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
)
282 /* See if we have remapped this type. */
283 node
= (tree
*) pointer_map_contains (id
->decl_map
, type
);
287 /* The type only needs remapping if it's variably modified. */
288 if (! variably_modified_type_p (type
, id
->src_fn
))
290 insert_decl_map (id
, type
, type
);
294 /* We do need a copy. build and register it now. If this is a pointer or
295 reference type, remap the designated type and make a new pointer or
297 if (TREE_CODE (type
) == POINTER_TYPE
)
299 new = build_pointer_type_for_mode (remap_type (TREE_TYPE (type
), id
),
301 TYPE_REF_CAN_ALIAS_ALL (type
));
302 insert_decl_map (id
, type
, new);
305 else if (TREE_CODE (type
) == REFERENCE_TYPE
)
307 new = build_reference_type_for_mode (remap_type (TREE_TYPE (type
), id
),
309 TYPE_REF_CAN_ALIAS_ALL (type
));
310 insert_decl_map (id
, type
, new);
314 new = copy_node (type
);
316 insert_decl_map (id
, type
, new);
318 /* This is a new type, not a copy of an old type. Need to reassociate
319 variants. We can handle everything except the main variant lazily. */
320 t
= TYPE_MAIN_VARIANT (type
);
323 t
= remap_type (t
, id
);
324 TYPE_MAIN_VARIANT (new) = t
;
325 TYPE_NEXT_VARIANT (new) = TYPE_NEXT_VARIANT (t
);
326 TYPE_NEXT_VARIANT (t
) = new;
330 TYPE_MAIN_VARIANT (new) = new;
331 TYPE_NEXT_VARIANT (new) = NULL
;
334 if (TYPE_STUB_DECL (type
))
335 TYPE_STUB_DECL (new) = remap_decl (TYPE_STUB_DECL (type
), id
);
337 /* Lazily create pointer and reference types. */
338 TYPE_POINTER_TO (new) = NULL
;
339 TYPE_REFERENCE_TO (new) = NULL
;
341 switch (TREE_CODE (new))
345 case FIXED_POINT_TYPE
:
348 t
= TYPE_MIN_VALUE (new);
349 if (t
&& TREE_CODE (t
) != INTEGER_CST
)
350 walk_tree (&TYPE_MIN_VALUE (new), copy_body_r
, id
, NULL
);
352 t
= TYPE_MAX_VALUE (new);
353 if (t
&& TREE_CODE (t
) != INTEGER_CST
)
354 walk_tree (&TYPE_MAX_VALUE (new), copy_body_r
, id
, NULL
);
358 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id
);
359 walk_tree (&TYPE_ARG_TYPES (new), copy_body_r
, id
, NULL
);
363 TREE_TYPE (new) = remap_type (TREE_TYPE (new), id
);
364 TYPE_DOMAIN (new) = remap_type (TYPE_DOMAIN (new), id
);
369 case QUAL_UNION_TYPE
:
373 for (f
= TYPE_FIELDS (new); f
; f
= TREE_CHAIN (f
))
375 t
= remap_decl (f
, id
);
376 DECL_CONTEXT (t
) = new;
380 TYPE_FIELDS (new) = nreverse (nf
);
386 /* Shouldn't have been thought variable sized. */
390 walk_tree (&TYPE_SIZE (new), copy_body_r
, id
, NULL
);
391 walk_tree (&TYPE_SIZE_UNIT (new), copy_body_r
, id
, NULL
);
397 remap_type (tree type
, copy_body_data
*id
)
404 /* See if we have remapped this type. */
405 node
= (tree
*) pointer_map_contains (id
->decl_map
, type
);
409 /* The type only needs remapping if it's variably modified. */
410 if (! variably_modified_type_p (type
, id
->src_fn
))
412 insert_decl_map (id
, type
, type
);
416 return remap_type_1 (type
, id
);
420 remap_decls (tree decls
, copy_body_data
*id
)
423 tree new_decls
= NULL_TREE
;
425 /* Remap its variables. */
426 for (old_var
= decls
; old_var
; old_var
= TREE_CHAIN (old_var
))
430 /* We can not chain the local static declarations into the unexpanded_var_list
431 as we can't duplicate them or break one decl rule. Go ahead and link
432 them into unexpanded_var_list. */
433 if (!auto_var_in_fn_p (old_var
, id
->src_fn
)
434 && !DECL_EXTERNAL (old_var
))
436 cfun
->unexpanded_var_list
= tree_cons (NULL_TREE
, old_var
,
437 cfun
->unexpanded_var_list
);
441 /* Remap the variable. */
442 new_var
= remap_decl (old_var
, id
);
444 /* If we didn't remap this variable, so we can't mess with its
445 TREE_CHAIN. If we remapped this variable to the return slot, it's
446 already declared somewhere else, so don't declare it here. */
447 if (!new_var
|| new_var
== id
->retvar
)
451 gcc_assert (DECL_P (new_var
));
452 TREE_CHAIN (new_var
) = new_decls
;
457 return nreverse (new_decls
);
460 /* Copy the BLOCK to contain remapped versions of the variables
461 therein. And hook the new block into the block-tree. */
464 remap_block (tree
*block
, copy_body_data
*id
)
470 /* Make the new block. */
472 new_block
= make_node (BLOCK
);
473 TREE_USED (new_block
) = TREE_USED (old_block
);
474 BLOCK_ABSTRACT_ORIGIN (new_block
) = old_block
;
475 BLOCK_SOURCE_LOCATION (new_block
) = BLOCK_SOURCE_LOCATION (old_block
);
478 /* Remap its variables. */
479 BLOCK_VARS (new_block
) = remap_decls (BLOCK_VARS (old_block
), id
);
483 if (id
->transform_lang_insert_block
)
484 lang_hooks
.decls
.insert_block (new_block
);
486 /* Remember the remapped block. */
487 insert_decl_map (id
, old_block
, new_block
);
490 /* Copy the whole block tree and root it in id->block. */
492 remap_blocks (tree block
, copy_body_data
*id
)
500 remap_block (&new, id
);
501 gcc_assert (new != block
);
502 for (t
= BLOCK_SUBBLOCKS (block
); t
; t
= BLOCK_CHAIN (t
))
503 add_lexical_block (new, remap_blocks (t
, id
));
508 copy_statement_list (tree
*tp
)
510 tree_stmt_iterator oi
, ni
;
513 new = alloc_stmt_list ();
514 ni
= tsi_start (new);
515 oi
= tsi_start (*tp
);
518 for (; !tsi_end_p (oi
); tsi_next (&oi
))
519 tsi_link_after (&ni
, tsi_stmt (oi
), TSI_NEW_STMT
);
523 copy_bind_expr (tree
*tp
, int *walk_subtrees
, copy_body_data
*id
)
525 tree block
= BIND_EXPR_BLOCK (*tp
);
526 /* Copy (and replace) the statement. */
527 copy_tree_r (tp
, walk_subtrees
, NULL
);
530 remap_block (&block
, id
);
531 BIND_EXPR_BLOCK (*tp
) = block
;
534 if (BIND_EXPR_VARS (*tp
))
535 /* This will remap a lot of the same decls again, but this should be
537 BIND_EXPR_VARS (*tp
) = remap_decls (BIND_EXPR_VARS (*tp
), id
);
540 /* Called from copy_body_id via walk_tree. DATA is really an
541 `copy_body_data *'. */
544 copy_body_r (tree
*tp
, int *walk_subtrees
, void *data
)
546 copy_body_data
*id
= (copy_body_data
*) data
;
547 tree fn
= id
->src_fn
;
550 /* Begin by recognizing trees that we'll completely rewrite for the
551 inlining context. Our output for these trees is completely
552 different from out input (e.g. RETURN_EXPR is deleted, and morphs
553 into an edge). Further down, we'll handle trees that get
554 duplicated and/or tweaked. */
556 /* When requested, RETURN_EXPRs should be transformed to just the
557 contained GIMPLE_MODIFY_STMT. The branch semantics of the return will
558 be handled elsewhere by manipulating the CFG rather than a statement. */
559 if (TREE_CODE (*tp
) == RETURN_EXPR
&& id
->transform_return_to_modify
)
561 tree assignment
= TREE_OPERAND (*tp
, 0);
563 /* If we're returning something, just turn that into an
564 assignment into the equivalent of the original RESULT_DECL.
565 If the "assignment" is just the result decl, the result
566 decl has already been set (e.g. a recent "foo (&result_decl,
567 ...)"); just toss the entire RETURN_EXPR. */
568 if (assignment
&& TREE_CODE (assignment
) == GIMPLE_MODIFY_STMT
)
570 /* Replace the RETURN_EXPR with (a copy of) the
571 GIMPLE_MODIFY_STMT hanging underneath. */
572 *tp
= copy_node (assignment
);
574 else /* Else the RETURN_EXPR returns no value. */
577 return (tree
) (void *)1;
580 else if (TREE_CODE (*tp
) == SSA_NAME
)
582 *tp
= remap_ssa_name (*tp
, id
);
587 /* Local variables and labels need to be replaced by equivalent
588 variables. We don't want to copy static variables; there's only
589 one of those, no matter how many times we inline the containing
590 function. Similarly for globals from an outer function. */
591 else if (auto_var_in_fn_p (*tp
, fn
))
595 /* Remap the declaration. */
596 new_decl
= remap_decl (*tp
, id
);
597 gcc_assert (new_decl
);
598 /* Replace this variable with the copy. */
599 STRIP_TYPE_NOPS (new_decl
);
603 else if (TREE_CODE (*tp
) == STATEMENT_LIST
)
604 copy_statement_list (tp
);
605 else if (TREE_CODE (*tp
) == SAVE_EXPR
)
606 remap_save_expr (tp
, id
->decl_map
, walk_subtrees
);
607 else if (TREE_CODE (*tp
) == LABEL_DECL
608 && (! DECL_CONTEXT (*tp
)
609 || decl_function_context (*tp
) == id
->src_fn
))
610 /* These may need to be remapped for EH handling. */
611 *tp
= remap_decl (*tp
, id
);
612 else if (TREE_CODE (*tp
) == BIND_EXPR
)
613 copy_bind_expr (tp
, walk_subtrees
, id
);
614 /* Types may need remapping as well. */
615 else if (TYPE_P (*tp
))
616 *tp
= remap_type (*tp
, id
);
618 /* If this is a constant, we have to copy the node iff the type will be
619 remapped. copy_tree_r will not copy a constant. */
620 else if (CONSTANT_CLASS_P (*tp
))
622 tree new_type
= remap_type (TREE_TYPE (*tp
), id
);
624 if (new_type
== TREE_TYPE (*tp
))
627 else if (TREE_CODE (*tp
) == INTEGER_CST
)
628 *tp
= build_int_cst_wide (new_type
, TREE_INT_CST_LOW (*tp
),
629 TREE_INT_CST_HIGH (*tp
));
632 *tp
= copy_node (*tp
);
633 TREE_TYPE (*tp
) = new_type
;
637 /* Otherwise, just copy the node. Note that copy_tree_r already
638 knows not to copy VAR_DECLs, etc., so this is safe. */
641 /* Here we handle trees that are not completely rewritten.
642 First we detect some inlining-induced bogosities for
644 if (TREE_CODE (*tp
) == GIMPLE_MODIFY_STMT
645 && GIMPLE_STMT_OPERAND (*tp
, 0) == GIMPLE_STMT_OPERAND (*tp
, 1)
646 && (auto_var_in_fn_p (GIMPLE_STMT_OPERAND (*tp
, 0), fn
)))
648 /* Some assignments VAR = VAR; don't generate any rtl code
649 and thus don't count as variable modification. Avoid
650 keeping bogosities like 0 = 0. */
651 tree decl
= GIMPLE_STMT_OPERAND (*tp
, 0), value
;
654 n
= (tree
*) pointer_map_contains (id
->decl_map
, decl
);
658 STRIP_TYPE_NOPS (value
);
659 if (TREE_CONSTANT (value
) || TREE_READONLY_DECL_P (value
))
661 *tp
= build_empty_stmt ();
662 return copy_body_r (tp
, walk_subtrees
, data
);
666 else if (TREE_CODE (*tp
) == INDIRECT_REF
)
668 /* Get rid of *& from inline substitutions that can happen when a
669 pointer argument is an ADDR_EXPR. */
670 tree decl
= TREE_OPERAND (*tp
, 0);
673 n
= (tree
*) pointer_map_contains (id
->decl_map
, decl
);
678 /* If we happen to get an ADDR_EXPR in n->value, strip
679 it manually here as we'll eventually get ADDR_EXPRs
680 which lie about their types pointed to. In this case
681 build_fold_indirect_ref wouldn't strip the INDIRECT_REF,
682 but we absolutely rely on that. As fold_indirect_ref
683 does other useful transformations, try that first, though. */
684 tree type
= TREE_TYPE (TREE_TYPE (*n
));
685 new = unshare_expr (*n
);
687 *tp
= fold_indirect_ref_1 (type
, new);
690 if (TREE_CODE (new) == ADDR_EXPR
)
691 *tp
= TREE_OPERAND (new, 0);
694 *tp
= build1 (INDIRECT_REF
, type
, new);
695 TREE_THIS_VOLATILE (*tp
) = TREE_THIS_VOLATILE (old
);
703 /* Here is the "usual case". Copy this tree node, and then
704 tweak some special cases. */
705 copy_tree_r (tp
, walk_subtrees
, NULL
);
707 /* Global variables we didn't seen yet needs to go into referenced
709 if (gimple_in_ssa_p (cfun
) && TREE_CODE (*tp
) == VAR_DECL
)
710 add_referenced_var (*tp
);
712 /* If EXPR has block defined, map it to newly constructed block.
713 When inlining we want EXPRs without block appear in the block
715 if (EXPR_P (*tp
) || GIMPLE_STMT_P (*tp
))
717 new_block
= id
->block
;
718 if (TREE_BLOCK (*tp
))
721 n
= (tree
*) pointer_map_contains (id
->decl_map
,
726 TREE_BLOCK (*tp
) = new_block
;
729 if (TREE_CODE (*tp
) == RESX_EXPR
&& id
->eh_region_offset
)
730 TREE_OPERAND (*tp
, 0) =
733 id
->eh_region_offset
+ TREE_INT_CST_LOW (TREE_OPERAND (*tp
, 0)));
735 if (!GIMPLE_TUPLE_P (*tp
) && TREE_CODE (*tp
) != OMP_CLAUSE
)
736 TREE_TYPE (*tp
) = remap_type (TREE_TYPE (*tp
), id
);
738 /* The copied TARGET_EXPR has never been expanded, even if the
739 original node was expanded already. */
740 if (TREE_CODE (*tp
) == TARGET_EXPR
&& TREE_OPERAND (*tp
, 3))
742 TREE_OPERAND (*tp
, 1) = TREE_OPERAND (*tp
, 3);
743 TREE_OPERAND (*tp
, 3) = NULL_TREE
;
746 /* Variable substitution need not be simple. In particular, the
747 INDIRECT_REF substitution above. Make sure that TREE_CONSTANT
748 and friends are up-to-date. */
749 else if (TREE_CODE (*tp
) == ADDR_EXPR
)
751 walk_tree (&TREE_OPERAND (*tp
, 0), copy_body_r
, id
, NULL
);
752 /* Handle the case where we substituted an INDIRECT_REF
753 into the operand of the ADDR_EXPR. */
754 if (TREE_CODE (TREE_OPERAND (*tp
, 0)) == INDIRECT_REF
)
755 *tp
= TREE_OPERAND (TREE_OPERAND (*tp
, 0), 0);
757 recompute_tree_invariant_for_addr_expr (*tp
);
762 /* Keep iterating. */
766 /* Copy basic block, scale profile accordingly. Edges will be taken care of
770 copy_bb (copy_body_data
*id
, basic_block bb
, int frequency_scale
, int count_scale
)
772 block_stmt_iterator bsi
, copy_bsi
;
773 basic_block copy_basic_block
;
775 /* create_basic_block() will append every new block to
776 basic_block_info automatically. */
777 copy_basic_block
= create_basic_block (NULL
, (void *) 0,
778 (basic_block
) bb
->prev_bb
->aux
);
779 copy_basic_block
->count
= bb
->count
* count_scale
/ REG_BR_PROB_BASE
;
781 /* We are going to rebuild frequencies from scratch. These values have just
782 small importance to drive canonicalize_loop_headers. */
783 copy_basic_block
->frequency
= ((gcov_type
)bb
->frequency
784 * frequency_scale
/ REG_BR_PROB_BASE
);
785 if (copy_basic_block
->frequency
> BB_FREQ_MAX
)
786 copy_basic_block
->frequency
= BB_FREQ_MAX
;
787 copy_bsi
= bsi_start (copy_basic_block
);
789 for (bsi
= bsi_start (bb
);
790 !bsi_end_p (bsi
); bsi_next (&bsi
))
792 tree stmt
= bsi_stmt (bsi
);
793 tree orig_stmt
= stmt
;
795 walk_tree (&stmt
, copy_body_r
, id
, NULL
);
797 /* RETURN_EXPR might be removed,
798 this is signalled by making stmt pointer NULL. */
803 gimple_duplicate_stmt_histograms (cfun
, stmt
, id
->src_cfun
, orig_stmt
);
805 /* With return slot optimization we can end up with
806 non-gimple (foo *)&this->m, fix that here. */
807 if (TREE_CODE (stmt
) == GIMPLE_MODIFY_STMT
808 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 1)) == NOP_EXPR
809 && !is_gimple_val (TREE_OPERAND (GIMPLE_STMT_OPERAND (stmt
, 1), 0)))
810 gimplify_stmt (&stmt
);
812 bsi_insert_after (©_bsi
, stmt
, BSI_NEW_STMT
);
814 /* Process new statement. gimplify_stmt possibly turned statement
815 into multiple statements, we need to process all of them. */
816 while (!bsi_end_p (copy_bsi
))
818 tree
*stmtp
= bsi_stmt_ptr (copy_bsi
);
820 call
= get_call_expr_in (stmt
);
822 if (call
&& CALL_EXPR_VA_ARG_PACK (call
) && id
->call_expr
)
824 /* __builtin_va_arg_pack () should be replaced by
825 all arguments corresponding to ... in the caller. */
826 tree p
, *argarray
, new_call
, *call_ptr
;
827 int nargs
= call_expr_nargs (id
->call_expr
);
829 for (p
= DECL_ARGUMENTS (id
->src_fn
); p
; p
= TREE_CHAIN (p
))
832 argarray
= (tree
*) alloca ((nargs
+ call_expr_nargs (call
))
835 memcpy (argarray
, CALL_EXPR_ARGP (call
),
836 call_expr_nargs (call
) * sizeof (*argarray
));
837 memcpy (argarray
+ call_expr_nargs (call
),
838 CALL_EXPR_ARGP (id
->call_expr
)
839 + (call_expr_nargs (id
->call_expr
) - nargs
),
840 nargs
* sizeof (*argarray
));
842 new_call
= build_call_array (TREE_TYPE (call
),
844 nargs
+ call_expr_nargs (call
),
846 /* Copy all CALL_EXPR flags, locus and block, except
847 CALL_EXPR_VA_ARG_PACK flag. */
848 CALL_EXPR_STATIC_CHAIN (new_call
)
849 = CALL_EXPR_STATIC_CHAIN (call
);
850 CALL_EXPR_TAILCALL (new_call
) = CALL_EXPR_TAILCALL (call
);
851 CALL_EXPR_RETURN_SLOT_OPT (new_call
)
852 = CALL_EXPR_RETURN_SLOT_OPT (call
);
853 CALL_FROM_THUNK_P (new_call
) = CALL_FROM_THUNK_P (call
);
854 CALL_CANNOT_INLINE_P (new_call
)
855 = CALL_CANNOT_INLINE_P (call
);
856 TREE_NOTHROW (new_call
) = TREE_NOTHROW (call
);
857 SET_EXPR_LOCUS (new_call
, EXPR_LOCUS (call
));
858 TREE_BLOCK (new_call
) = TREE_BLOCK (call
);
861 if (TREE_CODE (*call_ptr
) == GIMPLE_MODIFY_STMT
)
862 call_ptr
= &GIMPLE_STMT_OPERAND (*call_ptr
, 1);
863 if (TREE_CODE (*call_ptr
) == WITH_SIZE_EXPR
)
864 call_ptr
= &TREE_OPERAND (*call_ptr
, 0);
865 gcc_assert (*call_ptr
== call
);
866 *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
);
1060 gcc_assert ((re
->flags
& (EDGE_EH
| EDGE_ABNORMAL
))
1061 == (e
->flags
& (EDGE_EH
| EDGE_ABNORMAL
)));
1063 SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi
, e
),
1064 USE_FROM_PTR (PHI_ARG_DEF_PTR_FROM_EDGE (phi
, re
)));
1069 /* Copy edges from BB into its copy constructed earlier, scale profile
1070 accordingly. Edges will be taken care of later. Assume aux
1071 pointers to point to the copies of each BB. */
1073 copy_edges_for_bb (basic_block bb
, int count_scale
, basic_block ret_bb
)
1075 basic_block new_bb
= (basic_block
) bb
->aux
;
1078 block_stmt_iterator bsi
;
1081 /* Use the indices from the original blocks to create edges for the
1083 FOR_EACH_EDGE (old_edge
, ei
, bb
->succs
)
1084 if (!(old_edge
->flags
& EDGE_EH
))
1088 flags
= old_edge
->flags
;
1090 /* Return edges do get a FALLTHRU flag when the get inlined. */
1091 if (old_edge
->dest
->index
== EXIT_BLOCK
&& !old_edge
->flags
1092 && old_edge
->dest
->aux
!= EXIT_BLOCK_PTR
)
1093 flags
|= EDGE_FALLTHRU
;
1094 new = make_edge (new_bb
, (basic_block
) old_edge
->dest
->aux
, flags
);
1095 new->count
= old_edge
->count
* count_scale
/ REG_BR_PROB_BASE
;
1096 new->probability
= old_edge
->probability
;
1099 if (bb
->index
== ENTRY_BLOCK
|| bb
->index
== EXIT_BLOCK
)
1102 for (bsi
= bsi_start (new_bb
); !bsi_end_p (bsi
);)
1105 bool can_throw
, nonlocal_goto
;
1107 copy_stmt
= bsi_stmt (bsi
);
1108 update_stmt (copy_stmt
);
1109 if (gimple_in_ssa_p (cfun
))
1110 mark_symbols_for_renaming (copy_stmt
);
1111 /* Do this before the possible split_block. */
1114 /* If this tree could throw an exception, there are two
1115 cases where we need to add abnormal edge(s): the
1116 tree wasn't in a region and there is a "current
1117 region" in the caller; or the original tree had
1118 EH edges. In both cases split the block after the tree,
1119 and add abnormal edge(s) as needed; we need both
1120 those from the callee and the caller.
1121 We check whether the copy can throw, because the const
1122 propagation can change an INDIRECT_REF which throws
1123 into a COMPONENT_REF which doesn't. If the copy
1124 can throw, the original could also throw. */
1126 can_throw
= tree_can_throw_internal (copy_stmt
);
1127 nonlocal_goto
= tree_can_make_abnormal_goto (copy_stmt
);
1129 if (can_throw
|| nonlocal_goto
)
1131 if (!bsi_end_p (bsi
))
1132 /* Note that bb's predecessor edges aren't necessarily
1133 right at this point; split_block doesn't care. */
1135 edge e
= split_block (new_bb
, copy_stmt
);
1138 new_bb
->aux
= e
->src
->aux
;
1139 bsi
= bsi_start (new_bb
);
1144 make_eh_edges (copy_stmt
);
1147 make_abnormal_goto_edges (bb_for_stmt (copy_stmt
), true);
1149 if ((can_throw
|| nonlocal_goto
)
1150 && gimple_in_ssa_p (cfun
))
1151 update_ssa_across_abnormal_edges (bb_for_stmt (copy_stmt
), ret_bb
,
1152 can_throw
, nonlocal_goto
);
1156 /* Copy the PHIs. All blocks and edges are copied, some blocks
1157 was possibly split and new outgoing EH edges inserted.
1158 BB points to the block of original function and AUX pointers links
1159 the original and newly copied blocks. */
1162 copy_phis_for_bb (basic_block bb
, copy_body_data
*id
)
1164 basic_block new_bb
= bb
->aux
;
1168 for (phi
= phi_nodes (bb
); phi
; phi
= PHI_CHAIN (phi
))
1170 tree res
= PHI_RESULT (phi
);
1175 if (is_gimple_reg (res
))
1177 walk_tree (&new_res
, copy_body_r
, id
, NULL
);
1178 SSA_NAME_DEF_STMT (new_res
)
1179 = new_phi
= create_phi_node (new_res
, new_bb
);
1180 FOR_EACH_EDGE (new_edge
, ei
, new_bb
->preds
)
1182 edge old_edge
= find_edge (new_edge
->src
->aux
, bb
);
1183 tree arg
= PHI_ARG_DEF_FROM_EDGE (phi
, old_edge
);
1186 walk_tree (&new_arg
, copy_body_r
, id
, NULL
);
1187 gcc_assert (new_arg
);
1188 add_phi_arg (new_phi
, new_arg
, new_edge
);
1194 /* Wrapper for remap_decl so it can be used as a callback. */
1196 remap_decl_1 (tree decl
, void *data
)
1198 return remap_decl (decl
, (copy_body_data
*) data
);
1201 /* Build struct function and associated datastructures for the new clone
1202 NEW_FNDECL to be build. CALLEE_FNDECL is the original */
1205 initialize_cfun (tree new_fndecl
, tree callee_fndecl
, gcov_type count
,
1208 struct function
*new_cfun
1209 = (struct function
*) ggc_alloc_cleared (sizeof (struct function
));
1210 struct function
*src_cfun
= DECL_STRUCT_FUNCTION (callee_fndecl
);
1211 int count_scale
, frequency_scale
;
1213 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->count
)
1214 count_scale
= (REG_BR_PROB_BASE
* count
1215 / ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->count
);
1219 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->frequency
)
1220 frequency_scale
= (REG_BR_PROB_BASE
* frequency
1222 ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->frequency
);
1224 frequency_scale
= count_scale
;
1226 /* Register specific tree functions. */
1227 tree_register_cfg_hooks ();
1228 *new_cfun
= *DECL_STRUCT_FUNCTION (callee_fndecl
);
1229 new_cfun
->funcdef_no
= get_next_funcdef_no ();
1230 VALUE_HISTOGRAMS (new_cfun
) = NULL
;
1231 new_cfun
->unexpanded_var_list
= NULL
;
1232 new_cfun
->cfg
= NULL
;
1233 new_cfun
->decl
= new_fndecl
/*= copy_node (callee_fndecl)*/;
1234 DECL_STRUCT_FUNCTION (new_fndecl
) = new_cfun
;
1235 push_cfun (new_cfun
);
1236 init_empty_tree_cfg ();
1238 ENTRY_BLOCK_PTR
->count
=
1239 (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->count
* count_scale
/
1241 ENTRY_BLOCK_PTR
->frequency
=
1242 (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->frequency
*
1243 frequency_scale
/ REG_BR_PROB_BASE
);
1244 EXIT_BLOCK_PTR
->count
=
1245 (EXIT_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->count
* count_scale
/
1247 EXIT_BLOCK_PTR
->frequency
=
1248 (EXIT_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->frequency
*
1249 frequency_scale
/ REG_BR_PROB_BASE
);
1251 init_eh_for_function ();
1253 if (src_cfun
->gimple_df
)
1256 cfun
->gimple_df
->in_ssa_p
= true;
1257 init_ssa_operands ();
1262 /* Make a copy of the body of FN so that it can be inserted inline in
1263 another function. Walks FN via CFG, returns new fndecl. */
1266 copy_cfg_body (copy_body_data
* id
, gcov_type count
, int frequency
,
1267 basic_block entry_block_map
, basic_block exit_block_map
)
1269 tree callee_fndecl
= id
->src_fn
;
1270 /* Original cfun for the callee, doesn't change. */
1271 struct function
*src_cfun
= DECL_STRUCT_FUNCTION (callee_fndecl
);
1272 struct function
*cfun_to_copy
;
1274 tree new_fndecl
= NULL
;
1275 int count_scale
, frequency_scale
;
1278 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->count
)
1279 count_scale
= (REG_BR_PROB_BASE
* count
1280 / ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->count
);
1284 if (ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->frequency
)
1285 frequency_scale
= (REG_BR_PROB_BASE
* frequency
1287 ENTRY_BLOCK_PTR_FOR_FUNCTION (src_cfun
)->frequency
);
1289 frequency_scale
= count_scale
;
1291 /* Register specific tree functions. */
1292 tree_register_cfg_hooks ();
1294 /* Must have a CFG here at this point. */
1295 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION
1296 (DECL_STRUCT_FUNCTION (callee_fndecl
)));
1298 cfun_to_copy
= id
->src_cfun
= DECL_STRUCT_FUNCTION (callee_fndecl
);
1301 ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy
)->aux
= entry_block_map
;
1302 EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy
)->aux
= exit_block_map
;
1303 entry_block_map
->aux
= ENTRY_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy
);
1304 exit_block_map
->aux
= EXIT_BLOCK_PTR_FOR_FUNCTION (cfun_to_copy
);
1306 /* Duplicate any exception-handling regions. */
1309 id
->eh_region_offset
1310 = duplicate_eh_regions (cfun_to_copy
, remap_decl_1
, id
,
1313 /* Use aux pointers to map the original blocks to copy. */
1314 FOR_EACH_BB_FN (bb
, cfun_to_copy
)
1316 basic_block
new = copy_bb (id
, bb
, frequency_scale
, count_scale
);
1321 last
= last_basic_block
;
1322 /* Now that we've duplicated the blocks, duplicate their edges. */
1323 FOR_ALL_BB_FN (bb
, cfun_to_copy
)
1324 copy_edges_for_bb (bb
, count_scale
, exit_block_map
);
1325 if (gimple_in_ssa_p (cfun
))
1326 FOR_ALL_BB_FN (bb
, cfun_to_copy
)
1327 copy_phis_for_bb (bb
, id
);
1328 FOR_ALL_BB_FN (bb
, cfun_to_copy
)
1330 ((basic_block
)bb
->aux
)->aux
= NULL
;
1333 /* Zero out AUX fields of newly created block during EH edge
1335 for (; last
< last_basic_block
; last
++)
1336 BASIC_BLOCK (last
)->aux
= NULL
;
1337 entry_block_map
->aux
= NULL
;
1338 exit_block_map
->aux
= NULL
;
1343 /* Make a copy of the body of FN so that it can be inserted inline in
1344 another function. */
1347 copy_generic_body (copy_body_data
*id
)
1350 tree fndecl
= id
->src_fn
;
1352 body
= DECL_SAVED_TREE (fndecl
);
1353 walk_tree (&body
, copy_body_r
, id
, NULL
);
1359 copy_body (copy_body_data
*id
, gcov_type count
, int frequency
,
1360 basic_block entry_block_map
, basic_block exit_block_map
)
1362 tree fndecl
= id
->src_fn
;
1365 /* If this body has a CFG, walk CFG and copy. */
1366 gcc_assert (ENTRY_BLOCK_PTR_FOR_FUNCTION (DECL_STRUCT_FUNCTION (fndecl
)));
1367 body
= copy_cfg_body (id
, count
, frequency
, entry_block_map
, exit_block_map
);
1372 /* Return true if VALUE is an ADDR_EXPR of an automatic variable
1373 defined in function FN, or of a data member thereof. */
1376 self_inlining_addr_expr (tree value
, tree fn
)
1380 if (TREE_CODE (value
) != ADDR_EXPR
)
1383 var
= get_base_address (TREE_OPERAND (value
, 0));
1385 return var
&& auto_var_in_fn_p (var
, fn
);
1389 setup_one_parameter (copy_body_data
*id
, tree p
, tree value
, tree fn
,
1390 basic_block bb
, tree
*vars
)
1396 tree def
= (gimple_in_ssa_p (cfun
)
1397 ? gimple_default_def (id
->src_cfun
, p
) : NULL
);
1400 && value
!= error_mark_node
1401 && !useless_type_conversion_p (TREE_TYPE (p
), TREE_TYPE (value
)))
1402 rhs
= fold_build1 (NOP_EXPR
, TREE_TYPE (p
), value
);
1404 /* If the parameter is never assigned to, has no SSA_NAMEs created,
1405 we may not need to create a new variable here at all. Instead, we may
1406 be able to just use the argument value. */
1407 if (TREE_READONLY (p
)
1408 && !TREE_ADDRESSABLE (p
)
1409 && value
&& !TREE_SIDE_EFFECTS (value
)
1412 /* We may produce non-gimple trees by adding NOPs or introduce
1413 invalid sharing when operand is not really constant.
1414 It is not big deal to prohibit constant propagation here as
1415 we will constant propagate in DOM1 pass anyway. */
1416 if (is_gimple_min_invariant (value
)
1417 && useless_type_conversion_p (TREE_TYPE (p
),
1419 /* We have to be very careful about ADDR_EXPR. Make sure
1420 the base variable isn't a local variable of the inlined
1421 function, e.g., when doing recursive inlining, direct or
1422 mutually-recursive or whatever, which is why we don't
1423 just test whether fn == current_function_decl. */
1424 && ! self_inlining_addr_expr (value
, fn
))
1426 insert_decl_map (id
, p
, value
);
1431 /* Make an equivalent VAR_DECL. Note that we must NOT remap the type
1432 here since the type of this decl must be visible to the calling
1434 var
= copy_decl_to_var (p
, id
);
1435 if (gimple_in_ssa_p (cfun
) && TREE_CODE (var
) == VAR_DECL
)
1438 add_referenced_var (var
);
1441 /* See if the frontend wants to pass this by invisible reference. If
1442 so, our new VAR_DECL will have REFERENCE_TYPE, and we need to
1443 replace uses of the PARM_DECL with dereferences. */
1444 if (TREE_TYPE (var
) != TREE_TYPE (p
)
1445 && POINTER_TYPE_P (TREE_TYPE (var
))
1446 && TREE_TYPE (TREE_TYPE (var
)) == TREE_TYPE (p
))
1448 insert_decl_map (id
, var
, var
);
1449 var_sub
= build_fold_indirect_ref (var
);
1454 /* Register the VAR_DECL as the equivalent for the PARM_DECL;
1455 that way, when the PARM_DECL is encountered, it will be
1456 automatically replaced by the VAR_DECL. */
1457 insert_decl_map (id
, p
, var_sub
);
1459 /* Declare this new variable. */
1460 TREE_CHAIN (var
) = *vars
;
1463 /* Make gimplifier happy about this variable. */
1464 DECL_SEEN_IN_BIND_EXPR_P (var
) = 1;
1466 /* Even if P was TREE_READONLY, the new VAR should not be.
1467 In the original code, we would have constructed a
1468 temporary, and then the function body would have never
1469 changed the value of P. However, now, we will be
1470 constructing VAR directly. The constructor body may
1471 change its value multiple times as it is being
1472 constructed. Therefore, it must not be TREE_READONLY;
1473 the back-end assumes that TREE_READONLY variable is
1474 assigned to only once. */
1475 if (TYPE_NEEDS_CONSTRUCTING (TREE_TYPE (p
)))
1476 TREE_READONLY (var
) = 0;
1478 /* If there is no setup required and we are in SSA, take the easy route
1479 replacing all SSA names representing the function parameter by the
1480 SSA name passed to function.
1482 We need to construct map for the variable anyway as it might be used
1483 in different SSA names when parameter is set in function.
1485 FIXME: This usually kills the last connection in between inlined
1486 function parameter and the actual value in debug info. Can we do
1487 better here? If we just inserted the statement, copy propagation
1488 would kill it anyway as it always did in older versions of GCC.
1490 We might want to introduce a notion that single SSA_NAME might
1491 represent multiple variables for purposes of debugging. */
1492 if (gimple_in_ssa_p (cfun
) && rhs
&& def
&& is_gimple_reg (p
)
1493 && (TREE_CODE (rhs
) == SSA_NAME
1494 || is_gimple_min_invariant (rhs
))
1495 && !SSA_NAME_OCCURS_IN_ABNORMAL_PHI (def
))
1497 insert_decl_map (id
, def
, rhs
);
1501 /* Initialize this VAR_DECL from the equivalent argument. Convert
1502 the argument to the proper type in case it was promoted. */
1505 block_stmt_iterator bsi
= bsi_last (bb
);
1507 if (rhs
== error_mark_node
)
1509 insert_decl_map (id
, p
, var_sub
);
1513 STRIP_USELESS_TYPE_CONVERSION (rhs
);
1515 /* We want to use GIMPLE_MODIFY_STMT, not INIT_EXPR here so that we
1516 keep our trees in gimple form. */
1517 if (def
&& gimple_in_ssa_p (cfun
) && is_gimple_reg (p
))
1519 def
= remap_ssa_name (def
, id
);
1520 init_stmt
= build_gimple_modify_stmt (def
, rhs
);
1521 SSA_NAME_DEF_STMT (def
) = init_stmt
;
1522 SSA_NAME_IS_DEFAULT_DEF (def
) = 0;
1523 set_default_def (var
, NULL
);
1526 init_stmt
= build_gimple_modify_stmt (var
, rhs
);
1528 /* If we did not create a gimple value and we did not create a gimple
1529 cast of a gimple value, then we will need to gimplify INIT_STMTS
1530 at the end. Note that is_gimple_cast only checks the outer
1531 tree code, not its operand. Thus the explicit check that its
1532 operand is a gimple value. */
1533 if ((!is_gimple_val (rhs
)
1534 && (!is_gimple_cast (rhs
)
1535 || !is_gimple_val (TREE_OPERAND (rhs
, 0))))
1536 || !is_gimple_reg (var
))
1538 tree_stmt_iterator i
;
1540 push_gimplify_context ();
1541 gimplify_stmt (&init_stmt
);
1542 if (gimple_in_ssa_p (cfun
)
1543 && init_stmt
&& TREE_CODE (init_stmt
) == STATEMENT_LIST
)
1545 /* The replacement can expose previously unreferenced
1547 for (i
= tsi_start (init_stmt
); !tsi_end_p (i
); tsi_next (&i
))
1548 find_new_referenced_vars (tsi_stmt_ptr (i
));
1550 pop_gimplify_context (NULL
);
1553 /* If VAR represents a zero-sized variable, it's possible that the
1554 assignment statment may result in no gimple statements. */
1556 bsi_insert_after (&bsi
, init_stmt
, BSI_NEW_STMT
);
1557 if (gimple_in_ssa_p (cfun
))
1558 for (;!bsi_end_p (bsi
); bsi_next (&bsi
))
1559 mark_symbols_for_renaming (bsi_stmt (bsi
));
1563 /* Generate code to initialize the parameters of the function at the
1564 top of the stack in ID from the CALL_EXPR EXP. */
1567 initialize_inlined_parameters (copy_body_data
*id
, tree exp
,
1568 tree fn
, basic_block bb
)
1573 tree vars
= NULL_TREE
;
1574 call_expr_arg_iterator iter
;
1575 tree static_chain
= CALL_EXPR_STATIC_CHAIN (exp
);
1577 /* Figure out what the parameters are. */
1578 parms
= DECL_ARGUMENTS (fn
);
1580 /* Loop through the parameter declarations, replacing each with an
1581 equivalent VAR_DECL, appropriately initialized. */
1582 for (p
= parms
, a
= first_call_expr_arg (exp
, &iter
); p
;
1583 a
= next_call_expr_arg (&iter
), p
= TREE_CHAIN (p
))
1584 setup_one_parameter (id
, p
, a
, fn
, bb
, &vars
);
1586 /* Initialize the static chain. */
1587 p
= DECL_STRUCT_FUNCTION (fn
)->static_chain_decl
;
1588 gcc_assert (fn
!= current_function_decl
);
1591 /* No static chain? Seems like a bug in tree-nested.c. */
1592 gcc_assert (static_chain
);
1594 setup_one_parameter (id
, p
, static_chain
, fn
, bb
, &vars
);
1597 declare_inline_vars (id
->block
, vars
);
1600 /* Declare a return variable to replace the RESULT_DECL for the
1601 function we are calling. An appropriate DECL_STMT is returned.
1602 The USE_STMT is filled to contain a use of the declaration to
1603 indicate the return value of the function.
1605 RETURN_SLOT, if non-null is place where to store the result. It
1606 is set only for CALL_EXPR_RETURN_SLOT_OPT. MODIFY_DEST, if non-null,
1607 was the LHS of the GIMPLE_MODIFY_STMT to which this call is the RHS.
1609 The return value is a (possibly null) value that is the result of the
1610 function as seen by the callee. *USE_P is a (possibly null) value that
1611 holds the result as seen by the caller. */
1614 declare_return_variable (copy_body_data
*id
, tree return_slot
, tree modify_dest
,
1617 tree callee
= id
->src_fn
;
1618 tree caller
= id
->dst_fn
;
1619 tree result
= DECL_RESULT (callee
);
1620 tree callee_type
= TREE_TYPE (result
);
1621 tree caller_type
= TREE_TYPE (TREE_TYPE (callee
));
1624 /* We don't need to do anything for functions that don't return
1626 if (!result
|| VOID_TYPE_P (callee_type
))
1632 /* If there was a return slot, then the return value is the
1633 dereferenced address of that object. */
1636 /* The front end shouldn't have used both return_slot and
1637 a modify expression. */
1638 gcc_assert (!modify_dest
);
1639 if (DECL_BY_REFERENCE (result
))
1641 tree return_slot_addr
= build_fold_addr_expr (return_slot
);
1642 STRIP_USELESS_TYPE_CONVERSION (return_slot_addr
);
1644 /* We are going to construct *&return_slot and we can't do that
1645 for variables believed to be not addressable.
1647 FIXME: This check possibly can match, because values returned
1648 via return slot optimization are not believed to have address
1649 taken by alias analysis. */
1650 gcc_assert (TREE_CODE (return_slot
) != SSA_NAME
);
1651 if (gimple_in_ssa_p (cfun
))
1653 HOST_WIDE_INT bitsize
;
1654 HOST_WIDE_INT bitpos
;
1656 enum machine_mode mode
;
1660 base
= get_inner_reference (return_slot
, &bitsize
, &bitpos
,
1662 &mode
, &unsignedp
, &volatilep
,
1664 if (TREE_CODE (base
) == INDIRECT_REF
)
1665 base
= TREE_OPERAND (base
, 0);
1666 if (TREE_CODE (base
) == SSA_NAME
)
1667 base
= SSA_NAME_VAR (base
);
1668 mark_sym_for_renaming (base
);
1670 var
= return_slot_addr
;
1675 gcc_assert (TREE_CODE (var
) != SSA_NAME
);
1677 if ((TREE_CODE (TREE_TYPE (result
)) == COMPLEX_TYPE
1678 || TREE_CODE (TREE_TYPE (result
)) == VECTOR_TYPE
)
1679 && !DECL_GIMPLE_REG_P (result
)
1681 DECL_GIMPLE_REG_P (var
) = 0;
1686 /* All types requiring non-trivial constructors should have been handled. */
1687 gcc_assert (!TREE_ADDRESSABLE (callee_type
));
1689 /* Attempt to avoid creating a new temporary variable. */
1691 && TREE_CODE (modify_dest
) != SSA_NAME
)
1693 bool use_it
= false;
1695 /* We can't use MODIFY_DEST if there's type promotion involved. */
1696 if (!useless_type_conversion_p (callee_type
, caller_type
))
1699 /* ??? If we're assigning to a variable sized type, then we must
1700 reuse the destination variable, because we've no good way to
1701 create variable sized temporaries at this point. */
1702 else if (TREE_CODE (TYPE_SIZE_UNIT (caller_type
)) != INTEGER_CST
)
1705 /* If the callee cannot possibly modify MODIFY_DEST, then we can
1706 reuse it as the result of the call directly. Don't do this if
1707 it would promote MODIFY_DEST to addressable. */
1708 else if (TREE_ADDRESSABLE (result
))
1712 tree base_m
= get_base_address (modify_dest
);
1714 /* If the base isn't a decl, then it's a pointer, and we don't
1715 know where that's going to go. */
1716 if (!DECL_P (base_m
))
1718 else if (is_global_var (base_m
))
1720 else if ((TREE_CODE (TREE_TYPE (result
)) == COMPLEX_TYPE
1721 || TREE_CODE (TREE_TYPE (result
)) == VECTOR_TYPE
)
1722 && !DECL_GIMPLE_REG_P (result
)
1723 && DECL_GIMPLE_REG_P (base_m
))
1725 else if (!TREE_ADDRESSABLE (base_m
))
1737 gcc_assert (TREE_CODE (TYPE_SIZE_UNIT (callee_type
)) == INTEGER_CST
);
1739 var
= copy_result_decl_to_var (result
, id
);
1740 if (gimple_in_ssa_p (cfun
))
1743 add_referenced_var (var
);
1746 DECL_SEEN_IN_BIND_EXPR_P (var
) = 1;
1747 DECL_STRUCT_FUNCTION (caller
)->unexpanded_var_list
1748 = tree_cons (NULL_TREE
, var
,
1749 DECL_STRUCT_FUNCTION (caller
)->unexpanded_var_list
);
1751 /* Do not have the rest of GCC warn about this variable as it should
1752 not be visible to the user. */
1753 TREE_NO_WARNING (var
) = 1;
1755 declare_inline_vars (id
->block
, var
);
1757 /* Build the use expr. If the return type of the function was
1758 promoted, convert it back to the expected type. */
1760 if (!useless_type_conversion_p (caller_type
, TREE_TYPE (var
)))
1761 use
= fold_convert (caller_type
, var
);
1763 STRIP_USELESS_TYPE_CONVERSION (use
);
1765 if (DECL_BY_REFERENCE (result
))
1766 var
= build_fold_addr_expr (var
);
1769 /* Register the VAR_DECL as the equivalent for the RESULT_DECL; that
1770 way, when the RESULT_DECL is encountered, it will be
1771 automatically replaced by the VAR_DECL. */
1772 insert_decl_map (id
, result
, var
);
1774 /* Remember this so we can ignore it in remap_decls. */
1781 /* Returns nonzero if a function can be inlined as a tree. */
1784 tree_inlinable_function_p (tree fn
)
1786 return inlinable_function_p (fn
);
1789 static const char *inline_forbidden_reason
;
1792 inline_forbidden_p_1 (tree
*nodep
, int *walk_subtrees ATTRIBUTE_UNUSED
,
1796 tree fn
= (tree
) fnp
;
1799 switch (TREE_CODE (node
))
1802 /* Refuse to inline alloca call unless user explicitly forced so as
1803 this may change program's memory overhead drastically when the
1804 function using alloca is called in loop. In GCC present in
1805 SPEC2000 inlining into schedule_block cause it to require 2GB of
1806 RAM instead of 256MB. */
1807 if (alloca_call_p (node
)
1808 && !lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn
)))
1810 inline_forbidden_reason
1811 = G_("function %q+F can never be inlined because it uses "
1812 "alloca (override using the always_inline attribute)");
1815 t
= get_callee_fndecl (node
);
1819 /* We cannot inline functions that call setjmp. */
1820 if (setjmp_call_p (t
))
1822 inline_forbidden_reason
1823 = G_("function %q+F can never be inlined because it uses setjmp");
1827 if (DECL_BUILT_IN_CLASS (t
) == BUILT_IN_NORMAL
)
1828 switch (DECL_FUNCTION_CODE (t
))
1830 /* We cannot inline functions that take a variable number of
1832 case BUILT_IN_VA_START
:
1833 case BUILT_IN_STDARG_START
:
1834 case BUILT_IN_NEXT_ARG
:
1835 case BUILT_IN_VA_END
:
1836 inline_forbidden_reason
1837 = G_("function %q+F can never be inlined because it "
1838 "uses variable argument lists");
1841 case BUILT_IN_LONGJMP
:
1842 /* We can't inline functions that call __builtin_longjmp at
1843 all. The non-local goto machinery really requires the
1844 destination be in a different function. If we allow the
1845 function calling __builtin_longjmp to be inlined into the
1846 function calling __builtin_setjmp, Things will Go Awry. */
1847 inline_forbidden_reason
1848 = G_("function %q+F can never be inlined because "
1849 "it uses setjmp-longjmp exception handling");
1852 case BUILT_IN_NONLOCAL_GOTO
:
1854 inline_forbidden_reason
1855 = G_("function %q+F can never be inlined because "
1856 "it uses non-local goto");
1859 case BUILT_IN_RETURN
:
1860 case BUILT_IN_APPLY_ARGS
:
1861 /* If a __builtin_apply_args caller would be inlined,
1862 it would be saving arguments of the function it has
1863 been inlined into. Similarly __builtin_return would
1864 return from the function the inline has been inlined into. */
1865 inline_forbidden_reason
1866 = G_("function %q+F can never be inlined because "
1867 "it uses __builtin_return or __builtin_apply_args");
1876 t
= TREE_OPERAND (node
, 0);
1878 /* We will not inline a function which uses computed goto. The
1879 addresses of its local labels, which may be tucked into
1880 global storage, are of course not constant across
1881 instantiations, which causes unexpected behavior. */
1882 if (TREE_CODE (t
) != LABEL_DECL
)
1884 inline_forbidden_reason
1885 = G_("function %q+F can never be inlined "
1886 "because it contains a computed goto");
1892 t
= TREE_OPERAND (node
, 0);
1893 if (DECL_NONLOCAL (t
))
1895 /* We cannot inline a function that receives a non-local goto
1896 because we cannot remap the destination label used in the
1897 function that is performing the non-local goto. */
1898 inline_forbidden_reason
1899 = G_("function %q+F can never be inlined "
1900 "because it receives a non-local goto");
1907 /* We cannot inline a function of the form
1909 void F (int i) { struct S { int ar[i]; } s; }
1911 Attempting to do so produces a catch-22.
1912 If walk_tree examines the TYPE_FIELDS chain of RECORD_TYPE/
1913 UNION_TYPE nodes, then it goes into infinite recursion on a
1914 structure containing a pointer to its own type. If it doesn't,
1915 then the type node for S doesn't get adjusted properly when
1918 ??? This is likely no longer true, but it's too late in the 4.0
1919 cycle to try to find out. This should be checked for 4.1. */
1920 for (t
= TYPE_FIELDS (node
); t
; t
= TREE_CHAIN (t
))
1921 if (variably_modified_type_p (TREE_TYPE (t
), NULL
))
1923 inline_forbidden_reason
1924 = G_("function %q+F can never be inlined "
1925 "because it uses variable sized variables");
1936 /* Return subexpression representing possible alloca call, if any. */
1938 inline_forbidden_p (tree fndecl
)
1940 location_t saved_loc
= input_location
;
1941 block_stmt_iterator bsi
;
1943 tree ret
= NULL_TREE
;
1945 FOR_EACH_BB_FN (bb
, DECL_STRUCT_FUNCTION (fndecl
))
1946 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
1948 ret
= walk_tree_without_duplicates (bsi_stmt_ptr (bsi
),
1949 inline_forbidden_p_1
, fndecl
);
1955 input_location
= saved_loc
;
1959 /* Returns nonzero if FN is a function that does not have any
1960 fundamental inline blocking properties. */
1963 inlinable_function_p (tree fn
)
1965 bool inlinable
= true;
1969 /* If we've already decided this function shouldn't be inlined,
1970 there's no need to check again. */
1971 if (DECL_UNINLINABLE (fn
))
1974 /* We only warn for functions declared `inline' by the user. */
1975 do_warning
= (warn_inline
1977 && DECL_DECLARED_INLINE_P (fn
)
1978 && !DECL_IN_SYSTEM_HEADER (fn
));
1980 always_inline
= lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn
));
1982 if (flag_really_no_inline
1983 && always_inline
== NULL
)
1986 warning (OPT_Winline
, "function %q+F can never be inlined because it "
1987 "is suppressed using -fno-inline", fn
);
1991 /* Don't auto-inline anything that might not be bound within
1992 this unit of translation. */
1993 else if (!DECL_DECLARED_INLINE_P (fn
)
1994 && DECL_REPLACEABLE_P (fn
))
1997 else if (!function_attribute_inlinable_p (fn
))
2000 warning (OPT_Winline
, "function %q+F can never be inlined because it "
2001 "uses attributes conflicting with inlining", fn
);
2005 /* If we don't have the function body available, we can't inline it.
2006 However, this should not be recorded since we also get here for
2007 forward declared inline functions. Therefore, return at once. */
2008 if (!DECL_SAVED_TREE (fn
))
2011 /* If we're not inlining at all, then we cannot inline this function. */
2012 else if (!flag_inline_trees
)
2015 /* Only try to inline functions if DECL_INLINE is set. This should be
2016 true for all functions declared `inline', and for all other functions
2017 as well with -finline-functions.
2019 Don't think of disregarding DECL_INLINE when flag_inline_trees == 2;
2020 it's the front-end that must set DECL_INLINE in this case, because
2021 dwarf2out loses if a function that does not have DECL_INLINE set is
2022 inlined anyway. That is why we have both DECL_INLINE and
2023 DECL_DECLARED_INLINE_P. */
2024 /* FIXME: When flag_inline_trees dies, the check for flag_unit_at_a_time
2025 here should be redundant. */
2026 else if (!DECL_INLINE (fn
) && !flag_unit_at_a_time
)
2029 else if (inline_forbidden_p (fn
))
2031 /* See if we should warn about uninlinable functions. Previously,
2032 some of these warnings would be issued while trying to expand
2033 the function inline, but that would cause multiple warnings
2034 about functions that would for example call alloca. But since
2035 this a property of the function, just one warning is enough.
2036 As a bonus we can now give more details about the reason why a
2037 function is not inlinable. */
2039 sorry (inline_forbidden_reason
, fn
);
2040 else if (do_warning
)
2041 warning (OPT_Winline
, inline_forbidden_reason
, fn
);
2046 /* Squirrel away the result so that we don't have to check again. */
2047 DECL_UNINLINABLE (fn
) = !inlinable
;
2052 /* Estimate the cost of a memory move. Use machine dependent
2053 word size and take possible memcpy call into account. */
2056 estimate_move_cost (tree type
)
2060 size
= int_size_in_bytes (type
);
2062 if (size
< 0 || size
> MOVE_MAX_PIECES
* MOVE_RATIO
)
2063 /* Cost of a memcpy call, 3 arguments and the call. */
2066 return ((size
+ MOVE_MAX_PIECES
- 1) / MOVE_MAX_PIECES
);
2069 /* Arguments for estimate_num_insns_1. */
2073 /* Used to return the number of insns. */
2076 /* Weights of various constructs. */
2077 eni_weights
*weights
;
2080 /* Used by estimate_num_insns. Estimate number of instructions seen
2081 by given statement. */
2084 estimate_num_insns_1 (tree
*tp
, int *walk_subtrees
, void *data
)
2086 struct eni_data
*d
= data
;
2090 if (IS_TYPE_OR_DECL_P (x
))
2095 /* Assume that constants and references counts nothing. These should
2096 be majorized by amount of operations among them we count later
2097 and are common target of CSE and similar optimizations. */
2098 else if (CONSTANT_CLASS_P (x
) || REFERENCE_CLASS_P (x
))
2101 switch (TREE_CODE (x
))
2103 /* Containers have no cost. */
2110 case ALIGN_INDIRECT_REF
:
2111 case MISALIGNED_INDIRECT_REF
:
2113 case ARRAY_RANGE_REF
:
2115 case EXC_PTR_EXPR
: /* ??? */
2116 case FILTER_EXPR
: /* ??? */
2119 case WITH_CLEANUP_EXPR
:
2122 case VIEW_CONVERT_EXPR
:
2127 case CASE_LABEL_EXPR
:
2130 case EH_FILTER_EXPR
:
2131 case STATEMENT_LIST
:
2133 case NON_LVALUE_EXPR
:
2136 case TRY_CATCH_EXPR
:
2137 case TRY_FINALLY_EXPR
:
2144 case WITH_SIZE_EXPR
:
2148 case OMP_SECTIONS_SWITCH
:
2151 /* We don't account constants for now. Assume that the cost is amortized
2152 by operations that do use them. We may re-consider this decision once
2153 we are able to optimize the tree before estimating its size and break
2154 out static initializers. */
2155 case IDENTIFIER_NODE
:
2165 /* CHANGE_DYNAMIC_TYPE_EXPR explicitly expands to nothing. */
2166 case CHANGE_DYNAMIC_TYPE_EXPR
:
2170 /* Try to estimate the cost of assignments. We have three cases to
2172 1) Simple assignments to registers;
2173 2) Stores to things that must live in memory. This includes
2174 "normal" stores to scalars, but also assignments of large
2175 structures, or constructors of big arrays;
2178 Let us look at the first two cases, assuming we have "a = b + C":
2179 <GIMPLE_MODIFY_STMT <var_decl "a">
2180 <plus_expr <var_decl "b"> <constant C>>
2181 If "a" is a GIMPLE register, the assignment to it is free on almost
2182 any target, because "a" usually ends up in a real register. Hence
2183 the only cost of this expression comes from the PLUS_EXPR, and we
2184 can ignore the GIMPLE_MODIFY_STMT.
2185 If "a" is not a GIMPLE register, the assignment to "a" will most
2186 likely be a real store, so the cost of the GIMPLE_MODIFY_STMT is the cost
2187 of moving something into "a", which we compute using the function
2190 The third case deals with TARGET_EXPRs, for which the semantics are
2191 that a temporary is assigned, unless the TARGET_EXPR itself is being
2192 assigned to something else. In the latter case we do not need the
2194 <GIMPLE_MODIFY_STMT <var_decl "a"> <target_expr>>, the
2195 GIMPLE_MODIFY_STMT is free. */
2197 case GIMPLE_MODIFY_STMT
:
2198 /* Is the right and side a TARGET_EXPR? */
2199 if (TREE_CODE (GENERIC_TREE_OPERAND (x
, 1)) == TARGET_EXPR
)
2201 /* ... fall through ... */
2204 x
= GENERIC_TREE_OPERAND (x
, 0);
2205 /* Is this an assignments to a register? */
2206 if (is_gimple_reg (x
))
2208 /* Otherwise it's a store, so fall through to compute the move cost. */
2211 d
->count
+= estimate_move_cost (TREE_TYPE (x
));
2214 /* Assign cost of 1 to usual operations.
2215 ??? We may consider mapping RTL costs to this. */
2220 case POINTER_PLUS_EXPR
:
2224 case FIXED_CONVERT_EXPR
:
2225 case FIX_TRUNC_EXPR
:
2237 case VEC_LSHIFT_EXPR
:
2238 case VEC_RSHIFT_EXPR
:
2245 case TRUTH_ANDIF_EXPR
:
2246 case TRUTH_ORIF_EXPR
:
2247 case TRUTH_AND_EXPR
:
2249 case TRUTH_XOR_EXPR
:
2250 case TRUTH_NOT_EXPR
:
2259 case UNORDERED_EXPR
:
2270 case PREDECREMENT_EXPR
:
2271 case PREINCREMENT_EXPR
:
2272 case POSTDECREMENT_EXPR
:
2273 case POSTINCREMENT_EXPR
:
2277 case REALIGN_LOAD_EXPR
:
2279 case REDUC_MAX_EXPR
:
2280 case REDUC_MIN_EXPR
:
2281 case REDUC_PLUS_EXPR
:
2282 case WIDEN_SUM_EXPR
:
2284 case VEC_WIDEN_MULT_HI_EXPR
:
2285 case VEC_WIDEN_MULT_LO_EXPR
:
2286 case VEC_UNPACK_HI_EXPR
:
2287 case VEC_UNPACK_LO_EXPR
:
2288 case VEC_UNPACK_FLOAT_HI_EXPR
:
2289 case VEC_UNPACK_FLOAT_LO_EXPR
:
2290 case VEC_PACK_TRUNC_EXPR
:
2291 case VEC_PACK_SAT_EXPR
:
2292 case VEC_PACK_FIX_TRUNC_EXPR
:
2294 case WIDEN_MULT_EXPR
:
2296 case VEC_EXTRACT_EVEN_EXPR
:
2297 case VEC_EXTRACT_ODD_EXPR
:
2298 case VEC_INTERLEAVE_HIGH_EXPR
:
2299 case VEC_INTERLEAVE_LOW_EXPR
:
2306 /* TODO: Cost of a switch should be derived from the number of
2308 d
->count
+= d
->weights
->switch_cost
;
2311 /* Few special cases of expensive operations. This is useful
2312 to avoid inlining on functions having too many of these. */
2313 case TRUNC_DIV_EXPR
:
2315 case FLOOR_DIV_EXPR
:
2316 case ROUND_DIV_EXPR
:
2317 case EXACT_DIV_EXPR
:
2318 case TRUNC_MOD_EXPR
:
2320 case FLOOR_MOD_EXPR
:
2321 case ROUND_MOD_EXPR
:
2323 d
->count
+= d
->weights
->div_mod_cost
;
2327 tree decl
= get_callee_fndecl (x
);
2329 if (decl
&& DECL_BUILT_IN_CLASS (decl
) == BUILT_IN_MD
)
2330 cost
= d
->weights
->target_builtin_call_cost
;
2332 cost
= d
->weights
->call_cost
;
2334 if (decl
&& DECL_BUILT_IN_CLASS (decl
) == BUILT_IN_NORMAL
)
2335 switch (DECL_FUNCTION_CODE (decl
))
2337 case BUILT_IN_CONSTANT_P
:
2340 case BUILT_IN_EXPECT
:
2342 /* Prefetch instruction is not expensive. */
2343 case BUILT_IN_PREFETCH
:
2350 /* Our cost must be kept in sync with cgraph_estimate_size_after_inlining
2351 that does use function declaration to figure out the arguments. */
2355 call_expr_arg_iterator iter
;
2356 FOR_EACH_CALL_EXPR_ARG (a
, iter
, x
)
2357 d
->count
+= estimate_move_cost (TREE_TYPE (a
));
2362 for (arg
= DECL_ARGUMENTS (decl
); arg
; arg
= TREE_CHAIN (arg
))
2363 d
->count
+= estimate_move_cost (TREE_TYPE (arg
));
2379 /* OpenMP directives are generally very expensive. */
2380 d
->count
+= d
->weights
->omp_cost
;
2389 /* Estimate number of instructions that will be created by expanding EXPR.
2390 WEIGHTS contains weights attributed to various constructs. */
2393 estimate_num_insns (tree expr
, eni_weights
*weights
)
2395 struct pointer_set_t
*visited_nodes
;
2397 block_stmt_iterator bsi
;
2398 struct function
*my_function
;
2399 struct eni_data data
;
2402 data
.weights
= weights
;
2404 /* If we're given an entire function, walk the CFG. */
2405 if (TREE_CODE (expr
) == FUNCTION_DECL
)
2407 my_function
= DECL_STRUCT_FUNCTION (expr
);
2408 gcc_assert (my_function
&& my_function
->cfg
);
2409 visited_nodes
= pointer_set_create ();
2410 FOR_EACH_BB_FN (bb
, my_function
)
2412 for (bsi
= bsi_start (bb
);
2416 walk_tree (bsi_stmt_ptr (bsi
), estimate_num_insns_1
,
2417 &data
, visited_nodes
);
2420 pointer_set_destroy (visited_nodes
);
2423 walk_tree_without_duplicates (&expr
, estimate_num_insns_1
, &data
);
2428 /* Initializes weights used by estimate_num_insns. */
2431 init_inline_once (void)
2433 eni_inlining_weights
.call_cost
= PARAM_VALUE (PARAM_INLINE_CALL_COST
);
2434 eni_inlining_weights
.target_builtin_call_cost
= 1;
2435 eni_inlining_weights
.div_mod_cost
= 10;
2436 eni_inlining_weights
.switch_cost
= 1;
2437 eni_inlining_weights
.omp_cost
= 40;
2439 eni_size_weights
.call_cost
= 1;
2440 eni_size_weights
.target_builtin_call_cost
= 1;
2441 eni_size_weights
.div_mod_cost
= 1;
2442 eni_size_weights
.switch_cost
= 10;
2443 eni_size_weights
.omp_cost
= 40;
2445 /* Estimating time for call is difficult, since we have no idea what the
2446 called function does. In the current uses of eni_time_weights,
2447 underestimating the cost does less harm than overestimating it, so
2448 we choose a rather small value here. */
2449 eni_time_weights
.call_cost
= 10;
2450 eni_time_weights
.target_builtin_call_cost
= 10;
2451 eni_time_weights
.div_mod_cost
= 10;
2452 eni_time_weights
.switch_cost
= 4;
2453 eni_time_weights
.omp_cost
= 40;
2456 /* Install new lexical TREE_BLOCK underneath 'current_block'. */
2458 add_lexical_block (tree current_block
, tree new_block
)
2462 /* Walk to the last sub-block. */
2463 for (blk_p
= &BLOCK_SUBBLOCKS (current_block
);
2465 blk_p
= &TREE_CHAIN (*blk_p
))
2468 BLOCK_SUPERCONTEXT (new_block
) = current_block
;
2471 /* If *TP is a CALL_EXPR, replace it with its inline expansion. */
2474 expand_call_inline (basic_block bb
, tree stmt
, tree
*tp
, void *data
)
2480 struct pointer_map_t
*st
;
2483 location_t saved_location
;
2484 struct cgraph_edge
*cg_edge
;
2486 basic_block return_block
;
2488 block_stmt_iterator bsi
, stmt_bsi
;
2489 bool successfully_inlined
= FALSE
;
2490 bool purge_dead_abnormal_edges
;
2494 /* See what we've got. */
2495 id
= (copy_body_data
*) data
;
2498 /* Set input_location here so we get the right instantiation context
2499 if we call instantiate_decl from inlinable_function_p. */
2500 saved_location
= input_location
;
2501 if (EXPR_HAS_LOCATION (t
))
2502 input_location
= EXPR_LOCATION (t
);
2504 /* From here on, we're only interested in CALL_EXPRs. */
2505 if (TREE_CODE (t
) != CALL_EXPR
)
2508 /* First, see if we can figure out what function is being called.
2509 If we cannot, then there is no hope of inlining the function. */
2510 fn
= get_callee_fndecl (t
);
2514 /* Turn forward declarations into real ones. */
2515 fn
= cgraph_node (fn
)->decl
;
2517 /* If fn is a declaration of a function in a nested scope that was
2518 globally declared inline, we don't set its DECL_INITIAL.
2519 However, we can't blindly follow DECL_ABSTRACT_ORIGIN because the
2520 C++ front-end uses it for cdtors to refer to their internal
2521 declarations, that are not real functions. Fortunately those
2522 don't have trees to be saved, so we can tell by checking their
2524 if (! DECL_INITIAL (fn
)
2525 && DECL_ABSTRACT_ORIGIN (fn
)
2526 && DECL_SAVED_TREE (DECL_ABSTRACT_ORIGIN (fn
)))
2527 fn
= DECL_ABSTRACT_ORIGIN (fn
);
2529 /* Objective C and fortran still calls tree_rest_of_compilation directly.
2530 Kill this check once this is fixed. */
2531 if (!id
->dst_node
->analyzed
)
2534 cg_edge
= cgraph_edge (id
->dst_node
, stmt
);
2536 /* Constant propagation on argument done during previous inlining
2537 may create new direct call. Produce an edge for it. */
2540 struct cgraph_node
*dest
= cgraph_node (fn
);
2542 /* We have missing edge in the callgraph. This can happen in one case
2543 where previous inlining turned indirect call into direct call by
2544 constant propagating arguments. In all other cases we hit a bug
2545 (incorrect node sharing is most common reason for missing edges. */
2546 gcc_assert (dest
->needed
|| !flag_unit_at_a_time
);
2547 cgraph_create_edge (id
->dst_node
, dest
, stmt
,
2548 bb
->count
, CGRAPH_FREQ_BASE
,
2549 bb
->loop_depth
)->inline_failed
2550 = N_("originally indirect function call not considered for inlining");
2553 fprintf (dump_file
, "Created new direct edge to %s",
2554 cgraph_node_name (dest
));
2559 /* Don't try to inline functions that are not well-suited to
2561 if (!cgraph_inline_p (cg_edge
, &reason
))
2563 if (lookup_attribute ("always_inline", DECL_ATTRIBUTES (fn
))
2564 /* Avoid warnings during early inline pass. */
2565 && (!flag_unit_at_a_time
|| cgraph_global_info_ready
))
2567 sorry ("inlining failed in call to %q+F: %s", fn
, reason
);
2568 sorry ("called from here");
2570 else if (warn_inline
&& DECL_DECLARED_INLINE_P (fn
)
2571 && !DECL_IN_SYSTEM_HEADER (fn
)
2573 && !lookup_attribute ("noinline", DECL_ATTRIBUTES (fn
))
2574 /* Avoid warnings during early inline pass. */
2575 && (!flag_unit_at_a_time
|| cgraph_global_info_ready
))
2577 warning (OPT_Winline
, "inlining failed in call to %q+F: %s",
2579 warning (OPT_Winline
, "called from here");
2583 fn
= cg_edge
->callee
->decl
;
2585 #ifdef ENABLE_CHECKING
2586 if (cg_edge
->callee
->decl
!= id
->dst_node
->decl
)
2587 verify_cgraph_node (cg_edge
->callee
);
2590 /* We will be inlining this callee. */
2591 id
->eh_region
= lookup_stmt_eh_region (stmt
);
2593 /* Split the block holding the CALL_EXPR. */
2594 e
= split_block (bb
, stmt
);
2596 return_block
= e
->dest
;
2599 /* split_block splits after the statement; work around this by
2600 moving the call into the second block manually. Not pretty,
2601 but seems easier than doing the CFG manipulation by hand
2602 when the CALL_EXPR is in the last statement of BB. */
2603 stmt_bsi
= bsi_last (bb
);
2604 bsi_remove (&stmt_bsi
, false);
2606 /* If the CALL_EXPR was in the last statement of BB, it may have
2607 been the source of abnormal edges. In this case, schedule
2608 the removal of dead abnormal edges. */
2609 bsi
= bsi_start (return_block
);
2610 if (bsi_end_p (bsi
))
2612 bsi_insert_after (&bsi
, stmt
, BSI_NEW_STMT
);
2613 purge_dead_abnormal_edges
= true;
2617 bsi_insert_before (&bsi
, stmt
, BSI_NEW_STMT
);
2618 purge_dead_abnormal_edges
= false;
2621 stmt_bsi
= bsi_start (return_block
);
2623 /* Build a block containing code to initialize the arguments, the
2624 actual inline expansion of the body, and a label for the return
2625 statements within the function to jump to. The type of the
2626 statement expression is the return type of the function call. */
2627 id
->block
= make_node (BLOCK
);
2628 BLOCK_ABSTRACT_ORIGIN (id
->block
) = fn
;
2629 BLOCK_SOURCE_LOCATION (id
->block
) = input_location
;
2630 add_lexical_block (TREE_BLOCK (stmt
), id
->block
);
2632 /* Local declarations will be replaced by their equivalents in this
2635 id
->decl_map
= pointer_map_create ();
2637 /* Record the function we are about to inline. */
2639 id
->src_node
= cg_edge
->callee
;
2640 id
->src_cfun
= DECL_STRUCT_FUNCTION (fn
);
2643 initialize_inlined_parameters (id
, t
, fn
, bb
);
2645 if (DECL_INITIAL (fn
))
2646 add_lexical_block (id
->block
, remap_blocks (DECL_INITIAL (fn
), id
));
2648 /* Return statements in the function body will be replaced by jumps
2649 to the RET_LABEL. */
2651 gcc_assert (DECL_INITIAL (fn
));
2652 gcc_assert (TREE_CODE (DECL_INITIAL (fn
)) == BLOCK
);
2654 /* Find the lhs to which the result of this call is assigned. */
2656 if (TREE_CODE (stmt
) == GIMPLE_MODIFY_STMT
)
2658 modify_dest
= GIMPLE_STMT_OPERAND (stmt
, 0);
2660 /* The function which we are inlining might not return a value,
2661 in which case we should issue a warning that the function
2662 does not return a value. In that case the optimizers will
2663 see that the variable to which the value is assigned was not
2664 initialized. We do not want to issue a warning about that
2665 uninitialized variable. */
2666 if (DECL_P (modify_dest
))
2667 TREE_NO_WARNING (modify_dest
) = 1;
2668 if (CALL_EXPR_RETURN_SLOT_OPT (t
))
2670 return_slot
= modify_dest
;
2677 /* Declare the return variable for the function. */
2678 declare_return_variable (id
, return_slot
,
2679 modify_dest
, &use_retvar
);
2681 /* This is it. Duplicate the callee body. Assume callee is
2682 pre-gimplified. Note that we must not alter the caller
2683 function in any way before this point, as this CALL_EXPR may be
2684 a self-referential call; if we're calling ourselves, we need to
2685 duplicate our body before altering anything. */
2686 copy_body (id
, bb
->count
, bb
->frequency
, bb
, return_block
);
2688 /* Add local vars in this inlined callee to caller. */
2689 t_step
= id
->src_cfun
->unexpanded_var_list
;
2690 for (; t_step
; t_step
= TREE_CHAIN (t_step
))
2692 var
= TREE_VALUE (t_step
);
2693 if (TREE_STATIC (var
) && !TREE_ASM_WRITTEN (var
))
2694 cfun
->unexpanded_var_list
= tree_cons (NULL_TREE
, var
,
2695 cfun
->unexpanded_var_list
);
2697 cfun
->unexpanded_var_list
= tree_cons (NULL_TREE
, remap_decl (var
, id
),
2698 cfun
->unexpanded_var_list
);
2702 pointer_map_destroy (id
->decl_map
);
2705 /* If the inlined function returns a result that we care about,
2706 clobber the CALL_EXPR with a reference to the return variable. */
2707 if (use_retvar
&& (TREE_CODE (bsi_stmt (stmt_bsi
)) != CALL_EXPR
))
2710 if (gimple_in_ssa_p (cfun
))
2713 mark_symbols_for_renaming (stmt
);
2715 maybe_clean_or_replace_eh_stmt (stmt
, stmt
);
2718 /* We're modifying a TSI owned by gimple_expand_calls_inline();
2719 tsi_delink() will leave the iterator in a sane state. */
2721 /* Handle case of inlining function that miss return statement so
2722 return value becomes undefined. */
2723 if (TREE_CODE (stmt
) == GIMPLE_MODIFY_STMT
2724 && TREE_CODE (GIMPLE_STMT_OPERAND (stmt
, 0)) == SSA_NAME
)
2726 tree name
= TREE_OPERAND (stmt
, 0);
2727 tree var
= SSA_NAME_VAR (TREE_OPERAND (stmt
, 0));
2728 tree def
= gimple_default_def (cfun
, var
);
2730 /* If the variable is used undefined, make this name undefined via
2734 TREE_OPERAND (stmt
, 1) = def
;
2737 /* Otherwise make this variable undefined. */
2740 bsi_remove (&stmt_bsi
, true);
2741 set_default_def (var
, name
);
2742 SSA_NAME_DEF_STMT (name
) = build_empty_stmt ();
2746 bsi_remove (&stmt_bsi
, true);
2749 if (purge_dead_abnormal_edges
)
2750 tree_purge_dead_abnormal_call_edges (return_block
);
2752 /* If the value of the new expression is ignored, that's OK. We
2753 don't warn about this for CALL_EXPRs, so we shouldn't warn about
2754 the equivalent inlined version either. */
2755 TREE_USED (*tp
) = 1;
2757 /* Output the inlining info for this abstract function, since it has been
2758 inlined. If we don't do this now, we can lose the information about the
2759 variables in the function when the blocks get blown away as soon as we
2760 remove the cgraph node. */
2761 (*debug_hooks
->outlining_inline_function
) (cg_edge
->callee
->decl
);
2763 /* Update callgraph if needed. */
2764 cgraph_remove_node (cg_edge
->callee
);
2766 id
->block
= NULL_TREE
;
2767 successfully_inlined
= TRUE
;
2770 input_location
= saved_location
;
2771 return successfully_inlined
;
2774 /* Expand call statements reachable from STMT_P.
2775 We can only have CALL_EXPRs as the "toplevel" tree code or nested
2776 in a GIMPLE_MODIFY_STMT. See tree-gimple.c:get_call_expr_in(). We can
2777 unfortunately not use that function here because we need a pointer
2778 to the CALL_EXPR, not the tree itself. */
2781 gimple_expand_calls_inline (basic_block bb
, copy_body_data
*id
)
2783 block_stmt_iterator bsi
;
2785 /* Register specific tree functions. */
2786 tree_register_cfg_hooks ();
2787 for (bsi
= bsi_start (bb
); !bsi_end_p (bsi
); bsi_next (&bsi
))
2789 tree
*expr_p
= bsi_stmt_ptr (bsi
);
2790 tree stmt
= *expr_p
;
2792 if (TREE_CODE (*expr_p
) == GIMPLE_MODIFY_STMT
)
2793 expr_p
= &GIMPLE_STMT_OPERAND (*expr_p
, 1);
2794 if (TREE_CODE (*expr_p
) == WITH_SIZE_EXPR
)
2795 expr_p
= &TREE_OPERAND (*expr_p
, 0);
2796 if (TREE_CODE (*expr_p
) == CALL_EXPR
)
2797 if (expand_call_inline (bb
, stmt
, expr_p
, id
))
2803 /* Walk all basic blocks created after FIRST and try to fold every statement
2804 in the STATEMENTS pointer set. */
2806 fold_marked_statements (int first
, struct pointer_set_t
*statements
)
2808 for (;first
< n_basic_blocks
;first
++)
2809 if (BASIC_BLOCK (first
))
2811 block_stmt_iterator bsi
;
2812 for (bsi
= bsi_start (BASIC_BLOCK (first
));
2813 !bsi_end_p (bsi
); bsi_next (&bsi
))
2814 if (pointer_set_contains (statements
, bsi_stmt (bsi
)))
2816 tree old_stmt
= bsi_stmt (bsi
);
2817 if (fold_stmt (bsi_stmt_ptr (bsi
)))
2819 update_stmt (bsi_stmt (bsi
));
2820 if (maybe_clean_or_replace_eh_stmt (old_stmt
, bsi_stmt (bsi
)))
2821 tree_purge_dead_eh_edges (BASIC_BLOCK (first
));
2827 /* Return true if BB has at least one abnormal outgoing edge. */
2830 has_abnormal_outgoing_edge_p (basic_block bb
)
2835 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
2836 if (e
->flags
& EDGE_ABNORMAL
)
2842 /* Expand calls to inline functions in the body of FN. */
2845 optimize_inline_calls (tree fn
)
2850 int last
= n_basic_blocks
;
2851 /* There is no point in performing inlining if errors have already
2852 occurred -- and we might crash if we try to inline invalid
2854 if (errorcount
|| sorrycount
)
2858 memset (&id
, 0, sizeof (id
));
2860 id
.src_node
= id
.dst_node
= cgraph_node (fn
);
2862 /* Or any functions that aren't finished yet. */
2863 prev_fn
= NULL_TREE
;
2864 if (current_function_decl
)
2866 id
.dst_fn
= current_function_decl
;
2867 prev_fn
= current_function_decl
;
2870 id
.copy_decl
= copy_decl_maybe_to_var
;
2871 id
.transform_call_graph_edges
= CB_CGE_DUPLICATE
;
2872 id
.transform_new_cfg
= false;
2873 id
.transform_return_to_modify
= true;
2874 id
.transform_lang_insert_block
= false;
2875 id
.statements_to_fold
= pointer_set_create ();
2877 push_gimplify_context ();
2879 /* We make no attempts to keep dominance info up-to-date. */
2880 free_dominance_info (CDI_DOMINATORS
);
2881 free_dominance_info (CDI_POST_DOMINATORS
);
2883 /* Reach the trees by walking over the CFG, and note the
2884 enclosing basic-blocks in the call edges. */
2885 /* We walk the blocks going forward, because inlined function bodies
2886 will split id->current_basic_block, and the new blocks will
2887 follow it; we'll trudge through them, processing their CALL_EXPRs
2890 gimple_expand_calls_inline (bb
, &id
);
2892 pop_gimplify_context (NULL
);
2894 #ifdef ENABLE_CHECKING
2896 struct cgraph_edge
*e
;
2898 verify_cgraph_node (id
.dst_node
);
2900 /* Double check that we inlined everything we are supposed to inline. */
2901 for (e
= id
.dst_node
->callees
; e
; e
= e
->next_callee
)
2902 gcc_assert (e
->inline_failed
);
2906 /* Fold the statements before compacting/renumbering the basic blocks. */
2907 fold_marked_statements (last
, id
.statements_to_fold
);
2908 pointer_set_destroy (id
.statements_to_fold
);
2910 /* Renumber the (code) basic_blocks consecutively. */
2912 /* Renumber the lexical scoping (non-code) blocks consecutively. */
2915 /* We are not going to maintain the cgraph edges up to date.
2916 Kill it so it won't confuse us. */
2917 cgraph_node_remove_callees (id
.dst_node
);
2919 fold_cond_expr_cond ();
2920 /* It would be nice to check SSA/CFG/statement consistency here, but it is
2921 not possible yet - the IPA passes might make various functions to not
2922 throw and they don't care to proactively update local EH info. This is
2923 done later in fixup_cfg pass that also execute the verification. */
2924 return (TODO_update_ssa
| TODO_cleanup_cfg
2925 | (gimple_in_ssa_p (cfun
) ? TODO_remove_unused_locals
: 0)
2926 | (profile_status
!= PROFILE_ABSENT
? TODO_rebuild_frequencies
: 0));
2929 /* FN is a function that has a complete body, and CLONE is a function whose
2930 body is to be set to a copy of FN, mapping argument declarations according
2931 to the ARG_MAP splay_tree. */
2934 clone_body (tree clone
, tree fn
, void *arg_map
)
2938 /* Clone the body, as if we were making an inline call. But, remap the
2939 parameters in the callee to the parameters of caller. */
2940 memset (&id
, 0, sizeof (id
));
2943 id
.src_cfun
= DECL_STRUCT_FUNCTION (fn
);
2944 id
.decl_map
= (struct pointer_map_t
*)arg_map
;
2946 id
.copy_decl
= copy_decl_no_change
;
2947 id
.transform_call_graph_edges
= CB_CGE_DUPLICATE
;
2948 id
.transform_new_cfg
= true;
2949 id
.transform_return_to_modify
= false;
2950 id
.transform_lang_insert_block
= true;
2952 /* We're not inside any EH region. */
2955 /* Actually copy the body. */
2956 append_to_statement_list_force (copy_generic_body (&id
), &DECL_SAVED_TREE (clone
));
2959 /* Passed to walk_tree. Copies the node pointed to, if appropriate. */
2962 copy_tree_r (tree
*tp
, int *walk_subtrees
, void *data ATTRIBUTE_UNUSED
)
2964 enum tree_code code
= TREE_CODE (*tp
);
2965 enum tree_code_class cl
= TREE_CODE_CLASS (code
);
2967 /* We make copies of most nodes. */
2968 if (IS_EXPR_CODE_CLASS (cl
)
2969 || IS_GIMPLE_STMT_CODE_CLASS (cl
)
2970 || code
== TREE_LIST
2972 || code
== TYPE_DECL
2973 || code
== OMP_CLAUSE
)
2975 /* Because the chain gets clobbered when we make a copy, we save it
2977 tree chain
= NULL_TREE
, new;
2979 if (!GIMPLE_TUPLE_P (*tp
))
2980 chain
= TREE_CHAIN (*tp
);
2982 /* Copy the node. */
2983 new = copy_node (*tp
);
2985 /* Propagate mudflap marked-ness. */
2986 if (flag_mudflap
&& mf_marked_p (*tp
))
2991 /* Now, restore the chain, if appropriate. That will cause
2992 walk_tree to walk into the chain as well. */
2993 if (code
== PARM_DECL
2994 || code
== TREE_LIST
2995 || code
== OMP_CLAUSE
)
2996 TREE_CHAIN (*tp
) = chain
;
2998 /* For now, we don't update BLOCKs when we make copies. So, we
2999 have to nullify all BIND_EXPRs. */
3000 if (TREE_CODE (*tp
) == BIND_EXPR
)
3001 BIND_EXPR_BLOCK (*tp
) = NULL_TREE
;
3003 else if (code
== CONSTRUCTOR
)
3005 /* CONSTRUCTOR nodes need special handling because
3006 we need to duplicate the vector of elements. */
3009 new = copy_node (*tp
);
3011 /* Propagate mudflap marked-ness. */
3012 if (flag_mudflap
&& mf_marked_p (*tp
))
3015 CONSTRUCTOR_ELTS (new) = VEC_copy (constructor_elt
, gc
,
3016 CONSTRUCTOR_ELTS (*tp
));
3019 else if (TREE_CODE_CLASS (code
) == tcc_type
)
3021 else if (TREE_CODE_CLASS (code
) == tcc_declaration
)
3023 else if (TREE_CODE_CLASS (code
) == tcc_constant
)
3026 gcc_assert (code
!= STATEMENT_LIST
);
3030 /* The SAVE_EXPR pointed to by TP is being copied. If ST contains
3031 information indicating to what new SAVE_EXPR this one should be mapped,
3032 use that one. Otherwise, create a new node and enter it in ST. FN is
3033 the function into which the copy will be placed. */
3036 remap_save_expr (tree
*tp
, void *st_
, int *walk_subtrees
)
3038 struct pointer_map_t
*st
= (struct pointer_map_t
*) st_
;
3042 /* See if we already encountered this SAVE_EXPR. */
3043 n
= (tree
*) pointer_map_contains (st
, *tp
);
3045 /* If we didn't already remap this SAVE_EXPR, do so now. */
3048 t
= copy_node (*tp
);
3050 /* Remember this SAVE_EXPR. */
3051 *pointer_map_insert (st
, *tp
) = t
;
3052 /* Make sure we don't remap an already-remapped SAVE_EXPR. */
3053 *pointer_map_insert (st
, t
) = t
;
3057 /* We've already walked into this SAVE_EXPR; don't do it again. */
3062 /* Replace this SAVE_EXPR with the copy. */
3066 /* Called via walk_tree. If *TP points to a DECL_STMT for a local label,
3067 copies the declaration and enters it in the splay_tree in DATA (which is
3068 really an `copy_body_data *'). */
3071 mark_local_for_remap_r (tree
*tp
, int *walk_subtrees ATTRIBUTE_UNUSED
,
3074 copy_body_data
*id
= (copy_body_data
*) data
;
3076 /* Don't walk into types. */
3080 else if (TREE_CODE (*tp
) == LABEL_EXPR
)
3082 tree decl
= TREE_OPERAND (*tp
, 0);
3084 /* Copy the decl and remember the copy. */
3085 insert_decl_map (id
, decl
, id
->copy_decl (decl
, id
));
3091 /* Perform any modifications to EXPR required when it is unsaved. Does
3092 not recurse into EXPR's subtrees. */
3095 unsave_expr_1 (tree expr
)
3097 switch (TREE_CODE (expr
))
3100 /* Don't mess with a TARGET_EXPR that hasn't been expanded.
3101 It's OK for this to happen if it was part of a subtree that
3102 isn't immediately expanded, such as operand 2 of another
3104 if (TREE_OPERAND (expr
, 1))
3107 TREE_OPERAND (expr
, 1) = TREE_OPERAND (expr
, 3);
3108 TREE_OPERAND (expr
, 3) = NULL_TREE
;
3116 /* Called via walk_tree when an expression is unsaved. Using the
3117 splay_tree pointed to by ST (which is really a `splay_tree'),
3118 remaps all local declarations to appropriate replacements. */
3121 unsave_r (tree
*tp
, int *walk_subtrees
, void *data
)
3123 copy_body_data
*id
= (copy_body_data
*) data
;
3124 struct pointer_map_t
*st
= id
->decl_map
;
3127 /* Only a local declaration (variable or label). */
3128 if ((TREE_CODE (*tp
) == VAR_DECL
&& !TREE_STATIC (*tp
))
3129 || TREE_CODE (*tp
) == LABEL_DECL
)
3131 /* Lookup the declaration. */
3132 n
= (tree
*) pointer_map_contains (st
, *tp
);
3134 /* If it's there, remap it. */
3139 else if (TREE_CODE (*tp
) == STATEMENT_LIST
)
3140 copy_statement_list (tp
);
3141 else if (TREE_CODE (*tp
) == BIND_EXPR
)
3142 copy_bind_expr (tp
, walk_subtrees
, id
);
3143 else if (TREE_CODE (*tp
) == SAVE_EXPR
)
3144 remap_save_expr (tp
, st
, walk_subtrees
);
3147 copy_tree_r (tp
, walk_subtrees
, NULL
);
3149 /* Do whatever unsaving is required. */
3150 unsave_expr_1 (*tp
);
3153 /* Keep iterating. */
3157 /* Copies everything in EXPR and replaces variables, labels
3158 and SAVE_EXPRs local to EXPR. */
3161 unsave_expr_now (tree expr
)
3165 /* There's nothing to do for NULL_TREE. */
3170 memset (&id
, 0, sizeof (id
));
3171 id
.src_fn
= current_function_decl
;
3172 id
.dst_fn
= current_function_decl
;
3173 id
.decl_map
= pointer_map_create ();
3175 id
.copy_decl
= copy_decl_no_change
;
3176 id
.transform_call_graph_edges
= CB_CGE_DUPLICATE
;
3177 id
.transform_new_cfg
= false;
3178 id
.transform_return_to_modify
= false;
3179 id
.transform_lang_insert_block
= false;
3181 /* Walk the tree once to find local labels. */
3182 walk_tree_without_duplicates (&expr
, mark_local_for_remap_r
, &id
);
3184 /* Walk the tree again, copying, remapping, and unsaving. */
3185 walk_tree (&expr
, unsave_r
, &id
, NULL
);
3188 pointer_map_destroy (id
.decl_map
);
3193 /* Allow someone to determine if SEARCH is a child of TOP from gdb. */
3196 debug_find_tree_1 (tree
*tp
, int *walk_subtrees ATTRIBUTE_UNUSED
, void *data
)
3205 debug_find_tree (tree top
, tree search
)
3207 return walk_tree_without_duplicates (&top
, debug_find_tree_1
, search
) != 0;
3211 /* Declare the variables created by the inliner. Add all the variables in
3212 VARS to BIND_EXPR. */
3215 declare_inline_vars (tree block
, tree vars
)
3218 for (t
= vars
; t
; t
= TREE_CHAIN (t
))
3220 DECL_SEEN_IN_BIND_EXPR_P (t
) = 1;
3221 gcc_assert (!TREE_STATIC (t
) && !TREE_ASM_WRITTEN (t
));
3222 cfun
->unexpanded_var_list
=
3223 tree_cons (NULL_TREE
, t
,
3224 cfun
->unexpanded_var_list
);
3228 BLOCK_VARS (block
) = chainon (BLOCK_VARS (block
), vars
);
3232 /* Copy NODE (which must be a DECL). The DECL originally was in the FROM_FN,
3233 but now it will be in the TO_FN. PARM_TO_VAR means enable PARM_DECL to
3234 VAR_DECL translation. */
3237 copy_decl_for_dup_finish (copy_body_data
*id
, tree decl
, tree copy
)
3239 /* Don't generate debug information for the copy if we wouldn't have
3240 generated it for the copy either. */
3241 DECL_ARTIFICIAL (copy
) = DECL_ARTIFICIAL (decl
);
3242 DECL_IGNORED_P (copy
) = DECL_IGNORED_P (decl
);
3244 /* Set the DECL_ABSTRACT_ORIGIN so the debugging routines know what
3245 declaration inspired this copy. */
3246 DECL_ABSTRACT_ORIGIN (copy
) = DECL_ORIGIN (decl
);
3248 /* The new variable/label has no RTL, yet. */
3249 if (CODE_CONTAINS_STRUCT (TREE_CODE (copy
), TS_DECL_WRTL
)
3250 && !TREE_STATIC (copy
) && !DECL_EXTERNAL (copy
))
3251 SET_DECL_RTL (copy
, NULL_RTX
);
3253 /* These args would always appear unused, if not for this. */
3254 TREE_USED (copy
) = 1;
3256 /* Set the context for the new declaration. */
3257 if (!DECL_CONTEXT (decl
))
3258 /* Globals stay global. */
3260 else if (DECL_CONTEXT (decl
) != id
->src_fn
)
3261 /* Things that weren't in the scope of the function we're inlining
3262 from aren't in the scope we're inlining to, either. */
3264 else if (TREE_STATIC (decl
))
3265 /* Function-scoped static variables should stay in the original
3269 /* Ordinary automatic local variables are now in the scope of the
3271 DECL_CONTEXT (copy
) = id
->dst_fn
;
3277 copy_decl_to_var (tree decl
, copy_body_data
*id
)
3281 gcc_assert (TREE_CODE (decl
) == PARM_DECL
3282 || TREE_CODE (decl
) == RESULT_DECL
);
3284 type
= TREE_TYPE (decl
);
3286 copy
= build_decl (VAR_DECL
, DECL_NAME (decl
), type
);
3287 TREE_ADDRESSABLE (copy
) = TREE_ADDRESSABLE (decl
);
3288 TREE_READONLY (copy
) = TREE_READONLY (decl
);
3289 TREE_THIS_VOLATILE (copy
) = TREE_THIS_VOLATILE (decl
);
3290 DECL_GIMPLE_REG_P (copy
) = DECL_GIMPLE_REG_P (decl
);
3291 DECL_NO_TBAA_P (copy
) = DECL_NO_TBAA_P (decl
);
3293 return copy_decl_for_dup_finish (id
, decl
, copy
);
3296 /* Like copy_decl_to_var, but create a return slot object instead of a
3297 pointer variable for return by invisible reference. */
3300 copy_result_decl_to_var (tree decl
, copy_body_data
*id
)
3304 gcc_assert (TREE_CODE (decl
) == PARM_DECL
3305 || TREE_CODE (decl
) == RESULT_DECL
);
3307 type
= TREE_TYPE (decl
);
3308 if (DECL_BY_REFERENCE (decl
))
3309 type
= TREE_TYPE (type
);
3311 copy
= build_decl (VAR_DECL
, DECL_NAME (decl
), type
);
3312 TREE_READONLY (copy
) = TREE_READONLY (decl
);
3313 TREE_THIS_VOLATILE (copy
) = TREE_THIS_VOLATILE (decl
);
3314 if (!DECL_BY_REFERENCE (decl
))
3316 TREE_ADDRESSABLE (copy
) = TREE_ADDRESSABLE (decl
);
3317 DECL_GIMPLE_REG_P (copy
) = DECL_GIMPLE_REG_P (decl
);
3318 DECL_NO_TBAA_P (copy
) = DECL_NO_TBAA_P (decl
);
3321 return copy_decl_for_dup_finish (id
, decl
, copy
);
3326 copy_decl_no_change (tree decl
, copy_body_data
*id
)
3330 copy
= copy_node (decl
);
3332 /* The COPY is not abstract; it will be generated in DST_FN. */
3333 DECL_ABSTRACT (copy
) = 0;
3334 lang_hooks
.dup_lang_specific_decl (copy
);
3336 /* TREE_ADDRESSABLE isn't used to indicate that a label's address has
3337 been taken; it's for internal bookkeeping in expand_goto_internal. */
3338 if (TREE_CODE (copy
) == LABEL_DECL
)
3340 TREE_ADDRESSABLE (copy
) = 0;
3341 LABEL_DECL_UID (copy
) = -1;
3344 return copy_decl_for_dup_finish (id
, decl
, copy
);
3348 copy_decl_maybe_to_var (tree decl
, copy_body_data
*id
)
3350 if (TREE_CODE (decl
) == PARM_DECL
|| TREE_CODE (decl
) == RESULT_DECL
)
3351 return copy_decl_to_var (decl
, id
);
3353 return copy_decl_no_change (decl
, id
);
3356 /* Return a copy of the function's argument tree. */
3358 copy_arguments_for_versioning (tree orig_parm
, copy_body_data
* id
)
3360 tree
*arg_copy
, *parg
;
3362 arg_copy
= &orig_parm
;
3363 for (parg
= arg_copy
; *parg
; parg
= &TREE_CHAIN (*parg
))
3365 tree
new = remap_decl (*parg
, id
);
3366 lang_hooks
.dup_lang_specific_decl (new);
3367 TREE_CHAIN (new) = TREE_CHAIN (*parg
);
3373 /* Return a copy of the function's static chain. */
3375 copy_static_chain (tree static_chain
, copy_body_data
* id
)
3377 tree
*chain_copy
, *pvar
;
3379 chain_copy
= &static_chain
;
3380 for (pvar
= chain_copy
; *pvar
; pvar
= &TREE_CHAIN (*pvar
))
3382 tree
new = remap_decl (*pvar
, id
);
3383 lang_hooks
.dup_lang_specific_decl (new);
3384 TREE_CHAIN (new) = TREE_CHAIN (*pvar
);
3387 return static_chain
;
3390 /* Return true if the function is allowed to be versioned.
3391 This is a guard for the versioning functionality. */
3393 tree_versionable_function_p (tree fndecl
)
3395 if (fndecl
== NULL_TREE
)
3397 /* ??? There are cases where a function is
3398 uninlinable but can be versioned. */
3399 if (!tree_inlinable_function_p (fndecl
))
3405 /* Create a copy of a function's tree.
3406 OLD_DECL and NEW_DECL are FUNCTION_DECL tree nodes
3407 of the original function and the new copied function
3408 respectively. In case we want to replace a DECL
3409 tree with another tree while duplicating the function's
3410 body, TREE_MAP represents the mapping between these
3411 trees. If UPDATE_CLONES is set, the call_stmt fields
3412 of edges of clones of the function will be updated. */
3414 tree_function_versioning (tree old_decl
, tree new_decl
, varray_type tree_map
,
3417 struct cgraph_node
*old_version_node
;
3418 struct cgraph_node
*new_version_node
;
3422 struct ipa_replace_map
*replace_info
;
3423 basic_block old_entry_block
;
3425 tree old_current_function_decl
= current_function_decl
;
3427 gcc_assert (TREE_CODE (old_decl
) == FUNCTION_DECL
3428 && TREE_CODE (new_decl
) == FUNCTION_DECL
);
3429 DECL_POSSIBLY_INLINED (old_decl
) = 1;
3431 old_version_node
= cgraph_node (old_decl
);
3432 new_version_node
= cgraph_node (new_decl
);
3434 DECL_ARTIFICIAL (new_decl
) = 1;
3435 DECL_ABSTRACT_ORIGIN (new_decl
) = DECL_ORIGIN (old_decl
);
3437 /* Prepare the data structures for the tree copy. */
3438 memset (&id
, 0, sizeof (id
));
3440 /* Generate a new name for the new version. */
3443 DECL_NAME (new_decl
) = create_tmp_var_name (NULL
);
3444 SET_DECL_ASSEMBLER_NAME (new_decl
, DECL_NAME (new_decl
));
3445 SET_DECL_RTL (new_decl
, NULL_RTX
);
3446 id
.statements_to_fold
= pointer_set_create ();
3449 id
.decl_map
= pointer_map_create ();
3450 id
.src_fn
= old_decl
;
3451 id
.dst_fn
= new_decl
;
3452 id
.src_node
= old_version_node
;
3453 id
.dst_node
= new_version_node
;
3454 id
.src_cfun
= DECL_STRUCT_FUNCTION (old_decl
);
3456 id
.copy_decl
= copy_decl_no_change
;
3457 id
.transform_call_graph_edges
3458 = update_clones
? CB_CGE_MOVE_CLONES
: CB_CGE_MOVE
;
3459 id
.transform_new_cfg
= true;
3460 id
.transform_return_to_modify
= false;
3461 id
.transform_lang_insert_block
= false;
3463 current_function_decl
= new_decl
;
3464 old_entry_block
= ENTRY_BLOCK_PTR_FOR_FUNCTION
3465 (DECL_STRUCT_FUNCTION (old_decl
));
3466 initialize_cfun (new_decl
, old_decl
,
3467 old_entry_block
->count
,
3468 old_entry_block
->frequency
);
3469 push_cfun (DECL_STRUCT_FUNCTION (new_decl
));
3471 /* Copy the function's static chain. */
3472 p
= DECL_STRUCT_FUNCTION (old_decl
)->static_chain_decl
;
3474 DECL_STRUCT_FUNCTION (new_decl
)->static_chain_decl
=
3475 copy_static_chain (DECL_STRUCT_FUNCTION (old_decl
)->static_chain_decl
,
3477 /* Copy the function's arguments. */
3478 if (DECL_ARGUMENTS (old_decl
) != NULL_TREE
)
3479 DECL_ARGUMENTS (new_decl
) =
3480 copy_arguments_for_versioning (DECL_ARGUMENTS (old_decl
), &id
);
3482 /* If there's a tree_map, prepare for substitution. */
3484 for (i
= 0; i
< VARRAY_ACTIVE_SIZE (tree_map
); i
++)
3486 replace_info
= VARRAY_GENERIC_PTR (tree_map
, i
);
3487 if (replace_info
->replace_p
)
3488 insert_decl_map (&id
, replace_info
->old_tree
,
3489 replace_info
->new_tree
);
3492 DECL_INITIAL (new_decl
) = remap_blocks (DECL_INITIAL (id
.src_fn
), &id
);
3494 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3495 number_blocks (id
.dst_fn
);
3497 if (DECL_STRUCT_FUNCTION (old_decl
)->unexpanded_var_list
!= NULL_TREE
)
3498 /* Add local vars. */
3499 for (t_step
= DECL_STRUCT_FUNCTION (old_decl
)->unexpanded_var_list
;
3500 t_step
; t_step
= TREE_CHAIN (t_step
))
3502 tree var
= TREE_VALUE (t_step
);
3503 if (TREE_STATIC (var
) && !TREE_ASM_WRITTEN (var
))
3504 cfun
->unexpanded_var_list
= tree_cons (NULL_TREE
, var
,
3505 cfun
->unexpanded_var_list
);
3507 cfun
->unexpanded_var_list
=
3508 tree_cons (NULL_TREE
, remap_decl (var
, &id
),
3509 cfun
->unexpanded_var_list
);
3512 /* Copy the Function's body. */
3513 copy_body (&id
, old_entry_block
->count
, old_entry_block
->frequency
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
);
3515 if (DECL_RESULT (old_decl
) != NULL_TREE
)
3517 tree
*res_decl
= &DECL_RESULT (old_decl
);
3518 DECL_RESULT (new_decl
) = remap_decl (*res_decl
, &id
);
3519 lang_hooks
.dup_lang_specific_decl (DECL_RESULT (new_decl
));
3522 /* Renumber the lexical scoping (non-code) blocks consecutively. */
3523 number_blocks (new_decl
);
3526 pointer_map_destroy (id
.decl_map
);
3529 fold_marked_statements (0, id
.statements_to_fold
);
3530 pointer_set_destroy (id
.statements_to_fold
);
3531 fold_cond_expr_cond ();
3533 if (gimple_in_ssa_p (cfun
))
3535 free_dominance_info (CDI_DOMINATORS
);
3536 free_dominance_info (CDI_POST_DOMINATORS
);
3538 delete_unreachable_blocks ();
3539 update_ssa (TODO_update_ssa
);
3542 fold_cond_expr_cond ();
3543 if (need_ssa_update_p ())
3544 update_ssa (TODO_update_ssa
);
3547 free_dominance_info (CDI_DOMINATORS
);
3548 free_dominance_info (CDI_POST_DOMINATORS
);
3550 current_function_decl
= old_current_function_decl
;
3551 gcc_assert (!current_function_decl
3552 || DECL_STRUCT_FUNCTION (current_function_decl
) == cfun
);
3556 /* Duplicate a type, fields and all. */
3559 build_duplicate_type (tree type
)
3561 struct copy_body_data id
;
3563 memset (&id
, 0, sizeof (id
));
3564 id
.src_fn
= current_function_decl
;
3565 id
.dst_fn
= current_function_decl
;
3567 id
.decl_map
= pointer_map_create ();
3569 type
= remap_type_1 (type
, &id
);
3571 pointer_map_destroy (id
.decl_map
);