory. This includes
2193 "normal" stores to scalars, but also assignments of large
2194 structures, or constructors of big arrays;
2197 Let us look at the first two cases, assuming we have "a = b + C":
2198 <GIMPLE_MODIFY_STMT <var_decl "a">
2199 <plus_expr <var_decl "b"> <constant C>>
2200 If "a" is a GIMPLE register, the assignment to it is free on almost
2201 any target, because "a" usually ends up in a real register. Hence
2202 the only cost of this expression comes from the PLUS_EXPR, and we
2203 can ignore the GIMPLE_MODIFY_STMT.
2204 If "a" is not a GIMPLE register, the assignment to "a" will most
2205 likely be a real store, so the cost of the GIMPLE_MODIFY_STMT is the cost
2206 of moving something into "a", which we compute using the function
2209 The third case deals with TARGET_EXPRs, for which the semantics are
2210 that a temporary is assigned, unless the TARGET_EXPR itself is being
2211 assigned to something else. In the latter case we do not need the
2213 <GIMPLE_MODIFY_STMT <var_decl "a"> <target_expr>>, the
2214 GIMPLE_MODIFY_STMT is free. */
2216 case GIMPLE_MODIFY_STMT:
2217 /* Is the right and side a TARGET_EXPR? */
2218 if (TREE_CODE (GENERIC_TREE_OPERAND (x, 1)) == TARGET_EXPR)
2220 /* ... fall through ... */
2223 x = GENERIC_TREE_OPERAND (x, 0);
2224 /* Is this an assignments to a register? */
2225 if (is_gimple_reg (x))
2227 /* Otherwise it's a store, so fall through to compute the move cost. */
2230 d->count += estimate_move_cost (TREE_TYPE (x));
2233 /* Assign cost of 1 to usual operations.
2234 ??? We may consider mapping RTL costs to this. */
2239 case POINTER_PLUS_EXPR:
2243 case FIXED_CONVERT_EXPR:
2244 case FIX_TRUNC_EXPR:
2256 case VEC_LSHIFT_EXPR:
2257 case VEC_RSHIFT_EXPR:
2264 case TRUTH_ANDIF_EXPR:
2265 case TRUTH_ORIF_EXPR:
2266 case TRUTH_AND_EXPR:
2268 case TRUTH_XOR_EXPR:
2269 case TRUTH_NOT_EXPR:
2278 case UNORDERED_EXPR:
2289 case PREDECREMENT_EXPR:
2290 case PREINCREMENT_EXPR:
2291 case POSTDECREMENT_EXPR:
2292 case POSTINCREMENT_EXPR:
2296 case REALIGN_LOAD_EXPR:
2298 case REDUC_MAX_EXPR:
2299 case REDUC_MIN_EXPR:
2300 case REDUC_PLUS_EXPR:
2301 case WIDEN_SUM_EXPR:
2303 case VEC_WIDEN_MULT_HI_EXPR:
2304 case VEC_WIDEN_MULT_LO_EXPR:
2305 case VEC_UNPACK_HI_EXPR:
2306 case VEC_UNPACK_LO_EXPR:
2307 case VEC_UNPACK_FLOAT_HI_EXPR:
2308 case VEC_UNPACK_FLOAT_LO_EXPR:
2309 case VEC_PACK_TRUNC_EXPR:
2310 case VEC_PACK_SAT_EXPR:
2311 case VEC_PACK_FIX_TRUNC_EXPR:
2313 case WIDEN_MULT_EXPR:
2315 case VEC_EXTRACT_EVEN_EXPR:
2316 case VEC_EXTRACT_ODD_EXPR:
2317 case VEC_INTERLEAVE_HIGH_EXPR:
2318 case VEC_INTERLEAVE_LOW_EXPR:
2325 /* TODO: Cost of a switch should be derived from the number of
2327 d->count += d->weights->switch_cost;
2330 /* Few special cases of expensive operations. This is useful
2331 to avoid inlining on functions having too many of these. */
2332 case TRUNC_DIV_EXPR:
2334 case FLOOR_DIV_EXPR:
2335 case ROUND_DIV_EXPR:
2336 case EXACT_DIV_EXPR:
2337 case TRUNC_MOD_EXPR:
2339 case FLOOR_MOD_EXPR:
2340 case ROUND_MOD_EXPR:
2342 d->count += d->weights->div_mod_cost;
2346 tree decl = get_callee_fndecl (x);
2348 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_MD)
2349 cost = d->weights->target_builtin_call_cost;
2351 cost = d->weights->call_cost;
2353 if (decl && DECL_BUILT_IN_CLASS (decl) == BUILT_IN_NORMAL)
2354 switch (DECL_FUNCTION_CODE (decl))
2356 case BUILT_IN_CONSTANT_P:
2359 case BUILT_IN_EXPECT:
2361 /* Prefetch instruction is not expensive. */
2362 case BUILT_IN_PREFETCH:
2369 /* Our cost must be kept in sync with cgraph_estimate_size_after_inlining
2370 that does use function declaration to figure out the arguments. */
2374 call_expr_arg_iterator iter;
2375 FOR_EACH_CALL_EXPR_ARG (a, iter, x)
2376 d->count += estimate_move_cost (TREE_TYPE (a));
2381 for (arg = DECL_ARGUMENTS (decl); arg; arg = TREE_CHAIN (arg))
2382 d->count += estimate_move_cost (TREE_TYPE (arg));
2398 case OMP_ATOMIC_LOAD:
2399 /* OpenMP directives are generally very expensive. */
2400 d->count += d->weights->omp_cost;
2409 /* Estimate number of instructions that will be created by expanding EXPR.
2410 WEIGHTS contains weights attributed to various constructs. */
2413 estimate_num_insns (tree expr, eni_weights *weights)
2415 struct pointer_set_t *visited_nodes;
2417 block_stmt_iterator bsi;
2418 struct function *my_function;
2419 struct eni_data data;
2422 data.weights = weights;
2424 /* If we're given an entire function, walk the CFG. */
2425 if (TREE_CODE (expr) == FUNCTION_DECL)
2427 my_function = DECL_STRUCT_FUNCTION (expr);
2428 gcc_assert (my_function && my_function->cfg);
2429 visited_nodes = pointer_set_create ();
2430 FOR_EACH_BB_FN (bb, my_function)
2432 for (bsi = bsi_start (bb);
2436 walk_tree (bsi_stmt_ptr (bsi), estimate_num_insns_1,
2437 &data, visited_nodes);
2440 pointer_set_destroy (visited_nodes);
2443 walk_tree_without_duplicates (&expr, estimate_num_insns_1, &data);
2448 /* Initializes weights used by estimate_num_insns. */
2451 init_inline_once (void)
2453 eni_inlining_weights.call_cost = PARAM_VALUE (PARAM_INLINE_CALL_COST);
2454 eni_inlining_weights.target_builtin_call_cost = 1;
2455 eni_inlining_weights.div_mod_cost = 10;
2456 eni_inlining_weights.switch_cost = 1;
2457 eni_inlining_weights.omp_cost = 40;
2459 eni_size_weights.call_cost = 1;
2460 eni_size_weights.target_builtin_call_cost = 1;
2461 eni_size_weights.div_mod_cost = 1;
2462 eni_size_weights.switch_cost = 10;
2463 eni_size_weights.omp_cost = 40;
2465 /* Estimating time for call is difficult, since we have no idea what the
2466 called function does. In the current uses of eni_time_weights,
2467 underestimating the cost does less harm than overestimating it, so
2468 we choose a rather small value here. */
2469 eni_time_weights.call_cost = 10;
2470 eni_time_weights.target_builtin_call_cost = 10;
2471 eni_time_weights.div_mod_cost = 10;
2472 eni_time_weights.switch_cost = 4;
2473 eni_time_weights.omp_cost = 40;
2476 /* Install new lexical TREE_BLOCK underneath 'current_block'. */
2478 add_lexical_block (tree current_block, tree new_block)
2482 /* Walk to the last sub-block. */
2483 for (blk_p = &BLOCK_SUBBLOCKS (current_block);
2485 blk_p = &BLOCK_CHAIN (*blk_p))
2488 BLOCK_SUPERCONTEXT (new_block) = current_block;
2491 /* If *TP is a CALL_EXPR, replace it with its inline expansion. */
2494 expand_call_inline (basic_block bb, tree stmt, tree *tp, void *data)
2500 struct pointer_map_t *st;
2503 location_t saved_location;
2504 struct cgraph_edge *cg_edge;
2506 basic_block return_block;
2508 block_stmt_iterator bsi, stmt_bsi;
2509 bool successfully_inlined = FALSE;
2510 bool purge_dead_abnormal_edges;
2514 /* See what we've got. */
2515 id = (copy_body_data *) data;
2518 /* Set input_location here so we get the right instantiation context
2519 if we call instantiate_decl from inlinable_function_p. */
2520 saved_location = input_location;
2521 if (EXPR_HAS_LOCATION (t))
2522 input_location = EXPR_LOCATION (t);
2524 /* From here on, we're only interested in CALL_EXPRs. */
2525 if (TREE_CODE (t) != CALL_EXPR)
2528 /* First, see if we can figure out what function is being called.
2529 If we cannot, then there is no hope of inlining the function. */
2530 fn = get_callee_fndecl (t);
2534 /* Turn forward declarations into real ones. */
2535 fn = cgraph_node (fn)->decl;
2537 /* If fn is a declaration of a function in a nested scope that was
2538 globally declared inline, we don't set its DECL_INITIAL.
2539 However, we can't blindly follow DECL_ABSTRACT_ORIGIN because the
2540 C++ front-end uses it for cdtors to refer to their internal
2541 declarations, that are not real functions. Fortunately those
2542 don't have trees to be saved, so we can tell by checking their
2544 if (! DECL_INITIAL (fn)
2545 && DECL_ABSTRACT_ORIGIN (fn)
2546 && DECL_SAVED_TREE (DECL_ABSTRACT_ORIGIN (fn)))
2547 fn = DECL_ABSTRACT_ORIGIN (fn);
2549 /* Objective C and fortran still calls tree_rest_of_compilation directly.
2550 Kill this check once this is fixed. */
2551 if (!id->dst_node->analyzed)
2554 cg_edge = cgraph_edge (id->dst_node, stmt);
2556 /* Constant propagation on argument done during previous inlining
2557 may create new direct call. Produce an edge for it. */
2560 struct cgraph_node *dest = cgraph_node (fn);
2562 /* We have missing edge in the callgraph. This can happen in one case
2563 where previous inlining turned indirect call into direct call by
2564 constant propagating arguments. In all other cases we hit a bug
2565 (incorrect node sharing is most common reason for missing edges. */
2566 gcc_assert (dest->needed || !flag_unit_at_a_time);
2567 cgraph_create_edge (id->dst_node, dest, stmt,
2568 bb->count, CGRAPH_FREQ_BASE,
2569 bb->loop_depth)->inline_failed
2570 = N_("originally indirect function call not considered for inlining");
2573 fprintf (dump_file, "Created new direct edge to %s",
2574 cgraph_node_name (dest));
2579 /* Don't try to inline functions that are not well-suited to
2581 if (!cgraph_inline_p (cg_edge, &reason))
2583 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn))
2584 /* Avoid warnings during early inline pass. */
2585 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2587 sorry ("inlining failed in call to %q+F: %s", fn, reason);
2588 sorry ("called from here");
2590 else if (warn_inline && DECL_DECLARED_INLINE_P (fn)
2591 && !DECL_IN_SYSTEM_HEADER (fn)
2593 && !lookup_attribute ("noinline", DECL_ATTRIBUTES (fn))
2594 /* Avoid warnings during early inline pass. */
2595 && (!flag_unit_at_a_time || cgraph_global_info_ready))
2597 warning (OPT_Winline, "inlining failed in call to %q+F: %s",
2599 warning (OPT_Winline, "called from here");
2603 fn = cg_edge->callee->decl;
2605 #ifdef ENABLE_CHECKING
2606 if (cg_edge->callee->decl != id->dst_node->decl)
2607 verify_cgraph_node (cg_edge->callee);
2610 /* We will be inlining this callee. */
2611 id->eh_region = lookup_stmt_eh_region (stmt);
2613 /* Split the block holding the CALL_EXPR. */
2614 e = split_block (bb, stmt);
2616 return_block = e->dest;
2619 /* split_block splits after the statement; work around this by
2620 moving the call into the second block manually. Not pretty,
2621 but seems easier than doing the CFG manipulation by hand
2622 when the CALL_EXPR is in the last statement of BB. */
2623 stmt_bsi = bsi_last (bb);
2624 bsi_remove (&stmt_bsi, false);
2626 /* If the CALL_EXPR was in the last statement of BB, it may have
2627 been the source of abnormal edges. In this case, schedule
2628 the removal of dead abnormal edges. */
2629 bsi = bsi_start (return_block);
2630 if (bsi_end_p (bsi))
2632 bsi_insert_after (&bsi, stmt, BSI_NEW_STMT);
2633 purge_dead_abnormal_edges = true;
2637 bsi_insert_before (&bsi, stmt, BSI_NEW_STMT);
2638 purge_dead_abnormal_edges = false;
2641 stmt_bsi = bsi_start (return_block);
2643 /* Build a block containing code to initialize the arguments, the
2644 actual inline expansion of the body, and a label for the return
2645 statements within the function to jump to. The type of the
2646 statement expression is the return type of the function call. */
2647 id->block = make_node (BLOCK);
2648 BLOCK_ABSTRACT_ORIGIN (id->block) = fn;
2649 BLOCK_SOURCE_LOCATION (id->block) = input_location;
2650 add_lexical_block (TREE_BLOCK (stmt), id->block);
2652 /* Local declarations will be replaced by their equivalents in this
2655 id->decl_map = pointer_map_create ();
2657 /* Record the function we are about to inline. */
2659 id->src_node = cg_edge->callee;
2660 id->src_cfun = DECL_STRUCT_FUNCTION (fn);
2663 gcc_assert (!id->src_cfun->after_inlining);
2665 initialize_inlined_parameters (id, t, fn, bb);
2667 if (DECL_INITIAL (fn))
2668 add_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn), id));
2670 /* Return statements in the function body will be replaced by jumps
2671 to the RET_LABEL. */
2673 gcc_assert (DECL_INITIAL (fn));
2674 gcc_assert (TREE_CODE (DECL_INITIAL (fn)) == BLOCK);
2676 /* Find the lhs to which the result of this call is assigned. */
2678 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT)
2680 modify_dest = GIMPLE_STMT_OPERAND (stmt, 0);
2682 /* The function which we are inlining might not return a value,
2683 in which case we should issue a warning that the function
2684 does not return a value. In that case the optimizers will
2685 see that the variable to which the value is assigned was not
2686 initialized. We do not want to issue a warning about that
2687 uninitialized variable. */
2688 if (DECL_P (modify_dest))
2689 TREE_NO_WARNING (modify_dest) = 1;
2690 if (CALL_EXPR_RETURN_SLOT_OPT (t))
2692 return_slot = modify_dest;
2699 /* Declare the return variable for the function. */
2700 declare_return_variable (id, return_slot,
2701 modify_dest, &use_retvar);
2703 /* This is it. Duplicate the callee body. Assume callee is
2704 pre-gimplified. Note that we must not alter the caller
2705 function in any way before this point, as this CALL_EXPR may be
2706 a self-referential call; if we're calling ourselves, we need to
2707 duplicate our body before altering anything. */
2708 copy_body (id, bb->count, bb->frequency, bb, return_block);
2710 /* Add local vars in this inlined callee to caller. */
2711 t_step = id->src_cfun->unexpanded_var_list;
2712 for (; t_step; t_step = TREE_CHAIN (t_step))
2714 var = TREE_VALUE (t_step);
2715 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
2716 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
2717 cfun->unexpanded_var_list);
2719 cfun->unexpanded_var_list = tree_cons (NULL_TREE, remap_decl (var, id),
2720 cfun->unexpanded_var_list);
2724 pointer_map_destroy (id->decl_map);
2727 /* If the inlined function returns a result that we care about,
2728 clobber the CALL_EXPR with a reference to the return variable. */
2729 if (use_retvar && (TREE_CODE (bsi_stmt (stmt_bsi)) != CALL_EXPR))
2732 if (gimple_in_ssa_p (cfun))
2735 mark_symbols_for_renaming (stmt);
2737 maybe_clean_or_replace_eh_stmt (stmt, stmt);
2740 /* We're modifying a TSI owned by gimple_expand_calls_inline();
2741 tsi_delink() will leave the iterator in a sane state. */
2743 /* Handle case of inlining function that miss return statement so
2744 return value becomes undefined. */
2745 if (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT
2746 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt, 0)) == SSA_NAME)
2748 tree name = TREE_OPERAND (stmt, 0);
2749 tree var = SSA_NAME_VAR (TREE_OPERAND (stmt, 0));
2750 tree def = gimple_default_def (cfun, var);
2752 /* If the variable is used undefined, make this name undefined via
2756 TREE_OPERAND (stmt, 1) = def;
2759 /* Otherwise make this variable undefined. */
2762 bsi_remove (&stmt_bsi, true);
2763 set_default_def (var, name);
2764 SSA_NAME_DEF_STMT (name) = build_empty_stmt ();
2768 bsi_remove (&stmt_bsi, true);
2771 if (purge_dead_abnormal_edges)
2772 tree_purge_dead_abnormal_call_edges (return_block);
2774 /* If the value of the new expression is ignored, that's OK. We
2775 don't warn about this for CALL_EXPRs, so we shouldn't warn about
2776 the equivalent inlined version either. */
2777 TREE_USED (*tp) = 1;
2779 /* Output the inlining info for this abstract function, since it has been
2780 inlined. If we don't do this now, we can lose the information about the
2781 variables in the function when the blocks get blown away as soon as we
2782 remove the cgraph node. */
2783 (*debug_hooks->outlining_inline_function) (cg_edge->callee->decl);
2785 /* Update callgraph if needed. */
2786 cgraph_remove_node (cg_edge->callee);
2788 id->block = NULL_TREE;
2789 successfully_inlined = TRUE;
2792 input_location = saved_location;
2793 return successfully_inlined;
2796 /* Expand call statements reachable from STMT_P.
2797 We can only have CALL_EXPRs as the "toplevel" tree code or nested
2798 in a GIMPLE_MODIFY_STMT. See tree-gimple.c:get_call_expr_in(). We can
2799 unfortunately not use that function here because we need a pointer
2800 to the CALL_EXPR, not the tree itself. */
2803 gimple_expand_calls_inline (basic_block bb, copy_body_data *id)
2805 block_stmt_iterator bsi;
2807 /* Register specific tree functions. */
2808 tree_register_cfg_hooks ();
2809 for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
2811 tree *expr_p = bsi_stmt_ptr (bsi);
2812 tree stmt = *expr_p;
2814 if (TREE_CODE (*expr_p) == GIMPLE_MODIFY_STMT)
2815 expr_p = &GIMPLE_STMT_OPERAND (*expr_p, 1);
2816 if (TREE_CODE (*expr_p) == WITH_SIZE_EXPR)
2817 expr_p = &TREE_OPERAND (*expr_p, 0);
2818 if (TREE_CODE (*expr_p) == CALL_EXPR)
2819 if (expand_call_inline (bb, stmt, expr_p, id))
2825 /* Walk all basic blocks created after FIRST and try to fold every statement
2826 in the STATEMENTS pointer set. */
2828 fold_marked_statements (int first, struct pointer_set_t *statements)
2830 for (;first < n_basic_blocks;first++)
2831 if (BASIC_BLOCK (first))
2833 block_stmt_iterator bsi;
2834 for (bsi = bsi_start (BASIC_BLOCK (first));
2835 !bsi_end_p (bsi); bsi_next (&bsi))
2836 if (pointer_set_contains (statements, bsi_stmt (bsi)))
2838 tree old_stmt = bsi_stmt (bsi);
2839 if (fold_stmt (bsi_stmt_ptr (bsi)))
2841 update_stmt (bsi_stmt (bsi));
2842 if (maybe_clean_or_replace_eh_stmt (old_stmt, bsi_stmt (bsi)))
2843 tree_purge_dead_eh_edges (BASIC_BLOCK (first));
2849 /* Return true if BB has at least one abnormal outgoing edge. */
2852 has_abnormal_outgoing_edge_p (basic_block bb)
2857 FOR_EACH_EDGE (e, ei, bb->succs)
2858 if (e->flags & EDGE_ABNORMAL)
2864 /* Expand calls to inline functions in the body of FN. */
2867 optimize_inline_calls (tree fn)
2872 int last = n_basic_blocks;
2873 /* There is no point in performing inlining if errors have already
2874 occurred -- and we might crash if we try to inline invalid
2876 if (errorcount || sorrycount)
2880 memset (&id, 0, sizeof (id));
2882 id.src_node = id.dst_node = cgraph_node (fn);
2884 /* Or any functions that aren't finished yet. */
2885 prev_fn = NULL_TREE;
2886 if (current_function_decl)
2888 id.dst_fn = current_function_decl;
2889 prev_fn = current_function_decl;
2892 id.copy_decl = copy_decl_maybe_to_var;
2893 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
2894 id.transform_new_cfg = false;
2895 id.transform_return_to_modify = true;
2896 id.transform_lang_insert_block = false;
2897 id.statements_to_fold = pointer_set_create ();
2899 push_gimplify_context ();
2901 /* We make no attempts to keep dominance info up-to-date. */
2902 free_dominance_info (CDI_DOMINATORS);
2903 free_dominance_info (CDI_POST_DOMINATORS);
2905 /* Reach the trees by walking over the CFG, and note the
2906 enclosing basic-blocks in the call edges. */
2907 /* We walk the blocks going forward, because inlined function bodies
2908 will split id->current_basic_block, and the new blocks will
2909 follow it; we'll trudge through them, processing their CALL_EXPRs
2912 gimple_expand_calls_inline (bb, &id);
2914 pop_gimplify_context (NULL);
2916 #ifdef ENABLE_CHECKING
2918 struct cgraph_edge *e;
2920 verify_cgraph_node (id.dst_node);
2922 /* Double check that we inlined everything we are supposed to inline. */
2923 for (e = id.dst_node->callees; e; e = e->next_callee)
2924 gcc_assert (e->inline_failed);
2928 /* Fold the statements before compacting/renumbering the basic blocks. */
2929 fold_marked_statements (last, id.statements_to_fold);
2930 pointer_set_destroy (id.statements_to_fold);
2932 /* Renumber the (code) basic_blocks consecutively. */
2934 /* Renumber the lexical scoping (non-code) blocks consecutively. */
2937 /* We are not going to maintain the cgraph edges up to date.
2938 Kill it so it won't confuse us. */
2939 cgraph_node_remove_callees (id.dst_node);
2941 fold_cond_expr_cond ();
2942 /* It would be nice to check SSA/CFG/statement consistency here, but it is
2943 not possible yet - the IPA passes might make various functions to not
2944 throw and they don't care to proactively update local EH info. This is
2945 done later in fixup_cfg pass that also execute the verification. */
2946 return (TODO_update_ssa | TODO_cleanup_cfg
2947 | (gimple_in_ssa_p (cfun) ? TODO_remove_unused_locals : 0)
2948 | (profile_status != PROFILE_ABSENT ? TODO_rebuild_frequencies : 0));
2951 /* FN is a function that has a complete body, and CLONE is a function whose
2952 body is to be set to a copy of FN, mapping argument declarations according
2953 to the ARG_MAP splay_tree. */
2956 clone_body (tree clone, tree fn, void *arg_map)
2960 /* Clone the body, as if we were making an inline call. But, remap the
2961 parameters in the callee to the parameters of caller. */
2962 memset (&id, 0, sizeof (id));
2965 id.src_cfun = DECL_STRUCT_FUNCTION (fn);
2966 id.decl_map = (struct pointer_map_t *)arg_map;
2968 id.copy_decl = copy_decl_no_change;
2969 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
2970 id.transform_new_cfg = true;
2971 id.transform_return_to_modify = false;
2972 id.transform_lang_insert_block = true;
2974 /* We're not inside any EH region. */
2977 /* Actually copy the body. */
2978 append_to_statement_list_force (copy_generic_body (&id), &DECL_SAVED_TREE (clone));
2981 /* Passed to walk_tree. Copies the node pointed to, if appropriate. */
2984 copy_tree_r (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
2986 enum tree_code code = TREE_CODE (*tp);
2987 enum tree_code_class cl = TREE_CODE_CLASS (code);
2989 /* We make copies of most nodes. */
2990 if (IS_EXPR_CODE_CLASS (cl)
2991 || IS_GIMPLE_STMT_CODE_CLASS (cl)
2992 || code == TREE_LIST
2994 || code == TYPE_DECL
2995 || code == OMP_CLAUSE)
2997 /* Because the chain gets clobbered when we make a copy, we save it
2999 tree chain = NULL_TREE, new;
3001 if (!GIMPLE_TUPLE_P (*tp))
3002 chain = TREE_CHAIN (*tp);
3004 /* Copy the node. */
3005 new = copy_node (*tp);
3007 /* Propagate mudflap marked-ness. */
3008 if (flag_mudflap && mf_marked_p (*tp))
3013 /* Now, restore the chain, if appropriate. That will cause
3014 walk_tree to walk into the chain as well. */
3015 if (code == PARM_DECL
3016 || code == TREE_LIST
3017 || code == OMP_CLAUSE)
3018 TREE_CHAIN (*tp) = chain;
3020 /* For now, we don't update BLOCKs when we make copies. So, we
3021 have to nullify all BIND_EXPRs. */
3022 if (TREE_CODE (*tp) == BIND_EXPR)
3023 BIND_EXPR_BLOCK (*tp) = NULL_TREE;
3025 else if (code == CONSTRUCTOR)
3027 /* CONSTRUCTOR nodes need special handling because
3028 we need to duplicate the vector of elements. */
3031 new = copy_node (*tp);
3033 /* Propagate mudflap marked-ness. */
3034 if (flag_mudflap && mf_marked_p (*tp))
3037 CONSTRUCTOR_ELTS (new) = VEC_copy (constructor_elt, gc,
3038 CONSTRUCTOR_ELTS (*tp));
3041 else if (TREE_CODE_CLASS (code) == tcc_type)
3043 else if (TREE_CODE_CLASS (code) == tcc_declaration)
3045 else if (TREE_CODE_CLASS (code) == tcc_constant)
3048 gcc_assert (code != STATEMENT_LIST);
3052 /* The SAVE_EXPR pointed to by TP is being copied. If ST contains
3053 information indicating to what new SAVE_EXPR this one should be mapped,
3054 use that one. Otherwise, create a new node and enter it in ST. FN is
3055 the function into which the copy will be placed. */
3058 remap_save_expr (tree *tp, void *st_, int *walk_subtrees)
3060 struct pointer_map_t *st = (struct pointer_map_t *) st_;
3064 /* See if we already encountered this SAVE_EXPR. */
3065 n = (tree *) pointer_map_contains (st, *tp);
3067 /* If we didn't already remap this SAVE_EXPR, do so now. */
3070 t = copy_node (*tp);
3072 /* Remember this SAVE_EXPR. */
3073 *pointer_map_insert (st, *tp) = t;
3074 /* Make sure we don't remap an already-remapped SAVE_EXPR. */
3075 *pointer_map_insert (st, t) = t;
3079 /* We've already walked into this SAVE_EXPR; don't do it again. */
3084 /* Replace this SAVE_EXPR with the copy. */
3088 /* Called via walk_tree. If *TP points to a DECL_STMT for a local label,
3089 copies the declaration and enters it in the splay_tree in DATA (which is
3090 really an `copy_body_data *'). */
3093 mark_local_for_remap_r (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED,
3096 copy_body_data *id = (copy_body_data *) data;
3098 /* Don't walk into types. */
3102 else if (TREE_CODE (*tp) == LABEL_EXPR)
3104 tree decl = TREE_OPERAND (*tp, 0);
3106 /* Copy the decl and remember the copy. */
3107 insert_decl_map (id, decl, id->copy_decl (decl, id));
3113 /* Perform any modifications to EXPR required when it is unsaved. Does
3114 not recurse into EXPR's subtrees. */
3117 unsave_expr_1 (tree expr)
3119 switch (TREE_CODE (expr))
3122 /* Don't mess with a TARGET_EXPR that hasn't been expanded.
3123 It's OK for this to happen if it was part of a subtree that
3124 isn't immediately expanded, such as operand 2 of another
3126 if (TREE_OPERAND (expr, 1))
3129 TREE_OPERAND (expr, 1) = TREE_OPERAND (expr, 3);
3130 TREE_OPERAND (expr, 3) = NULL_TREE;
3138 /* Called via walk_tree when an expression is unsaved. Using the
3139 splay_tree pointed to by ST (which is really a `splay_tree'),
3140 remaps all local declarations to appropriate replacements. */
3143 unsave_r (tree *tp, int *walk_subtrees, void *data)
3145 copy_body_data *id = (copy_body_data *) data;
3146 struct pointer_map_t *st = id->decl_map;
3149 /* Only a local declaration (variable or label). */
3150 if ((TREE_CODE (*tp) == VAR_DECL && !TREE_STATIC (*tp))
3151 || TREE_CODE (*tp) == LABEL_DECL)
3153 /* Lookup the declaration. */
3154 n = (tree *) pointer_map_contains (st, *tp);
3156 /* If it's there, remap it. */
3161 else if (TREE_CODE (*tp) == STATEMENT_LIST)
3162 copy_statement_list (tp);
3163 else if (TREE_CODE (*tp) == BIND_EXPR)
3164 copy_bind_expr (tp, walk_subtrees, id);
3165 else if (TREE_CODE (*tp) == SAVE_EXPR)
3166 remap_save_expr (tp, st, walk_subtrees);
3169 copy_tree_r (tp, walk_subtrees, NULL);
3171 /* Do whatever unsaving is required. */
3172 unsave_expr_1 (*tp);
3175 /* Keep iterating. */
3179 /* Copies everything in EXPR and replaces variables, labels
3180 and SAVE_EXPRs local to EXPR. */
3183 unsave_expr_now (tree expr)
3187 /* There's nothing to do for NULL_TREE. */
3192 memset (&id, 0, sizeof (id));
3193 id.src_fn = current_function_decl;
3194 id.dst_fn = current_function_decl;
3195 id.decl_map = pointer_map_create ();
3197 id.copy_decl = copy_decl_no_change;
3198 id.transform_call_graph_edges = CB_CGE_DUPLICATE;
3199 id.transform_new_cfg = false;
3200 id.transform_return_to_modify = false;
3201 id.transform_lang_insert_block = false;
3203 /* Walk the tree once to find local labels. */
3204 walk_tree_without_duplicates (&expr, mark_local_for_remap_r, &id);
3206 /* Walk the tree again, copying, remapping, and unsaving. */
3207 walk_tree (&expr, unsave_r, &id, NULL);
3210 pointer_map_destroy (id.decl_map);
3215 /* Allow someone to determine if SEARCH is a child of TOP from gdb. */
3218 debug_find_tree_1 (tree *tp, int *walk_subtrees ATTRIBUTE_UNUSED, void *data)
3227 debug_find_tree (tree top, tree search)
3229 return walk_tree_without_duplicates (&top, debug_find_tree_1, search) != 0;
3233 /* Declare the variables created by the inliner. Add all the variables in
3234 VARS to BIND_EXPR. */
3237 declare_inline_vars (tree block, tree vars)
3240 for (t = vars; t; t = TREE_CHAIN (t))
3242 DECL_SEEN_IN_BIND_EXPR_P (t) = 1;
3243 gcc_assert (!TREE_STATIC (t) && !TREE_ASM_WRITTEN (t));
3244 cfun->unexpanded_var_list =
3245 tree_cons (NULL_TREE, t,
3246 cfun->unexpanded_var_list);
3250 BLOCK_VARS (block) = chainon (BLOCK_VARS (block), vars);
3254 /* Copy NODE (which must be a DECL). The DECL originally was in the FROM_FN,
3255 but now it will be in the TO_FN. PARM_TO_VAR means enable PARM_DECL to
3256 VAR_DECL translation. */
3259 copy_decl_for_dup_finish (copy_body_data *id, tree decl, tree copy)
3261 /* Don't generate debug information for the copy if we wouldn't have
3262 generated it for the copy either. */
3263 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (decl);
3264 DECL_IGNORED_P (copy) = DECL_IGNORED_P (decl);
3266 /* Set the DECL_ABSTRACT_ORIGIN so the debugging routines know what
3267 declaration inspired this copy. */
3268 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (decl);
3270 /* The new variable/label has no RTL, yet. */
3271 if (CODE_CONTAINS_STRUCT (TREE_CODE (copy), TS_DECL_WRTL)
3272 && !TREE_STATIC (copy) && !DECL_EXTERNAL (copy))
3273 SET_DECL_RTL (copy, NULL_RTX);
3275 /* These args would always appear unused, if not for this. */
3276 TREE_USED (copy) = 1;
3278 /* Set the context for the new declaration. */
3279 if (!DECL_CONTEXT (decl))
3280 /* Globals stay global. */
3282 else if (DECL_CONTEXT (decl) != id->src_fn)
3283 /* Things that weren't in the scope of the function we're inlining
3284 from aren't in the scope we're inlining to, either. */
3286 else if (TREE_STATIC (decl))
3287 /* Function-scoped static variables should stay in the original
3291 /* Ordinary automatic local variables are now in the scope of the
3293 DECL_CONTEXT (copy) = id->dst_fn;
3299 copy_decl_to_var (tree decl, copy_body_data *id)
3303 gcc_assert (TREE_CODE (decl) == PARM_DECL
3304 || TREE_CODE (decl) == RESULT_DECL);
3306 type = TREE_TYPE (decl);
3308 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3309 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3310 TREE_READONLY (copy) = TREE_READONLY (decl);
3311 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3312 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3313 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3315 return copy_decl_for_dup_finish (id, decl, copy);
3318 /* Like copy_decl_to_var, but create a return slot object instead of a
3319 pointer variable for return by invisible reference. */
3322 copy_result_decl_to_var (tree decl, copy_body_data *id)
3326 gcc_assert (TREE_CODE (decl) == PARM_DECL
3327 || TREE_CODE (decl) == RESULT_DECL);
3329 type = TREE_TYPE (decl);
3330 if (DECL_BY_REFERENCE (decl))
3331 type = TREE_TYPE (type);
3333 copy = build_decl (VAR_DECL, DECL_NAME (decl), type);
3334 TREE_READONLY (copy) = TREE_READONLY (decl);
3335 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (decl);
3336 if (!DECL_BY_REFERENCE (decl))
3338 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (decl);
3339 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (decl);
3340 DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (decl);
3343 return copy_decl_for_dup_finish (id, decl, copy);
3348 copy_decl_no_change (tree decl, copy_body_data *id)
3352 copy = copy_node (decl);
3354 /* The COPY is not abstract; it will be generated in DST_FN. */
3355 DECL_ABSTRACT (copy) = 0;
3356 lang_hooks.dup_lang_specific_decl (copy);
3358 /* TREE_ADDRESSABLE isn't used to indicate that a label's address has
3359 been taken; it's for internal bookkeeping in expand_goto_internal. */
3360 if (TREE_CODE (copy) == LABEL_DECL)
3362 TREE_ADDRESSABLE (copy) = 0;
3363 LABEL_DECL_UID (copy) = -1;
3366 return copy_decl_for_dup_finish (id, decl, copy);
3370 copy_decl_maybe_to_var (tree decl, copy_body_data *id)
3372 if (TREE_CODE (decl) == PARM_DECL || TREE_CODE (decl) == RESULT_DECL)
3373 return copy_decl_to_var (decl, id);
3375 return copy_decl_no_change (decl, id);
3378 /* Return a copy of the function's argument tree. */
3380 copy_arguments_for_versioning (tree orig_parm, copy_body_data * id)
3382 tree *arg_copy, *parg;
3384 arg_copy = &orig_parm;
3385 for (parg = arg_copy; *parg; parg = &TREE_CHAIN (*parg))
3387 tree new = remap_decl (*parg, id);
3388 lang_hooks.dup_lang_specific_decl (new);
3389 TREE_CHAIN (new) = TREE_CHAIN (*parg);
3395 /* Return a copy of the function's static chain. */
3397 copy_static_chain (tree static_chain, copy_body_data * id)
3399 tree *chain_copy, *pvar;
3401 chain_copy = &static_chain;
3402 for (pvar = chain_copy; *pvar; pvar = &TREE_CHAIN (*pvar))
3404 tree new = remap_decl (*pvar, id);
3405 lang_hooks.dup_lang_specific_decl (new);
3406 TREE_CHAIN (new) = TREE_CHAIN (*pvar);
3409 return static_chain;
3412 /* Return true if the function is allowed to be versioned.
3413 This is a guard for the versioning functionality. */
3415 tree_versionable_function_p (tree fndecl)
3417 if (fndecl == NULL_TREE)
3419 /* ??? There are cases where a function is
3420 uninlinable but can be versioned. */
3421 if (!tree_inlinable_function_p (fndecl))
3427 /* Create a copy of a function's tree.
3428 OLD_DECL and NEW_DECL are FUNCTION_DECL tree nodes
3429 of the original function and the new copied function
3430 respectively. In case we want to replace a DECL
3431 tree with another tree while duplicating the function's
3432 body, TREE_MAP represents the mapping between these
3433 trees. If UPDATE_CLONES is set, the call_stmt fields
3434 of edges of clones of the function will be updated. */
3436 tree_function_versioning (tree old_decl, tree new_decl, varray_type tree_map,
3439 struct cgraph_node *old_version_node;
3440 struct cgraph_node *new_version_node;
3444 struct ipa_replace_map *replace_info;
3445 basic_block old_entry_block;
3447 tree old_current_function_decl = current_function_decl;
3449 gcc_assert (TREE_CODE (old_decl) == FUNCTION_DECL
3450 && TREE_CODE (new_decl) == FUNCTION_DECL);
3451 DECL_POSSIBLY_INLINED (old_decl) = 1;
3453 old_version_node = cgraph_node (old_decl);
3454 new_version_node = cgraph_node (new_decl);
3456 DECL_ARTIFICIAL (new_decl) = 1;
3457 DECL_ABSTRACT_ORIGIN (new_decl) = DECL_ORIGIN (old_decl);
3459 /* Prepare the data structures for the tree copy. */
3460 memset (&id, 0, sizeof (id));
3462 /* Generate a new name for the new version. */
3465 DECL_NAME (new_decl) = create_tmp_var_name (NULL);
3466 SET_DECL_ASSEMBLER_NAME (new_decl, DECL_NAME (new_decl));
3467 SET_DECL_RTL (new_decl, NULL_RTX);
3468 id.statements_to_fold = pointer_set_create ();
3471 id.decl_map = pointer_map_create ();
3472 id.src_fn = old_decl;
3473 id.dst_fn = new_decl;
3474 id.src_node = old_version_node;
3475 id.dst_node = new_version_node;
3476 id.src_cfun = DECL_STRUCT_FUNCTION (old_decl);
3478 id.copy_decl = copy_decl_no_change;
3479 id.transform_call_graph_edges
3480 = update_clones ? CB_CGE_MOVE_CLONES : CB_CGE_MOVE;
3481 id.transform_new_cfg = true;
3482 id.transform_return_to_modify = false;
3483 id.transform_lang_insert_block = false;
3485 current_function_decl = new_decl;
3486 old_entry_block = ENTRY_BLOCK_PTR_FOR_FUNCTION
3487 (DECL_STRUCT_FUNCTION (old_decl));
3488 initialize_cfun (new_decl, old_decl,
3489 old_entry_block->count,
3490 old_entry_block->frequency);
3491 push_cfun (DECL_STRUCT_FUNCTION (new_decl));
3493 /* Copy the function's static chain. */
3494 p = DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl;
3496 DECL_STRUCT_FUNCTION (new_decl)->static_chain_decl =
3497 copy_static_chain (DECL_STRUCT_FUNCTION (old_decl)->static_chain_decl,
3499 /* Copy the function's arguments. */
3500 if (DECL_ARGUMENTS (old_decl) != NULL_TREE)
3501 DECL_ARGUMENTS (new_decl) =
3502 copy_arguments_for_versioning (DECL_ARGUMENTS (old_decl), &id);
3504 /* If there's a tree_map, prepare for substitution. */
3506 for (i = 0; i < VARRAY_ACTIVE_SIZE (tree_map); i++)
3508 replace_info = VARRAY_GENERIC_PTR (tree_map, i);
3509 if (replace_info->replace_p)
3510 insert_decl_map (&id, replace_info->old_tree,
3511 replace_info->new_tree);
3514 DECL_INITIAL (new_decl) = remap_blocks (DECL_INITIAL (id.src_fn), &id);
3516 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3517 number_blocks (id.dst_fn);
3519 if (DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list != NULL_TREE)
3520 /* Add local vars. */
3521 for (t_step = DECL_STRUCT_FUNCTION (old_decl)->unexpanded_var_list;
3522 t_step; t_step = TREE_CHAIN (t_step))
3524 tree var = TREE_VALUE (t_step);
3525 if (TREE_STATIC (var) && !TREE_ASM_WRITTEN (var))
3526 cfun->unexpanded_var_list = tree_cons (NULL_TREE, var,
3527 cfun->unexpanded_var_list);
3529 cfun->unexpanded_var_list =
3530 tree_cons (NULL_TREE, remap_decl (var, &id),
3531 cfun->unexpanded_var_list);
3534 /* Copy the Function's body. */
3535 copy_body (&id, old_entry_block->count, old_entry_block->frequency, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR);
3537 if (DECL_RESULT (old_decl) != NULL_TREE)
3539 tree *res_decl = &DECL_RESULT (old_decl);
3540 DECL_RESULT (new_decl) = remap_decl (*res_decl, &id);
3541 lang_hooks.dup_lang_specific_decl (DECL_RESULT (new_decl));
3544 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3545 number_blocks (new_decl);
3548 pointer_map_destroy (id.decl_map);
3551 fold_marked_statements (0, id.statements_to_fold);
3552 pointer_set_destroy (id.statements_to_fold);
3553 fold_cond_expr_cond ();
3555 if (gimple_in_ssa_p (cfun))
3557 free_dominance_info (CDI_DOMINATORS);
3558 free_dominance_info (CDI_POST_DOMINATORS);
3560 delete_unreachable_blocks ();
3561 update_ssa (TODO_update_ssa);
3564 fold_cond_expr_cond ();
3565 if (need_ssa_update_p ())
3566 update_ssa (TODO_update_ssa);
3569 free_dominance_info (CDI_DOMINATORS);
3570 free_dominance_info (CDI_POST_DOMINATORS);
3572 current_function_decl = old_current_function_decl;
3573 gcc_assert (!current_function_decl
3574 || DECL_STRUCT_FUNCTION (current_function_decl) == cfun);
3578 /* Duplicate a type, fields and all. */
3581 build_duplicate_type (tree type)
3583 struct copy_body_data id;
3585 memset (&id, 0, sizeof (id));
3586 id.src_fn = current_function_decl;
3587 id.dst_fn = current_function_decl;
3589 id.decl_map = pointer_map_create ();
3590 id.copy_decl = copy_decl_no_change;
3592 type = remap_type_1 (type, &id);
3594 pointer_map_destroy (id.decl_map);