1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
4 Copyright (C) 2008-2014 Free Software Foundation, Inc.
5 Contributed by Martin Jambor <mjambor@suse.cz>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* This file implements Scalar Reduction of Aggregates (SRA). SRA is run
24 twice, once in the early stages of compilation (early SRA) and once in the
25 late stages (late SRA). The aim of both is to turn references to scalar
26 parts of aggregates into uses of independent scalar variables.
28 The two passes are nearly identical, the only difference is that early SRA
29 does not scalarize unions which are used as the result in a GIMPLE_RETURN
30 statement because together with inlining this can lead to weird type
33 Both passes operate in four stages:
35 1. The declarations that have properties which make them candidates for
36 scalarization are identified in function find_var_candidates(). The
37 candidates are stored in candidate_bitmap.
39 2. The function body is scanned. In the process, declarations which are
40 used in a manner that prevent their scalarization are removed from the
41 candidate bitmap. More importantly, for every access into an aggregate,
42 an access structure (struct access) is created by create_access() and
43 stored in a vector associated with the aggregate. Among other
44 information, the aggregate declaration, the offset and size of the access
45 and its type are stored in the structure.
47 On a related note, assign_link structures are created for every assign
48 statement between candidate aggregates and attached to the related
51 3. The vectors of accesses are analyzed. They are first sorted according to
52 their offset and size and then scanned for partially overlapping accesses
53 (i.e. those which overlap but one is not entirely within another). Such
54 an access disqualifies the whole aggregate from being scalarized.
56 If there is no such inhibiting overlap, a representative access structure
57 is chosen for every unique combination of offset and size. Afterwards,
58 the pass builds a set of trees from these structures, in which children
59 of an access are within their parent (in terms of offset and size).
61 Then accesses are propagated whenever possible (i.e. in cases when it
62 does not create a partially overlapping access) across assign_links from
63 the right hand side to the left hand side.
65 Then the set of trees for each declaration is traversed again and those
66 accesses which should be replaced by a scalar are identified.
68 4. The function is traversed again, and for every reference into an
69 aggregate that has some component which is about to be scalarized,
70 statements are amended and new statements are created as necessary.
71 Finally, if a parameter got scalarized, the scalar replacements are
72 initialized with values from respective parameter aggregates. */
76 #include "coretypes.h"
77 #include "hash-table.h"
78 #include "alloc-pool.h"
81 #include "pointer-set.h"
82 #include "basic-block.h"
83 #include "tree-ssa-alias.h"
84 #include "internal-fn.h"
86 #include "gimple-expr.h"
89 #include "stor-layout.h"
91 #include "gimple-iterator.h"
92 #include "gimplify-me.h"
93 #include "gimple-walk.h"
95 #include "gimple-ssa.h"
97 #include "tree-phinodes.h"
98 #include "ssa-iterators.h"
99 #include "stringpool.h"
100 #include "tree-ssanames.h"
102 #include "tree-dfa.h"
103 #include "tree-ssa.h"
104 #include "tree-pass.h"
105 #include "ipa-prop.h"
106 #include "statistics.h"
111 #include "tree-inline.h"
112 #include "gimple-pretty-print.h"
113 #include "ipa-inline.h"
114 #include "ipa-utils.h"
116 /* Enumeration of all aggregate reductions we can do. */
117 enum sra_mode
{ SRA_MODE_EARLY_IPA
, /* early call regularization */
118 SRA_MODE_EARLY_INTRA
, /* early intraprocedural SRA */
119 SRA_MODE_INTRA
}; /* late intraprocedural SRA */
121 /* Global variable describing which aggregate reduction we are performing at
123 static enum sra_mode sra_mode
;
127 /* ACCESS represents each access to an aggregate variable (as a whole or a
128 part). It can also represent a group of accesses that refer to exactly the
129 same fragment of an aggregate (i.e. those that have exactly the same offset
130 and size). Such representatives for a single aggregate, once determined,
131 are linked in a linked list and have the group fields set.
133 Moreover, when doing intraprocedural SRA, a tree is built from those
134 representatives (by the means of first_child and next_sibling pointers), in
135 which all items in a subtree are "within" the root, i.e. their offset is
136 greater or equal to offset of the root and offset+size is smaller or equal
137 to offset+size of the root. Children of an access are sorted by offset.
139 Note that accesses to parts of vector and complex number types always
140 represented by an access to the whole complex number or a vector. It is a
141 duty of the modifying functions to replace them appropriately. */
145 /* Values returned by `get_ref_base_and_extent' for each component reference
146 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
147 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
148 HOST_WIDE_INT offset
;
152 /* Expression. It is context dependent so do not use it to create new
153 expressions to access the original aggregate. See PR 42154 for a
159 /* The statement this access belongs to. */
162 /* Next group representative for this aggregate. */
163 struct access
*next_grp
;
165 /* Pointer to the group representative. Pointer to itself if the struct is
166 the representative. */
167 struct access
*group_representative
;
169 /* If this access has any children (in terms of the definition above), this
170 points to the first one. */
171 struct access
*first_child
;
173 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
174 described above. In IPA-SRA this is a pointer to the next access
175 belonging to the same group (having the same representative). */
176 struct access
*next_sibling
;
178 /* Pointers to the first and last element in the linked list of assign
180 struct assign_link
*first_link
, *last_link
;
182 /* Pointer to the next access in the work queue. */
183 struct access
*next_queued
;
185 /* Replacement variable for this access "region." Never to be accessed
186 directly, always only by the means of get_access_replacement() and only
187 when grp_to_be_replaced flag is set. */
188 tree replacement_decl
;
190 /* Is this particular access write access? */
193 /* Is this access an access to a non-addressable field? */
194 unsigned non_addressable
: 1;
196 /* Is this access currently in the work queue? */
197 unsigned grp_queued
: 1;
199 /* Does this group contain a write access? This flag is propagated down the
201 unsigned grp_write
: 1;
203 /* Does this group contain a read access? This flag is propagated down the
205 unsigned grp_read
: 1;
207 /* Does this group contain a read access that comes from an assignment
208 statement? This flag is propagated down the access tree. */
209 unsigned grp_assignment_read
: 1;
211 /* Does this group contain a write access that comes from an assignment
212 statement? This flag is propagated down the access tree. */
213 unsigned grp_assignment_write
: 1;
215 /* Does this group contain a read access through a scalar type? This flag is
216 not propagated in the access tree in any direction. */
217 unsigned grp_scalar_read
: 1;
219 /* Does this group contain a write access through a scalar type? This flag
220 is not propagated in the access tree in any direction. */
221 unsigned grp_scalar_write
: 1;
223 /* Is this access an artificial one created to scalarize some record
225 unsigned grp_total_scalarization
: 1;
227 /* Other passes of the analysis use this bit to make function
228 analyze_access_subtree create scalar replacements for this group if
230 unsigned grp_hint
: 1;
232 /* Is the subtree rooted in this access fully covered by scalar
234 unsigned grp_covered
: 1;
236 /* If set to true, this access and all below it in an access tree must not be
238 unsigned grp_unscalarizable_region
: 1;
240 /* Whether data have been written to parts of the aggregate covered by this
241 access which is not to be scalarized. This flag is propagated up in the
243 unsigned grp_unscalarized_data
: 1;
245 /* Does this access and/or group contain a write access through a
247 unsigned grp_partial_lhs
: 1;
249 /* Set when a scalar replacement should be created for this variable. */
250 unsigned grp_to_be_replaced
: 1;
252 /* Set when we want a replacement for the sole purpose of having it in
253 generated debug statements. */
254 unsigned grp_to_be_debug_replaced
: 1;
256 /* Should TREE_NO_WARNING of a replacement be set? */
257 unsigned grp_no_warning
: 1;
259 /* Is it possible that the group refers to data which might be (directly or
260 otherwise) modified? */
261 unsigned grp_maybe_modified
: 1;
263 /* Set when this is a representative of a pointer to scalar (i.e. by
264 reference) parameter which we consider for turning into a plain scalar
265 (i.e. a by value parameter). */
266 unsigned grp_scalar_ptr
: 1;
268 /* Set when we discover that this pointer is not safe to dereference in the
270 unsigned grp_not_necessarilly_dereferenced
: 1;
273 typedef struct access
*access_p
;
276 /* Alloc pool for allocating access structures. */
277 static alloc_pool access_pool
;
279 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
280 are used to propagate subaccesses from rhs to lhs as long as they don't
281 conflict with what is already there. */
284 struct access
*lacc
, *racc
;
285 struct assign_link
*next
;
288 /* Alloc pool for allocating assign link structures. */
289 static alloc_pool link_pool
;
291 /* Base (tree) -> Vector (vec<access_p> *) map. */
292 static struct pointer_map_t
*base_access_vec
;
294 /* Candidate hash table helpers. */
296 struct uid_decl_hasher
: typed_noop_remove
<tree_node
>
298 typedef tree_node value_type
;
299 typedef tree_node compare_type
;
300 static inline hashval_t
hash (const value_type
*);
301 static inline bool equal (const value_type
*, const compare_type
*);
304 /* Hash a tree in a uid_decl_map. */
307 uid_decl_hasher::hash (const value_type
*item
)
309 return item
->decl_minimal
.uid
;
312 /* Return true if the DECL_UID in both trees are equal. */
315 uid_decl_hasher::equal (const value_type
*a
, const compare_type
*b
)
317 return (a
->decl_minimal
.uid
== b
->decl_minimal
.uid
);
320 /* Set of candidates. */
321 static bitmap candidate_bitmap
;
322 static hash_table
<uid_decl_hasher
> candidates
;
324 /* For a candidate UID return the candidates decl. */
327 candidate (unsigned uid
)
330 t
.decl_minimal
.uid
= uid
;
331 return candidates
.find_with_hash (&t
, static_cast <hashval_t
> (uid
));
334 /* Bitmap of candidates which we should try to entirely scalarize away and
335 those which cannot be (because they are and need be used as a whole). */
336 static bitmap should_scalarize_away_bitmap
, cannot_scalarize_away_bitmap
;
338 /* Obstack for creation of fancy names. */
339 static struct obstack name_obstack
;
341 /* Head of a linked list of accesses that need to have its subaccesses
342 propagated to their assignment counterparts. */
343 static struct access
*work_queue_head
;
345 /* Number of parameters of the analyzed function when doing early ipa SRA. */
346 static int func_param_count
;
348 /* scan_function sets the following to true if it encounters a call to
349 __builtin_apply_args. */
350 static bool encountered_apply_args
;
352 /* Set by scan_function when it finds a recursive call. */
353 static bool encountered_recursive_call
;
355 /* Set by scan_function when it finds a recursive call with less actual
356 arguments than formal parameters.. */
357 static bool encountered_unchangable_recursive_call
;
359 /* This is a table in which for each basic block and parameter there is a
360 distance (offset + size) in that parameter which is dereferenced and
361 accessed in that BB. */
362 static HOST_WIDE_INT
*bb_dereferences
;
363 /* Bitmap of BBs that can cause the function to "stop" progressing by
364 returning, throwing externally, looping infinitely or calling a function
365 which might abort etc.. */
366 static bitmap final_bbs
;
368 /* Representative of no accesses at all. */
369 static struct access no_accesses_representant
;
371 /* Predicate to test the special value. */
374 no_accesses_p (struct access
*access
)
376 return access
== &no_accesses_representant
;
379 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
380 representative fields are dumped, otherwise those which only describe the
381 individual access are. */
385 /* Number of processed aggregates is readily available in
386 analyze_all_variable_accesses and so is not stored here. */
388 /* Number of created scalar replacements. */
391 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
395 /* Number of statements created by generate_subtree_copies. */
398 /* Number of statements created by load_assign_lhs_subreplacements. */
401 /* Number of times sra_modify_assign has deleted a statement. */
404 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
405 RHS reparately due to type conversions or nonexistent matching
407 int separate_lhs_rhs_handling
;
409 /* Number of parameters that were removed because they were unused. */
410 int deleted_unused_parameters
;
412 /* Number of scalars passed as parameters by reference that have been
413 converted to be passed by value. */
414 int scalar_by_ref_to_by_val
;
416 /* Number of aggregate parameters that were replaced by one or more of their
418 int aggregate_params_reduced
;
420 /* Numbber of components created when splitting aggregate parameters. */
421 int param_reductions_created
;
425 dump_access (FILE *f
, struct access
*access
, bool grp
)
427 fprintf (f
, "access { ");
428 fprintf (f
, "base = (%d)'", DECL_UID (access
->base
));
429 print_generic_expr (f
, access
->base
, 0);
430 fprintf (f
, "', offset = " HOST_WIDE_INT_PRINT_DEC
, access
->offset
);
431 fprintf (f
, ", size = " HOST_WIDE_INT_PRINT_DEC
, access
->size
);
432 fprintf (f
, ", expr = ");
433 print_generic_expr (f
, access
->expr
, 0);
434 fprintf (f
, ", type = ");
435 print_generic_expr (f
, access
->type
, 0);
437 fprintf (f
, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
438 "grp_assignment_write = %d, grp_scalar_read = %d, "
439 "grp_scalar_write = %d, grp_total_scalarization = %d, "
440 "grp_hint = %d, grp_covered = %d, "
441 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
442 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
443 "grp_to_be_debug_replaced = %d, grp_maybe_modified = %d, "
444 "grp_not_necessarilly_dereferenced = %d\n",
445 access
->grp_read
, access
->grp_write
, access
->grp_assignment_read
,
446 access
->grp_assignment_write
, access
->grp_scalar_read
,
447 access
->grp_scalar_write
, access
->grp_total_scalarization
,
448 access
->grp_hint
, access
->grp_covered
,
449 access
->grp_unscalarizable_region
, access
->grp_unscalarized_data
,
450 access
->grp_partial_lhs
, access
->grp_to_be_replaced
,
451 access
->grp_to_be_debug_replaced
, access
->grp_maybe_modified
,
452 access
->grp_not_necessarilly_dereferenced
);
454 fprintf (f
, ", write = %d, grp_total_scalarization = %d, "
455 "grp_partial_lhs = %d\n",
456 access
->write
, access
->grp_total_scalarization
,
457 access
->grp_partial_lhs
);
460 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
463 dump_access_tree_1 (FILE *f
, struct access
*access
, int level
)
469 for (i
= 0; i
< level
; i
++)
470 fputs ("* ", dump_file
);
472 dump_access (f
, access
, true);
474 if (access
->first_child
)
475 dump_access_tree_1 (f
, access
->first_child
, level
+ 1);
477 access
= access
->next_sibling
;
482 /* Dump all access trees for a variable, given the pointer to the first root in
486 dump_access_tree (FILE *f
, struct access
*access
)
488 for (; access
; access
= access
->next_grp
)
489 dump_access_tree_1 (f
, access
, 0);
492 /* Return true iff ACC is non-NULL and has subaccesses. */
495 access_has_children_p (struct access
*acc
)
497 return acc
&& acc
->first_child
;
500 /* Return true iff ACC is (partly) covered by at least one replacement. */
503 access_has_replacements_p (struct access
*acc
)
505 struct access
*child
;
506 if (acc
->grp_to_be_replaced
)
508 for (child
= acc
->first_child
; child
; child
= child
->next_sibling
)
509 if (access_has_replacements_p (child
))
514 /* Return a vector of pointers to accesses for the variable given in BASE or
515 NULL if there is none. */
517 static vec
<access_p
> *
518 get_base_access_vector (tree base
)
522 slot
= pointer_map_contains (base_access_vec
, base
);
526 return *(vec
<access_p
> **) slot
;
529 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
530 in ACCESS. Return NULL if it cannot be found. */
532 static struct access
*
533 find_access_in_subtree (struct access
*access
, HOST_WIDE_INT offset
,
536 while (access
&& (access
->offset
!= offset
|| access
->size
!= size
))
538 struct access
*child
= access
->first_child
;
540 while (child
&& (child
->offset
+ child
->size
<= offset
))
541 child
= child
->next_sibling
;
548 /* Return the first group representative for DECL or NULL if none exists. */
550 static struct access
*
551 get_first_repr_for_decl (tree base
)
553 vec
<access_p
> *access_vec
;
555 access_vec
= get_base_access_vector (base
);
559 return (*access_vec
)[0];
562 /* Find an access representative for the variable BASE and given OFFSET and
563 SIZE. Requires that access trees have already been built. Return NULL if
564 it cannot be found. */
566 static struct access
*
567 get_var_base_offset_size_access (tree base
, HOST_WIDE_INT offset
,
570 struct access
*access
;
572 access
= get_first_repr_for_decl (base
);
573 while (access
&& (access
->offset
+ access
->size
<= offset
))
574 access
= access
->next_grp
;
578 return find_access_in_subtree (access
, offset
, size
);
581 /* Add LINK to the linked list of assign links of RACC. */
583 add_link_to_rhs (struct access
*racc
, struct assign_link
*link
)
585 gcc_assert (link
->racc
== racc
);
587 if (!racc
->first_link
)
589 gcc_assert (!racc
->last_link
);
590 racc
->first_link
= link
;
593 racc
->last_link
->next
= link
;
595 racc
->last_link
= link
;
599 /* Move all link structures in their linked list in OLD_RACC to the linked list
602 relink_to_new_repr (struct access
*new_racc
, struct access
*old_racc
)
604 if (!old_racc
->first_link
)
606 gcc_assert (!old_racc
->last_link
);
610 if (new_racc
->first_link
)
612 gcc_assert (!new_racc
->last_link
->next
);
613 gcc_assert (!old_racc
->last_link
|| !old_racc
->last_link
->next
);
615 new_racc
->last_link
->next
= old_racc
->first_link
;
616 new_racc
->last_link
= old_racc
->last_link
;
620 gcc_assert (!new_racc
->last_link
);
622 new_racc
->first_link
= old_racc
->first_link
;
623 new_racc
->last_link
= old_racc
->last_link
;
625 old_racc
->first_link
= old_racc
->last_link
= NULL
;
628 /* Add ACCESS to the work queue (which is actually a stack). */
631 add_access_to_work_queue (struct access
*access
)
633 if (!access
->grp_queued
)
635 gcc_assert (!access
->next_queued
);
636 access
->next_queued
= work_queue_head
;
637 access
->grp_queued
= 1;
638 work_queue_head
= access
;
642 /* Pop an access from the work queue, and return it, assuming there is one. */
644 static struct access
*
645 pop_access_from_work_queue (void)
647 struct access
*access
= work_queue_head
;
649 work_queue_head
= access
->next_queued
;
650 access
->next_queued
= NULL
;
651 access
->grp_queued
= 0;
656 /* Allocate necessary structures. */
659 sra_initialize (void)
661 candidate_bitmap
= BITMAP_ALLOC (NULL
);
662 candidates
.create (vec_safe_length (cfun
->local_decls
) / 2);
663 should_scalarize_away_bitmap
= BITMAP_ALLOC (NULL
);
664 cannot_scalarize_away_bitmap
= BITMAP_ALLOC (NULL
);
665 gcc_obstack_init (&name_obstack
);
666 access_pool
= create_alloc_pool ("SRA accesses", sizeof (struct access
), 16);
667 link_pool
= create_alloc_pool ("SRA links", sizeof (struct assign_link
), 16);
668 base_access_vec
= pointer_map_create ();
669 memset (&sra_stats
, 0, sizeof (sra_stats
));
670 encountered_apply_args
= false;
671 encountered_recursive_call
= false;
672 encountered_unchangable_recursive_call
= false;
675 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
678 delete_base_accesses (const void *key ATTRIBUTE_UNUSED
, void **value
,
679 void *data ATTRIBUTE_UNUSED
)
681 vec
<access_p
> *access_vec
= (vec
<access_p
> *) *value
;
682 vec_free (access_vec
);
686 /* Deallocate all general structures. */
689 sra_deinitialize (void)
691 BITMAP_FREE (candidate_bitmap
);
692 candidates
.dispose ();
693 BITMAP_FREE (should_scalarize_away_bitmap
);
694 BITMAP_FREE (cannot_scalarize_away_bitmap
);
695 free_alloc_pool (access_pool
);
696 free_alloc_pool (link_pool
);
697 obstack_free (&name_obstack
, NULL
);
699 pointer_map_traverse (base_access_vec
, delete_base_accesses
, NULL
);
700 pointer_map_destroy (base_access_vec
);
703 /* Remove DECL from candidates for SRA and write REASON to the dump file if
706 disqualify_candidate (tree decl
, const char *reason
)
708 if (bitmap_clear_bit (candidate_bitmap
, DECL_UID (decl
)))
709 candidates
.clear_slot (candidates
.find_slot_with_hash (decl
,
713 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
715 fprintf (dump_file
, "! Disqualifying ");
716 print_generic_expr (dump_file
, decl
, 0);
717 fprintf (dump_file
, " - %s\n", reason
);
721 /* Return true iff the type contains a field or an element which does not allow
725 type_internals_preclude_sra_p (tree type
, const char **msg
)
730 switch (TREE_CODE (type
))
734 case QUAL_UNION_TYPE
:
735 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
736 if (TREE_CODE (fld
) == FIELD_DECL
)
738 tree ft
= TREE_TYPE (fld
);
740 if (TREE_THIS_VOLATILE (fld
))
742 *msg
= "volatile structure field";
745 if (!DECL_FIELD_OFFSET (fld
))
747 *msg
= "no structure field offset";
750 if (!DECL_SIZE (fld
))
752 *msg
= "zero structure field size";
755 if (!tree_fits_uhwi_p (DECL_FIELD_OFFSET (fld
)))
757 *msg
= "structure field offset not fixed";
760 if (!tree_fits_uhwi_p (DECL_SIZE (fld
)))
762 *msg
= "structure field size not fixed";
765 if (!tree_fits_shwi_p (bit_position (fld
)))
767 *msg
= "structure field size too big";
770 if (AGGREGATE_TYPE_P (ft
)
771 && int_bit_position (fld
) % BITS_PER_UNIT
!= 0)
773 *msg
= "structure field is bit field";
777 if (AGGREGATE_TYPE_P (ft
) && type_internals_preclude_sra_p (ft
, msg
))
784 et
= TREE_TYPE (type
);
786 if (TYPE_VOLATILE (et
))
788 *msg
= "element type is volatile";
792 if (AGGREGATE_TYPE_P (et
) && type_internals_preclude_sra_p (et
, msg
))
802 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
803 base variable if it is. Return T if it is not an SSA_NAME. */
806 get_ssa_base_param (tree t
)
808 if (TREE_CODE (t
) == SSA_NAME
)
810 if (SSA_NAME_IS_DEFAULT_DEF (t
))
811 return SSA_NAME_VAR (t
);
818 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
819 belongs to, unless the BB has already been marked as a potentially
823 mark_parm_dereference (tree base
, HOST_WIDE_INT dist
, gimple stmt
)
825 basic_block bb
= gimple_bb (stmt
);
826 int idx
, parm_index
= 0;
829 if (bitmap_bit_p (final_bbs
, bb
->index
))
832 for (parm
= DECL_ARGUMENTS (current_function_decl
);
833 parm
&& parm
!= base
;
834 parm
= DECL_CHAIN (parm
))
837 gcc_assert (parm_index
< func_param_count
);
839 idx
= bb
->index
* func_param_count
+ parm_index
;
840 if (bb_dereferences
[idx
] < dist
)
841 bb_dereferences
[idx
] = dist
;
844 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
845 the three fields. Also add it to the vector of accesses corresponding to
846 the base. Finally, return the new access. */
848 static struct access
*
849 create_access_1 (tree base
, HOST_WIDE_INT offset
, HOST_WIDE_INT size
)
852 struct access
*access
;
855 access
= (struct access
*) pool_alloc (access_pool
);
856 memset (access
, 0, sizeof (struct access
));
858 access
->offset
= offset
;
861 slot
= pointer_map_contains (base_access_vec
, base
);
863 v
= (vec
<access_p
> *) *slot
;
867 v
->safe_push (access
);
870 pointer_map_insert (base_access_vec
, base
)) = v
;
875 /* Create and insert access for EXPR. Return created access, or NULL if it is
878 static struct access
*
879 create_access (tree expr
, gimple stmt
, bool write
)
881 struct access
*access
;
882 HOST_WIDE_INT offset
, size
, max_size
;
884 bool ptr
, unscalarizable_region
= false;
886 base
= get_ref_base_and_extent (expr
, &offset
, &size
, &max_size
);
888 if (sra_mode
== SRA_MODE_EARLY_IPA
889 && TREE_CODE (base
) == MEM_REF
)
891 base
= get_ssa_base_param (TREE_OPERAND (base
, 0));
899 if (!DECL_P (base
) || !bitmap_bit_p (candidate_bitmap
, DECL_UID (base
)))
902 if (sra_mode
== SRA_MODE_EARLY_IPA
)
904 if (size
< 0 || size
!= max_size
)
906 disqualify_candidate (base
, "Encountered a variable sized access.");
909 if (TREE_CODE (expr
) == COMPONENT_REF
910 && DECL_BIT_FIELD (TREE_OPERAND (expr
, 1)))
912 disqualify_candidate (base
, "Encountered a bit-field access.");
915 gcc_checking_assert ((offset
% BITS_PER_UNIT
) == 0);
918 mark_parm_dereference (base
, offset
+ size
, stmt
);
922 if (size
!= max_size
)
925 unscalarizable_region
= true;
929 disqualify_candidate (base
, "Encountered an unconstrained access.");
934 access
= create_access_1 (base
, offset
, size
);
936 access
->type
= TREE_TYPE (expr
);
937 access
->write
= write
;
938 access
->grp_unscalarizable_region
= unscalarizable_region
;
941 if (TREE_CODE (expr
) == COMPONENT_REF
942 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr
, 1)))
943 access
->non_addressable
= 1;
949 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
950 register types or (recursively) records with only these two kinds of fields.
951 It also returns false if any of these records contains a bit-field. */
954 type_consists_of_records_p (tree type
)
958 if (TREE_CODE (type
) != RECORD_TYPE
)
961 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
962 if (TREE_CODE (fld
) == FIELD_DECL
)
964 tree ft
= TREE_TYPE (fld
);
966 if (DECL_BIT_FIELD (fld
))
969 if (!is_gimple_reg_type (ft
)
970 && !type_consists_of_records_p (ft
))
977 /* Create total_scalarization accesses for all scalar type fields in DECL that
978 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
979 must be the top-most VAR_DECL representing the variable, OFFSET must be the
980 offset of DECL within BASE. REF must be the memory reference expression for
984 completely_scalarize_record (tree base
, tree decl
, HOST_WIDE_INT offset
,
987 tree fld
, decl_type
= TREE_TYPE (decl
);
989 for (fld
= TYPE_FIELDS (decl_type
); fld
; fld
= DECL_CHAIN (fld
))
990 if (TREE_CODE (fld
) == FIELD_DECL
)
992 HOST_WIDE_INT pos
= offset
+ int_bit_position (fld
);
993 tree ft
= TREE_TYPE (fld
);
994 tree nref
= build3 (COMPONENT_REF
, TREE_TYPE (fld
), ref
, fld
,
997 if (is_gimple_reg_type (ft
))
999 struct access
*access
;
1002 size
= tree_to_uhwi (DECL_SIZE (fld
));
1003 access
= create_access_1 (base
, pos
, size
);
1004 access
->expr
= nref
;
1006 access
->grp_total_scalarization
= 1;
1007 /* Accesses for intraprocedural SRA can have their stmt NULL. */
1010 completely_scalarize_record (base
, fld
, pos
, nref
);
1014 /* Create total_scalarization accesses for all scalar type fields in VAR and
1015 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
1016 type_consists_of_records_p. */
1019 completely_scalarize_var (tree var
)
1021 HOST_WIDE_INT size
= tree_to_uhwi (DECL_SIZE (var
));
1022 struct access
*access
;
1024 access
= create_access_1 (var
, 0, size
);
1026 access
->type
= TREE_TYPE (var
);
1027 access
->grp_total_scalarization
= 1;
1029 completely_scalarize_record (var
, var
, 0, var
);
1032 /* Return true if REF has an VIEW_CONVERT_EXPR somewhere in it. */
1035 contains_view_convert_expr_p (const_tree ref
)
1037 while (handled_component_p (ref
))
1039 if (TREE_CODE (ref
) == VIEW_CONVERT_EXPR
)
1041 ref
= TREE_OPERAND (ref
, 0);
1047 /* Search the given tree for a declaration by skipping handled components and
1048 exclude it from the candidates. */
1051 disqualify_base_of_expr (tree t
, const char *reason
)
1053 t
= get_base_address (t
);
1054 if (sra_mode
== SRA_MODE_EARLY_IPA
1055 && TREE_CODE (t
) == MEM_REF
)
1056 t
= get_ssa_base_param (TREE_OPERAND (t
, 0));
1058 if (t
&& DECL_P (t
))
1059 disqualify_candidate (t
, reason
);
1062 /* Scan expression EXPR and create access structures for all accesses to
1063 candidates for scalarization. Return the created access or NULL if none is
1066 static struct access
*
1067 build_access_from_expr_1 (tree expr
, gimple stmt
, bool write
)
1069 struct access
*ret
= NULL
;
1072 if (TREE_CODE (expr
) == BIT_FIELD_REF
1073 || TREE_CODE (expr
) == IMAGPART_EXPR
1074 || TREE_CODE (expr
) == REALPART_EXPR
)
1076 expr
= TREE_OPERAND (expr
, 0);
1080 partial_ref
= false;
1082 /* We need to dive through V_C_Es in order to get the size of its parameter
1083 and not the result type. Ada produces such statements. We are also
1084 capable of handling the topmost V_C_E but not any of those buried in other
1085 handled components. */
1086 if (TREE_CODE (expr
) == VIEW_CONVERT_EXPR
)
1087 expr
= TREE_OPERAND (expr
, 0);
1089 if (contains_view_convert_expr_p (expr
))
1091 disqualify_base_of_expr (expr
, "V_C_E under a different handled "
1096 switch (TREE_CODE (expr
))
1099 if (TREE_CODE (TREE_OPERAND (expr
, 0)) != ADDR_EXPR
1100 && sra_mode
!= SRA_MODE_EARLY_IPA
)
1108 case ARRAY_RANGE_REF
:
1109 ret
= create_access (expr
, stmt
, write
);
1116 if (write
&& partial_ref
&& ret
)
1117 ret
->grp_partial_lhs
= 1;
1122 /* Scan expression EXPR and create access structures for all accesses to
1123 candidates for scalarization. Return true if any access has been inserted.
1124 STMT must be the statement from which the expression is taken, WRITE must be
1125 true if the expression is a store and false otherwise. */
1128 build_access_from_expr (tree expr
, gimple stmt
, bool write
)
1130 struct access
*access
;
1132 access
= build_access_from_expr_1 (expr
, stmt
, write
);
1135 /* This means the aggregate is accesses as a whole in a way other than an
1136 assign statement and thus cannot be removed even if we had a scalar
1137 replacement for everything. */
1138 if (cannot_scalarize_away_bitmap
)
1139 bitmap_set_bit (cannot_scalarize_away_bitmap
, DECL_UID (access
->base
));
1145 /* Return the single non-EH successor edge of BB or NULL if there is none or
1149 single_non_eh_succ (basic_block bb
)
1154 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1155 if (!(e
->flags
& EDGE_EH
))
1165 /* Disqualify LHS and RHS for scalarization if STMT has to terminate its BB and
1166 there is no alternative spot where to put statements SRA might need to
1167 generate after it. The spot we are looking for is an edge leading to a
1168 single non-EH successor, if it exists and is indeed single. RHS may be
1169 NULL, in that case ignore it. */
1172 disqualify_if_bad_bb_terminating_stmt (gimple stmt
, tree lhs
, tree rhs
)
1174 if ((sra_mode
== SRA_MODE_EARLY_INTRA
|| sra_mode
== SRA_MODE_INTRA
)
1175 && stmt_ends_bb_p (stmt
))
1177 if (single_non_eh_succ (gimple_bb (stmt
)))
1180 disqualify_base_of_expr (lhs
, "LHS of a throwing stmt.");
1182 disqualify_base_of_expr (rhs
, "RHS of a throwing stmt.");
1188 /* Scan expressions occurring in STMT, create access structures for all accesses
1189 to candidates for scalarization and remove those candidates which occur in
1190 statements or expressions that prevent them from being split apart. Return
1191 true if any access has been inserted. */
1194 build_accesses_from_assign (gimple stmt
)
1197 struct access
*lacc
, *racc
;
1199 if (!gimple_assign_single_p (stmt
)
1200 /* Scope clobbers don't influence scalarization. */
1201 || gimple_clobber_p (stmt
))
1204 lhs
= gimple_assign_lhs (stmt
);
1205 rhs
= gimple_assign_rhs1 (stmt
);
1207 if (disqualify_if_bad_bb_terminating_stmt (stmt
, lhs
, rhs
))
1210 racc
= build_access_from_expr_1 (rhs
, stmt
, false);
1211 lacc
= build_access_from_expr_1 (lhs
, stmt
, true);
1214 lacc
->grp_assignment_write
= 1;
1218 racc
->grp_assignment_read
= 1;
1219 if (should_scalarize_away_bitmap
&& !gimple_has_volatile_ops (stmt
)
1220 && !is_gimple_reg_type (racc
->type
))
1221 bitmap_set_bit (should_scalarize_away_bitmap
, DECL_UID (racc
->base
));
1225 && (sra_mode
== SRA_MODE_EARLY_INTRA
|| sra_mode
== SRA_MODE_INTRA
)
1226 && !lacc
->grp_unscalarizable_region
1227 && !racc
->grp_unscalarizable_region
1228 && AGGREGATE_TYPE_P (TREE_TYPE (lhs
))
1229 && lacc
->size
== racc
->size
1230 && useless_type_conversion_p (lacc
->type
, racc
->type
))
1232 struct assign_link
*link
;
1234 link
= (struct assign_link
*) pool_alloc (link_pool
);
1235 memset (link
, 0, sizeof (struct assign_link
));
1240 add_link_to_rhs (racc
, link
);
1243 return lacc
|| racc
;
1246 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1247 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1250 asm_visit_addr (gimple
, tree op
, tree
, void *)
1252 op
= get_base_address (op
);
1255 disqualify_candidate (op
, "Non-scalarizable GIMPLE_ASM operand.");
1260 /* Return true iff callsite CALL has at least as many actual arguments as there
1261 are formal parameters of the function currently processed by IPA-SRA and
1262 that their types match. */
1265 callsite_arguments_match_p (gimple call
)
1267 if (gimple_call_num_args (call
) < (unsigned) func_param_count
)
1272 for (parm
= DECL_ARGUMENTS (current_function_decl
), i
= 0;
1274 parm
= DECL_CHAIN (parm
), i
++)
1276 tree arg
= gimple_call_arg (call
, i
);
1277 if (!useless_type_conversion_p (TREE_TYPE (parm
), TREE_TYPE (arg
)))
1283 /* Scan function and look for interesting expressions and create access
1284 structures for them. Return true iff any access is created. */
1287 scan_function (void)
1292 FOR_EACH_BB_FN (bb
, cfun
)
1294 gimple_stmt_iterator gsi
;
1295 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1297 gimple stmt
= gsi_stmt (gsi
);
1301 if (final_bbs
&& stmt_can_throw_external (stmt
))
1302 bitmap_set_bit (final_bbs
, bb
->index
);
1303 switch (gimple_code (stmt
))
1306 t
= gimple_return_retval (stmt
);
1308 ret
|= build_access_from_expr (t
, stmt
, false);
1310 bitmap_set_bit (final_bbs
, bb
->index
);
1314 ret
|= build_accesses_from_assign (stmt
);
1318 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
1319 ret
|= build_access_from_expr (gimple_call_arg (stmt
, i
),
1322 if (sra_mode
== SRA_MODE_EARLY_IPA
)
1324 tree dest
= gimple_call_fndecl (stmt
);
1325 int flags
= gimple_call_flags (stmt
);
1329 if (DECL_BUILT_IN_CLASS (dest
) == BUILT_IN_NORMAL
1330 && DECL_FUNCTION_CODE (dest
) == BUILT_IN_APPLY_ARGS
)
1331 encountered_apply_args
= true;
1332 if (recursive_call_p (current_function_decl
, dest
))
1334 encountered_recursive_call
= true;
1335 if (!callsite_arguments_match_p (stmt
))
1336 encountered_unchangable_recursive_call
= true;
1341 && (flags
& (ECF_CONST
| ECF_PURE
)) == 0)
1342 bitmap_set_bit (final_bbs
, bb
->index
);
1345 t
= gimple_call_lhs (stmt
);
1346 if (t
&& !disqualify_if_bad_bb_terminating_stmt (stmt
, t
, NULL
))
1347 ret
|= build_access_from_expr (t
, stmt
, true);
1351 walk_stmt_load_store_addr_ops (stmt
, NULL
, NULL
, NULL
,
1354 bitmap_set_bit (final_bbs
, bb
->index
);
1356 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
1358 t
= TREE_VALUE (gimple_asm_input_op (stmt
, i
));
1359 ret
|= build_access_from_expr (t
, stmt
, false);
1361 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
1363 t
= TREE_VALUE (gimple_asm_output_op (stmt
, i
));
1364 ret
|= build_access_from_expr (t
, stmt
, true);
1377 /* Helper of QSORT function. There are pointers to accesses in the array. An
1378 access is considered smaller than another if it has smaller offset or if the
1379 offsets are the same but is size is bigger. */
1382 compare_access_positions (const void *a
, const void *b
)
1384 const access_p
*fp1
= (const access_p
*) a
;
1385 const access_p
*fp2
= (const access_p
*) b
;
1386 const access_p f1
= *fp1
;
1387 const access_p f2
= *fp2
;
1389 if (f1
->offset
!= f2
->offset
)
1390 return f1
->offset
< f2
->offset
? -1 : 1;
1392 if (f1
->size
== f2
->size
)
1394 if (f1
->type
== f2
->type
)
1396 /* Put any non-aggregate type before any aggregate type. */
1397 else if (!is_gimple_reg_type (f1
->type
)
1398 && is_gimple_reg_type (f2
->type
))
1400 else if (is_gimple_reg_type (f1
->type
)
1401 && !is_gimple_reg_type (f2
->type
))
1403 /* Put any complex or vector type before any other scalar type. */
1404 else if (TREE_CODE (f1
->type
) != COMPLEX_TYPE
1405 && TREE_CODE (f1
->type
) != VECTOR_TYPE
1406 && (TREE_CODE (f2
->type
) == COMPLEX_TYPE
1407 || TREE_CODE (f2
->type
) == VECTOR_TYPE
))
1409 else if ((TREE_CODE (f1
->type
) == COMPLEX_TYPE
1410 || TREE_CODE (f1
->type
) == VECTOR_TYPE
)
1411 && TREE_CODE (f2
->type
) != COMPLEX_TYPE
1412 && TREE_CODE (f2
->type
) != VECTOR_TYPE
)
1414 /* Put the integral type with the bigger precision first. */
1415 else if (INTEGRAL_TYPE_P (f1
->type
)
1416 && INTEGRAL_TYPE_P (f2
->type
))
1417 return TYPE_PRECISION (f2
->type
) - TYPE_PRECISION (f1
->type
);
1418 /* Put any integral type with non-full precision last. */
1419 else if (INTEGRAL_TYPE_P (f1
->type
)
1420 && (TREE_INT_CST_LOW (TYPE_SIZE (f1
->type
))
1421 != TYPE_PRECISION (f1
->type
)))
1423 else if (INTEGRAL_TYPE_P (f2
->type
)
1424 && (TREE_INT_CST_LOW (TYPE_SIZE (f2
->type
))
1425 != TYPE_PRECISION (f2
->type
)))
1427 /* Stabilize the sort. */
1428 return TYPE_UID (f1
->type
) - TYPE_UID (f2
->type
);
1431 /* We want the bigger accesses first, thus the opposite operator in the next
1433 return f1
->size
> f2
->size
? -1 : 1;
1437 /* Append a name of the declaration to the name obstack. A helper function for
1441 make_fancy_decl_name (tree decl
)
1445 tree name
= DECL_NAME (decl
);
1447 obstack_grow (&name_obstack
, IDENTIFIER_POINTER (name
),
1448 IDENTIFIER_LENGTH (name
));
1451 sprintf (buffer
, "D%u", DECL_UID (decl
));
1452 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1456 /* Helper for make_fancy_name. */
1459 make_fancy_name_1 (tree expr
)
1466 make_fancy_decl_name (expr
);
1470 switch (TREE_CODE (expr
))
1473 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1474 obstack_1grow (&name_obstack
, '$');
1475 make_fancy_decl_name (TREE_OPERAND (expr
, 1));
1479 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1480 obstack_1grow (&name_obstack
, '$');
1481 /* Arrays with only one element may not have a constant as their
1483 index
= TREE_OPERAND (expr
, 1);
1484 if (TREE_CODE (index
) != INTEGER_CST
)
1486 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
, TREE_INT_CST_LOW (index
));
1487 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1491 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1495 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1496 if (!integer_zerop (TREE_OPERAND (expr
, 1)))
1498 obstack_1grow (&name_obstack
, '$');
1499 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
,
1500 TREE_INT_CST_LOW (TREE_OPERAND (expr
, 1)));
1501 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1508 gcc_unreachable (); /* we treat these as scalars. */
1515 /* Create a human readable name for replacement variable of ACCESS. */
1518 make_fancy_name (tree expr
)
1520 make_fancy_name_1 (expr
);
1521 obstack_1grow (&name_obstack
, '\0');
1522 return XOBFINISH (&name_obstack
, char *);
1525 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1526 EXP_TYPE at the given OFFSET. If BASE is something for which
1527 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1528 to insert new statements either before or below the current one as specified
1529 by INSERT_AFTER. This function is not capable of handling bitfields.
1531 BASE must be either a declaration or a memory reference that has correct
1532 alignment ifformation embeded in it (e.g. a pre-existing one in SRA). */
1535 build_ref_for_offset (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1536 tree exp_type
, gimple_stmt_iterator
*gsi
,
1539 tree prev_base
= base
;
1542 HOST_WIDE_INT base_offset
;
1543 unsigned HOST_WIDE_INT misalign
;
1546 gcc_checking_assert (offset
% BITS_PER_UNIT
== 0);
1547 get_object_alignment_1 (base
, &align
, &misalign
);
1548 base
= get_addr_base_and_unit_offset (base
, &base_offset
);
1550 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1551 offset such as array[var_index]. */
1557 gcc_checking_assert (gsi
);
1558 tmp
= make_ssa_name (build_pointer_type (TREE_TYPE (prev_base
)), NULL
);
1559 addr
= build_fold_addr_expr (unshare_expr (prev_base
));
1560 STRIP_USELESS_TYPE_CONVERSION (addr
);
1561 stmt
= gimple_build_assign (tmp
, addr
);
1562 gimple_set_location (stmt
, loc
);
1564 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
1566 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1568 off
= build_int_cst (reference_alias_ptr_type (prev_base
),
1569 offset
/ BITS_PER_UNIT
);
1572 else if (TREE_CODE (base
) == MEM_REF
)
1574 off
= build_int_cst (TREE_TYPE (TREE_OPERAND (base
, 1)),
1575 base_offset
+ offset
/ BITS_PER_UNIT
);
1576 off
= int_const_binop (PLUS_EXPR
, TREE_OPERAND (base
, 1), off
);
1577 base
= unshare_expr (TREE_OPERAND (base
, 0));
1581 off
= build_int_cst (reference_alias_ptr_type (base
),
1582 base_offset
+ offset
/ BITS_PER_UNIT
);
1583 base
= build_fold_addr_expr (unshare_expr (base
));
1586 misalign
= (misalign
+ offset
) & (align
- 1);
1588 align
= (misalign
& -misalign
);
1589 if (align
< TYPE_ALIGN (exp_type
))
1590 exp_type
= build_aligned_type (exp_type
, align
);
1592 mem_ref
= fold_build2_loc (loc
, MEM_REF
, exp_type
, base
, off
);
1593 if (TREE_THIS_VOLATILE (prev_base
))
1594 TREE_THIS_VOLATILE (mem_ref
) = 1;
1595 if (TREE_SIDE_EFFECTS (prev_base
))
1596 TREE_SIDE_EFFECTS (mem_ref
) = 1;
1600 /* Construct a memory reference to a part of an aggregate BASE at the given
1601 OFFSET and of the same type as MODEL. In case this is a reference to a
1602 bit-field, the function will replicate the last component_ref of model's
1603 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1604 build_ref_for_offset. */
1607 build_ref_for_model (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1608 struct access
*model
, gimple_stmt_iterator
*gsi
,
1611 if (TREE_CODE (model
->expr
) == COMPONENT_REF
1612 && DECL_BIT_FIELD (TREE_OPERAND (model
->expr
, 1)))
1614 /* This access represents a bit-field. */
1615 tree t
, exp_type
, fld
= TREE_OPERAND (model
->expr
, 1);
1617 offset
-= int_bit_position (fld
);
1618 exp_type
= TREE_TYPE (TREE_OPERAND (model
->expr
, 0));
1619 t
= build_ref_for_offset (loc
, base
, offset
, exp_type
, gsi
, insert_after
);
1620 return fold_build3_loc (loc
, COMPONENT_REF
, TREE_TYPE (fld
), t
, fld
,
1624 return build_ref_for_offset (loc
, base
, offset
, model
->type
,
1628 /* Attempt to build a memory reference that we could but into a gimple
1629 debug_bind statement. Similar to build_ref_for_model but punts if it has to
1630 create statements and return s NULL instead. This function also ignores
1631 alignment issues and so its results should never end up in non-debug
1635 build_debug_ref_for_model (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1636 struct access
*model
)
1638 HOST_WIDE_INT base_offset
;
1641 if (TREE_CODE (model
->expr
) == COMPONENT_REF
1642 && DECL_BIT_FIELD (TREE_OPERAND (model
->expr
, 1)))
1645 base
= get_addr_base_and_unit_offset (base
, &base_offset
);
1648 if (TREE_CODE (base
) == MEM_REF
)
1650 off
= build_int_cst (TREE_TYPE (TREE_OPERAND (base
, 1)),
1651 base_offset
+ offset
/ BITS_PER_UNIT
);
1652 off
= int_const_binop (PLUS_EXPR
, TREE_OPERAND (base
, 1), off
);
1653 base
= unshare_expr (TREE_OPERAND (base
, 0));
1657 off
= build_int_cst (reference_alias_ptr_type (base
),
1658 base_offset
+ offset
/ BITS_PER_UNIT
);
1659 base
= build_fold_addr_expr (unshare_expr (base
));
1662 return fold_build2_loc (loc
, MEM_REF
, model
->type
, base
, off
);
1665 /* Construct a memory reference consisting of component_refs and array_refs to
1666 a part of an aggregate *RES (which is of type TYPE). The requested part
1667 should have type EXP_TYPE at be the given OFFSET. This function might not
1668 succeed, it returns true when it does and only then *RES points to something
1669 meaningful. This function should be used only to build expressions that we
1670 might need to present to user (e.g. in warnings). In all other situations,
1671 build_ref_for_model or build_ref_for_offset should be used instead. */
1674 build_user_friendly_ref_for_offset (tree
*res
, tree type
, HOST_WIDE_INT offset
,
1680 tree tr_size
, index
, minidx
;
1681 HOST_WIDE_INT el_size
;
1683 if (offset
== 0 && exp_type
1684 && types_compatible_p (exp_type
, type
))
1687 switch (TREE_CODE (type
))
1690 case QUAL_UNION_TYPE
:
1692 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
1694 HOST_WIDE_INT pos
, size
;
1695 tree tr_pos
, expr
, *expr_ptr
;
1697 if (TREE_CODE (fld
) != FIELD_DECL
)
1700 tr_pos
= bit_position (fld
);
1701 if (!tr_pos
|| !tree_fits_uhwi_p (tr_pos
))
1703 pos
= tree_to_uhwi (tr_pos
);
1704 gcc_assert (TREE_CODE (type
) == RECORD_TYPE
|| pos
== 0);
1705 tr_size
= DECL_SIZE (fld
);
1706 if (!tr_size
|| !tree_fits_uhwi_p (tr_size
))
1708 size
= tree_to_uhwi (tr_size
);
1714 else if (pos
> offset
|| (pos
+ size
) <= offset
)
1717 expr
= build3 (COMPONENT_REF
, TREE_TYPE (fld
), *res
, fld
,
1720 if (build_user_friendly_ref_for_offset (expr_ptr
, TREE_TYPE (fld
),
1721 offset
- pos
, exp_type
))
1730 tr_size
= TYPE_SIZE (TREE_TYPE (type
));
1731 if (!tr_size
|| !tree_fits_uhwi_p (tr_size
))
1733 el_size
= tree_to_uhwi (tr_size
);
1735 minidx
= TYPE_MIN_VALUE (TYPE_DOMAIN (type
));
1736 if (TREE_CODE (minidx
) != INTEGER_CST
|| el_size
== 0)
1738 index
= build_int_cst (TYPE_DOMAIN (type
), offset
/ el_size
);
1739 if (!integer_zerop (minidx
))
1740 index
= int_const_binop (PLUS_EXPR
, index
, minidx
);
1741 *res
= build4 (ARRAY_REF
, TREE_TYPE (type
), *res
, index
,
1742 NULL_TREE
, NULL_TREE
);
1743 offset
= offset
% el_size
;
1744 type
= TREE_TYPE (type
);
1759 /* Return true iff TYPE is stdarg va_list type. */
1762 is_va_list_type (tree type
)
1764 return TYPE_MAIN_VARIANT (type
) == TYPE_MAIN_VARIANT (va_list_type_node
);
1767 /* Print message to dump file why a variable was rejected. */
1770 reject (tree var
, const char *msg
)
1772 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1774 fprintf (dump_file
, "Rejected (%d): %s: ", DECL_UID (var
), msg
);
1775 print_generic_expr (dump_file
, var
, 0);
1776 fprintf (dump_file
, "\n");
1780 /* Return true if VAR is a candidate for SRA. */
1783 maybe_add_sra_candidate (tree var
)
1785 tree type
= TREE_TYPE (var
);
1789 if (!AGGREGATE_TYPE_P (type
))
1791 reject (var
, "not aggregate");
1794 if (needs_to_live_in_memory (var
))
1796 reject (var
, "needs to live in memory");
1799 if (TREE_THIS_VOLATILE (var
))
1801 reject (var
, "is volatile");
1804 if (!COMPLETE_TYPE_P (type
))
1806 reject (var
, "has incomplete type");
1809 if (!tree_fits_uhwi_p (TYPE_SIZE (type
)))
1811 reject (var
, "type size not fixed");
1814 if (tree_to_uhwi (TYPE_SIZE (type
)) == 0)
1816 reject (var
, "type size is zero");
1819 if (type_internals_preclude_sra_p (type
, &msg
))
1824 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1825 we also want to schedule it rather late. Thus we ignore it in
1827 (sra_mode
== SRA_MODE_EARLY_INTRA
1828 && is_va_list_type (type
)))
1830 reject (var
, "is va_list");
1834 bitmap_set_bit (candidate_bitmap
, DECL_UID (var
));
1835 slot
= candidates
.find_slot_with_hash (var
, DECL_UID (var
), INSERT
);
1838 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1840 fprintf (dump_file
, "Candidate (%d): ", DECL_UID (var
));
1841 print_generic_expr (dump_file
, var
, 0);
1842 fprintf (dump_file
, "\n");
1848 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1849 those with type which is suitable for scalarization. */
1852 find_var_candidates (void)
1858 for (parm
= DECL_ARGUMENTS (current_function_decl
);
1860 parm
= DECL_CHAIN (parm
))
1861 ret
|= maybe_add_sra_candidate (parm
);
1863 FOR_EACH_LOCAL_DECL (cfun
, i
, var
)
1865 if (TREE_CODE (var
) != VAR_DECL
)
1868 ret
|= maybe_add_sra_candidate (var
);
1874 /* Sort all accesses for the given variable, check for partial overlaps and
1875 return NULL if there are any. If there are none, pick a representative for
1876 each combination of offset and size and create a linked list out of them.
1877 Return the pointer to the first representative and make sure it is the first
1878 one in the vector of accesses. */
1880 static struct access
*
1881 sort_and_splice_var_accesses (tree var
)
1883 int i
, j
, access_count
;
1884 struct access
*res
, **prev_acc_ptr
= &res
;
1885 vec
<access_p
> *access_vec
;
1887 HOST_WIDE_INT low
= -1, high
= 0;
1889 access_vec
= get_base_access_vector (var
);
1892 access_count
= access_vec
->length ();
1894 /* Sort by <OFFSET, SIZE>. */
1895 access_vec
->qsort (compare_access_positions
);
1898 while (i
< access_count
)
1900 struct access
*access
= (*access_vec
)[i
];
1901 bool grp_write
= access
->write
;
1902 bool grp_read
= !access
->write
;
1903 bool grp_scalar_write
= access
->write
1904 && is_gimple_reg_type (access
->type
);
1905 bool grp_scalar_read
= !access
->write
1906 && is_gimple_reg_type (access
->type
);
1907 bool grp_assignment_read
= access
->grp_assignment_read
;
1908 bool grp_assignment_write
= access
->grp_assignment_write
;
1909 bool multiple_scalar_reads
= false;
1910 bool total_scalarization
= access
->grp_total_scalarization
;
1911 bool grp_partial_lhs
= access
->grp_partial_lhs
;
1912 bool first_scalar
= is_gimple_reg_type (access
->type
);
1913 bool unscalarizable_region
= access
->grp_unscalarizable_region
;
1915 if (first
|| access
->offset
>= high
)
1918 low
= access
->offset
;
1919 high
= access
->offset
+ access
->size
;
1921 else if (access
->offset
> low
&& access
->offset
+ access
->size
> high
)
1924 gcc_assert (access
->offset
>= low
1925 && access
->offset
+ access
->size
<= high
);
1928 while (j
< access_count
)
1930 struct access
*ac2
= (*access_vec
)[j
];
1931 if (ac2
->offset
!= access
->offset
|| ac2
->size
!= access
->size
)
1936 grp_scalar_write
= (grp_scalar_write
1937 || is_gimple_reg_type (ac2
->type
));
1942 if (is_gimple_reg_type (ac2
->type
))
1944 if (grp_scalar_read
)
1945 multiple_scalar_reads
= true;
1947 grp_scalar_read
= true;
1950 grp_assignment_read
|= ac2
->grp_assignment_read
;
1951 grp_assignment_write
|= ac2
->grp_assignment_write
;
1952 grp_partial_lhs
|= ac2
->grp_partial_lhs
;
1953 unscalarizable_region
|= ac2
->grp_unscalarizable_region
;
1954 total_scalarization
|= ac2
->grp_total_scalarization
;
1955 relink_to_new_repr (access
, ac2
);
1957 /* If there are both aggregate-type and scalar-type accesses with
1958 this combination of size and offset, the comparison function
1959 should have put the scalars first. */
1960 gcc_assert (first_scalar
|| !is_gimple_reg_type (ac2
->type
));
1961 ac2
->group_representative
= access
;
1967 access
->group_representative
= access
;
1968 access
->grp_write
= grp_write
;
1969 access
->grp_read
= grp_read
;
1970 access
->grp_scalar_read
= grp_scalar_read
;
1971 access
->grp_scalar_write
= grp_scalar_write
;
1972 access
->grp_assignment_read
= grp_assignment_read
;
1973 access
->grp_assignment_write
= grp_assignment_write
;
1974 access
->grp_hint
= multiple_scalar_reads
|| total_scalarization
;
1975 access
->grp_total_scalarization
= total_scalarization
;
1976 access
->grp_partial_lhs
= grp_partial_lhs
;
1977 access
->grp_unscalarizable_region
= unscalarizable_region
;
1978 if (access
->first_link
)
1979 add_access_to_work_queue (access
);
1981 *prev_acc_ptr
= access
;
1982 prev_acc_ptr
= &access
->next_grp
;
1985 gcc_assert (res
== (*access_vec
)[0]);
1989 /* Create a variable for the given ACCESS which determines the type, name and a
1990 few other properties. Return the variable declaration and store it also to
1991 ACCESS->replacement. */
1994 create_access_replacement (struct access
*access
)
1998 if (access
->grp_to_be_debug_replaced
)
2000 repl
= create_tmp_var_raw (access
->type
, NULL
);
2001 DECL_CONTEXT (repl
) = current_function_decl
;
2004 repl
= create_tmp_var (access
->type
, "SR");
2005 if (TREE_CODE (access
->type
) == COMPLEX_TYPE
2006 || TREE_CODE (access
->type
) == VECTOR_TYPE
)
2008 if (!access
->grp_partial_lhs
)
2009 DECL_GIMPLE_REG_P (repl
) = 1;
2011 else if (access
->grp_partial_lhs
2012 && is_gimple_reg_type (access
->type
))
2013 TREE_ADDRESSABLE (repl
) = 1;
2015 DECL_SOURCE_LOCATION (repl
) = DECL_SOURCE_LOCATION (access
->base
);
2016 DECL_ARTIFICIAL (repl
) = 1;
2017 DECL_IGNORED_P (repl
) = DECL_IGNORED_P (access
->base
);
2019 if (DECL_NAME (access
->base
)
2020 && !DECL_IGNORED_P (access
->base
)
2021 && !DECL_ARTIFICIAL (access
->base
))
2023 char *pretty_name
= make_fancy_name (access
->expr
);
2024 tree debug_expr
= unshare_expr_without_location (access
->expr
), d
;
2027 DECL_NAME (repl
) = get_identifier (pretty_name
);
2028 obstack_free (&name_obstack
, pretty_name
);
2030 /* Get rid of any SSA_NAMEs embedded in debug_expr,
2031 as DECL_DEBUG_EXPR isn't considered when looking for still
2032 used SSA_NAMEs and thus they could be freed. All debug info
2033 generation cares is whether something is constant or variable
2034 and that get_ref_base_and_extent works properly on the
2035 expression. It cannot handle accesses at a non-constant offset
2036 though, so just give up in those cases. */
2037 for (d
= debug_expr
;
2038 !fail
&& (handled_component_p (d
) || TREE_CODE (d
) == MEM_REF
);
2039 d
= TREE_OPERAND (d
, 0))
2040 switch (TREE_CODE (d
))
2043 case ARRAY_RANGE_REF
:
2044 if (TREE_OPERAND (d
, 1)
2045 && TREE_CODE (TREE_OPERAND (d
, 1)) != INTEGER_CST
)
2047 if (TREE_OPERAND (d
, 3)
2048 && TREE_CODE (TREE_OPERAND (d
, 3)) != INTEGER_CST
)
2052 if (TREE_OPERAND (d
, 2)
2053 && TREE_CODE (TREE_OPERAND (d
, 2)) != INTEGER_CST
)
2057 if (TREE_CODE (TREE_OPERAND (d
, 0)) != ADDR_EXPR
)
2060 d
= TREE_OPERAND (d
, 0);
2067 SET_DECL_DEBUG_EXPR (repl
, debug_expr
);
2068 DECL_HAS_DEBUG_EXPR_P (repl
) = 1;
2070 if (access
->grp_no_warning
)
2071 TREE_NO_WARNING (repl
) = 1;
2073 TREE_NO_WARNING (repl
) = TREE_NO_WARNING (access
->base
);
2076 TREE_NO_WARNING (repl
) = 1;
2080 if (access
->grp_to_be_debug_replaced
)
2082 fprintf (dump_file
, "Created a debug-only replacement for ");
2083 print_generic_expr (dump_file
, access
->base
, 0);
2084 fprintf (dump_file
, " offset: %u, size: %u\n",
2085 (unsigned) access
->offset
, (unsigned) access
->size
);
2089 fprintf (dump_file
, "Created a replacement for ");
2090 print_generic_expr (dump_file
, access
->base
, 0);
2091 fprintf (dump_file
, " offset: %u, size: %u: ",
2092 (unsigned) access
->offset
, (unsigned) access
->size
);
2093 print_generic_expr (dump_file
, repl
, 0);
2094 fprintf (dump_file
, "\n");
2097 sra_stats
.replacements
++;
2102 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
2105 get_access_replacement (struct access
*access
)
2107 gcc_checking_assert (access
->replacement_decl
);
2108 return access
->replacement_decl
;
2112 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
2113 linked list along the way. Stop when *ACCESS is NULL or the access pointed
2114 to it is not "within" the root. Return false iff some accesses partially
2118 build_access_subtree (struct access
**access
)
2120 struct access
*root
= *access
, *last_child
= NULL
;
2121 HOST_WIDE_INT limit
= root
->offset
+ root
->size
;
2123 *access
= (*access
)->next_grp
;
2124 while (*access
&& (*access
)->offset
+ (*access
)->size
<= limit
)
2127 root
->first_child
= *access
;
2129 last_child
->next_sibling
= *access
;
2130 last_child
= *access
;
2132 if (!build_access_subtree (access
))
2136 if (*access
&& (*access
)->offset
< limit
)
2142 /* Build a tree of access representatives, ACCESS is the pointer to the first
2143 one, others are linked in a list by the next_grp field. Return false iff
2144 some accesses partially overlap. */
2147 build_access_trees (struct access
*access
)
2151 struct access
*root
= access
;
2153 if (!build_access_subtree (&access
))
2155 root
->next_grp
= access
;
2160 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2164 expr_with_var_bounded_array_refs_p (tree expr
)
2166 while (handled_component_p (expr
))
2168 if (TREE_CODE (expr
) == ARRAY_REF
2169 && !tree_fits_shwi_p (array_ref_low_bound (expr
)))
2171 expr
= TREE_OPERAND (expr
, 0);
2176 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2177 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2178 sorts of access flags appropriately along the way, notably always set
2179 grp_read and grp_assign_read according to MARK_READ and grp_write when
2182 Creating a replacement for a scalar access is considered beneficial if its
2183 grp_hint is set (this means we are either attempting total scalarization or
2184 there is more than one direct read access) or according to the following
2187 Access written to through a scalar type (once or more times)
2189 | Written to in an assignment statement
2191 | | Access read as scalar _once_
2193 | | | Read in an assignment statement
2195 | | | | Scalarize Comment
2196 -----------------------------------------------------------------------------
2197 0 0 0 0 No access for the scalar
2198 0 0 0 1 No access for the scalar
2199 0 0 1 0 No Single read - won't help
2200 0 0 1 1 No The same case
2201 0 1 0 0 No access for the scalar
2202 0 1 0 1 No access for the scalar
2203 0 1 1 0 Yes s = *g; return s.i;
2204 0 1 1 1 Yes The same case as above
2205 1 0 0 0 No Won't help
2206 1 0 0 1 Yes s.i = 1; *g = s;
2207 1 0 1 0 Yes s.i = 5; g = s.i;
2208 1 0 1 1 Yes The same case as above
2209 1 1 0 0 No Won't help.
2210 1 1 0 1 Yes s.i = 1; *g = s;
2211 1 1 1 0 Yes s = *g; return s.i;
2212 1 1 1 1 Yes Any of the above yeses */
2215 analyze_access_subtree (struct access
*root
, struct access
*parent
,
2216 bool allow_replacements
)
2218 struct access
*child
;
2219 HOST_WIDE_INT limit
= root
->offset
+ root
->size
;
2220 HOST_WIDE_INT covered_to
= root
->offset
;
2221 bool scalar
= is_gimple_reg_type (root
->type
);
2222 bool hole
= false, sth_created
= false;
2226 if (parent
->grp_read
)
2228 if (parent
->grp_assignment_read
)
2229 root
->grp_assignment_read
= 1;
2230 if (parent
->grp_write
)
2231 root
->grp_write
= 1;
2232 if (parent
->grp_assignment_write
)
2233 root
->grp_assignment_write
= 1;
2234 if (parent
->grp_total_scalarization
)
2235 root
->grp_total_scalarization
= 1;
2238 if (root
->grp_unscalarizable_region
)
2239 allow_replacements
= false;
2241 if (allow_replacements
&& expr_with_var_bounded_array_refs_p (root
->expr
))
2242 allow_replacements
= false;
2244 for (child
= root
->first_child
; child
; child
= child
->next_sibling
)
2246 hole
|= covered_to
< child
->offset
;
2247 sth_created
|= analyze_access_subtree (child
, root
,
2248 allow_replacements
&& !scalar
);
2250 root
->grp_unscalarized_data
|= child
->grp_unscalarized_data
;
2251 root
->grp_total_scalarization
&= child
->grp_total_scalarization
;
2252 if (child
->grp_covered
)
2253 covered_to
+= child
->size
;
2258 if (allow_replacements
&& scalar
&& !root
->first_child
2260 || ((root
->grp_scalar_read
|| root
->grp_assignment_read
)
2261 && (root
->grp_scalar_write
|| root
->grp_assignment_write
))))
2263 /* Always create access replacements that cover the whole access.
2264 For integral types this means the precision has to match.
2265 Avoid assumptions based on the integral type kind, too. */
2266 if (INTEGRAL_TYPE_P (root
->type
)
2267 && (TREE_CODE (root
->type
) != INTEGER_TYPE
2268 || TYPE_PRECISION (root
->type
) != root
->size
)
2269 /* But leave bitfield accesses alone. */
2270 && (TREE_CODE (root
->expr
) != COMPONENT_REF
2271 || !DECL_BIT_FIELD (TREE_OPERAND (root
->expr
, 1))))
2273 tree rt
= root
->type
;
2274 gcc_assert ((root
->offset
% BITS_PER_UNIT
) == 0
2275 && (root
->size
% BITS_PER_UNIT
) == 0);
2276 root
->type
= build_nonstandard_integer_type (root
->size
,
2277 TYPE_UNSIGNED (rt
));
2278 root
->expr
= build_ref_for_offset (UNKNOWN_LOCATION
,
2279 root
->base
, root
->offset
,
2280 root
->type
, NULL
, false);
2282 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2284 fprintf (dump_file
, "Changing the type of a replacement for ");
2285 print_generic_expr (dump_file
, root
->base
, 0);
2286 fprintf (dump_file
, " offset: %u, size: %u ",
2287 (unsigned) root
->offset
, (unsigned) root
->size
);
2288 fprintf (dump_file
, " to an integer.\n");
2292 root
->grp_to_be_replaced
= 1;
2293 root
->replacement_decl
= create_access_replacement (root
);
2299 if (allow_replacements
2300 && scalar
&& !root
->first_child
2301 && (root
->grp_scalar_write
|| root
->grp_assignment_write
)
2302 && !bitmap_bit_p (cannot_scalarize_away_bitmap
,
2303 DECL_UID (root
->base
)))
2305 gcc_checking_assert (!root
->grp_scalar_read
2306 && !root
->grp_assignment_read
);
2308 if (MAY_HAVE_DEBUG_STMTS
)
2310 root
->grp_to_be_debug_replaced
= 1;
2311 root
->replacement_decl
= create_access_replacement (root
);
2315 if (covered_to
< limit
)
2318 root
->grp_total_scalarization
= 0;
2321 if (!hole
|| root
->grp_total_scalarization
)
2322 root
->grp_covered
= 1;
2323 else if (root
->grp_write
|| TREE_CODE (root
->base
) == PARM_DECL
)
2324 root
->grp_unscalarized_data
= 1; /* not covered and written to */
2328 /* Analyze all access trees linked by next_grp by the means of
2329 analyze_access_subtree. */
2331 analyze_access_trees (struct access
*access
)
2337 if (analyze_access_subtree (access
, NULL
, true))
2339 access
= access
->next_grp
;
2345 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2346 SIZE would conflict with an already existing one. If exactly such a child
2347 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2350 child_would_conflict_in_lacc (struct access
*lacc
, HOST_WIDE_INT norm_offset
,
2351 HOST_WIDE_INT size
, struct access
**exact_match
)
2353 struct access
*child
;
2355 for (child
= lacc
->first_child
; child
; child
= child
->next_sibling
)
2357 if (child
->offset
== norm_offset
&& child
->size
== size
)
2359 *exact_match
= child
;
2363 if (child
->offset
< norm_offset
+ size
2364 && child
->offset
+ child
->size
> norm_offset
)
2371 /* Create a new child access of PARENT, with all properties just like MODEL
2372 except for its offset and with its grp_write false and grp_read true.
2373 Return the new access or NULL if it cannot be created. Note that this access
2374 is created long after all splicing and sorting, it's not located in any
2375 access vector and is automatically a representative of its group. */
2377 static struct access
*
2378 create_artificial_child_access (struct access
*parent
, struct access
*model
,
2379 HOST_WIDE_INT new_offset
)
2381 struct access
*access
;
2382 struct access
**child
;
2383 tree expr
= parent
->base
;
2385 gcc_assert (!model
->grp_unscalarizable_region
);
2387 access
= (struct access
*) pool_alloc (access_pool
);
2388 memset (access
, 0, sizeof (struct access
));
2389 if (!build_user_friendly_ref_for_offset (&expr
, TREE_TYPE (expr
), new_offset
,
2392 access
->grp_no_warning
= true;
2393 expr
= build_ref_for_model (EXPR_LOCATION (parent
->base
), parent
->base
,
2394 new_offset
, model
, NULL
, false);
2397 access
->base
= parent
->base
;
2398 access
->expr
= expr
;
2399 access
->offset
= new_offset
;
2400 access
->size
= model
->size
;
2401 access
->type
= model
->type
;
2402 access
->grp_write
= true;
2403 access
->grp_read
= false;
2405 child
= &parent
->first_child
;
2406 while (*child
&& (*child
)->offset
< new_offset
)
2407 child
= &(*child
)->next_sibling
;
2409 access
->next_sibling
= *child
;
2416 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2417 true if any new subaccess was created. Additionally, if RACC is a scalar
2418 access but LACC is not, change the type of the latter, if possible. */
2421 propagate_subaccesses_across_link (struct access
*lacc
, struct access
*racc
)
2423 struct access
*rchild
;
2424 HOST_WIDE_INT norm_delta
= lacc
->offset
- racc
->offset
;
2427 if (is_gimple_reg_type (lacc
->type
)
2428 || lacc
->grp_unscalarizable_region
2429 || racc
->grp_unscalarizable_region
)
2432 if (is_gimple_reg_type (racc
->type
))
2434 if (!lacc
->first_child
&& !racc
->first_child
)
2436 tree t
= lacc
->base
;
2438 lacc
->type
= racc
->type
;
2439 if (build_user_friendly_ref_for_offset (&t
, TREE_TYPE (t
),
2440 lacc
->offset
, racc
->type
))
2444 lacc
->expr
= build_ref_for_model (EXPR_LOCATION (lacc
->base
),
2445 lacc
->base
, lacc
->offset
,
2447 lacc
->grp_no_warning
= true;
2453 for (rchild
= racc
->first_child
; rchild
; rchild
= rchild
->next_sibling
)
2455 struct access
*new_acc
= NULL
;
2456 HOST_WIDE_INT norm_offset
= rchild
->offset
+ norm_delta
;
2458 if (rchild
->grp_unscalarizable_region
)
2461 if (child_would_conflict_in_lacc (lacc
, norm_offset
, rchild
->size
,
2466 rchild
->grp_hint
= 1;
2467 new_acc
->grp_hint
|= new_acc
->grp_read
;
2468 if (rchild
->first_child
)
2469 ret
|= propagate_subaccesses_across_link (new_acc
, rchild
);
2474 rchild
->grp_hint
= 1;
2475 new_acc
= create_artificial_child_access (lacc
, rchild
, norm_offset
);
2479 if (racc
->first_child
)
2480 propagate_subaccesses_across_link (new_acc
, rchild
);
2487 /* Propagate all subaccesses across assignment links. */
2490 propagate_all_subaccesses (void)
2492 while (work_queue_head
)
2494 struct access
*racc
= pop_access_from_work_queue ();
2495 struct assign_link
*link
;
2497 gcc_assert (racc
->first_link
);
2499 for (link
= racc
->first_link
; link
; link
= link
->next
)
2501 struct access
*lacc
= link
->lacc
;
2503 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (lacc
->base
)))
2505 lacc
= lacc
->group_representative
;
2506 if (propagate_subaccesses_across_link (lacc
, racc
)
2507 && lacc
->first_link
)
2508 add_access_to_work_queue (lacc
);
2513 /* Go through all accesses collected throughout the (intraprocedural) analysis
2514 stage, exclude overlapping ones, identify representatives and build trees
2515 out of them, making decisions about scalarization on the way. Return true
2516 iff there are any to-be-scalarized variables after this stage. */
2519 analyze_all_variable_accesses (void)
2522 bitmap tmp
= BITMAP_ALLOC (NULL
);
2524 unsigned i
, max_total_scalarization_size
;
2526 max_total_scalarization_size
= UNITS_PER_WORD
* BITS_PER_UNIT
2527 * MOVE_RATIO (optimize_function_for_speed_p (cfun
));
2529 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap
, 0, i
, bi
)
2530 if (bitmap_bit_p (should_scalarize_away_bitmap
, i
)
2531 && !bitmap_bit_p (cannot_scalarize_away_bitmap
, i
))
2533 tree var
= candidate (i
);
2535 if (TREE_CODE (var
) == VAR_DECL
2536 && type_consists_of_records_p (TREE_TYPE (var
)))
2538 if (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (var
)))
2539 <= max_total_scalarization_size
)
2541 completely_scalarize_var (var
);
2542 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2544 fprintf (dump_file
, "Will attempt to totally scalarize ");
2545 print_generic_expr (dump_file
, var
, 0);
2546 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (var
));
2549 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2551 fprintf (dump_file
, "Too big to totally scalarize: ");
2552 print_generic_expr (dump_file
, var
, 0);
2553 fprintf (dump_file
, " (UID: %u)\n", DECL_UID (var
));
2558 bitmap_copy (tmp
, candidate_bitmap
);
2559 EXECUTE_IF_SET_IN_BITMAP (tmp
, 0, i
, bi
)
2561 tree var
= candidate (i
);
2562 struct access
*access
;
2564 access
= sort_and_splice_var_accesses (var
);
2565 if (!access
|| !build_access_trees (access
))
2566 disqualify_candidate (var
,
2567 "No or inhibitingly overlapping accesses.");
2570 propagate_all_subaccesses ();
2572 bitmap_copy (tmp
, candidate_bitmap
);
2573 EXECUTE_IF_SET_IN_BITMAP (tmp
, 0, i
, bi
)
2575 tree var
= candidate (i
);
2576 struct access
*access
= get_first_repr_for_decl (var
);
2578 if (analyze_access_trees (access
))
2581 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2583 fprintf (dump_file
, "\nAccess trees for ");
2584 print_generic_expr (dump_file
, var
, 0);
2585 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (var
));
2586 dump_access_tree (dump_file
, access
);
2587 fprintf (dump_file
, "\n");
2591 disqualify_candidate (var
, "No scalar replacements to be created.");
2598 statistics_counter_event (cfun
, "Scalarized aggregates", res
);
2605 /* Generate statements copying scalar replacements of accesses within a subtree
2606 into or out of AGG. ACCESS, all its children, siblings and their children
2607 are to be processed. AGG is an aggregate type expression (can be a
2608 declaration but does not have to be, it can for example also be a mem_ref or
2609 a series of handled components). TOP_OFFSET is the offset of the processed
2610 subtree which has to be subtracted from offsets of individual accesses to
2611 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2612 replacements in the interval <start_offset, start_offset + chunk_size>,
2613 otherwise copy all. GSI is a statement iterator used to place the new
2614 statements. WRITE should be true when the statements should write from AGG
2615 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2616 statements will be added after the current statement in GSI, they will be
2617 added before the statement otherwise. */
2620 generate_subtree_copies (struct access
*access
, tree agg
,
2621 HOST_WIDE_INT top_offset
,
2622 HOST_WIDE_INT start_offset
, HOST_WIDE_INT chunk_size
,
2623 gimple_stmt_iterator
*gsi
, bool write
,
2624 bool insert_after
, location_t loc
)
2628 if (chunk_size
&& access
->offset
>= start_offset
+ chunk_size
)
2631 if (access
->grp_to_be_replaced
2633 || access
->offset
+ access
->size
> start_offset
))
2635 tree expr
, repl
= get_access_replacement (access
);
2638 expr
= build_ref_for_model (loc
, agg
, access
->offset
- top_offset
,
2639 access
, gsi
, insert_after
);
2643 if (access
->grp_partial_lhs
)
2644 expr
= force_gimple_operand_gsi (gsi
, expr
, true, NULL_TREE
,
2646 insert_after
? GSI_NEW_STMT
2648 stmt
= gimple_build_assign (repl
, expr
);
2652 TREE_NO_WARNING (repl
) = 1;
2653 if (access
->grp_partial_lhs
)
2654 repl
= force_gimple_operand_gsi (gsi
, repl
, true, NULL_TREE
,
2656 insert_after
? GSI_NEW_STMT
2658 stmt
= gimple_build_assign (expr
, repl
);
2660 gimple_set_location (stmt
, loc
);
2663 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2665 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2667 sra_stats
.subtree_copies
++;
2670 && access
->grp_to_be_debug_replaced
2672 || access
->offset
+ access
->size
> start_offset
))
2675 tree drhs
= build_debug_ref_for_model (loc
, agg
,
2676 access
->offset
- top_offset
,
2678 ds
= gimple_build_debug_bind (get_access_replacement (access
),
2679 drhs
, gsi_stmt (*gsi
));
2681 gsi_insert_after (gsi
, ds
, GSI_NEW_STMT
);
2683 gsi_insert_before (gsi
, ds
, GSI_SAME_STMT
);
2686 if (access
->first_child
)
2687 generate_subtree_copies (access
->first_child
, agg
, top_offset
,
2688 start_offset
, chunk_size
, gsi
,
2689 write
, insert_after
, loc
);
2691 access
= access
->next_sibling
;
2696 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2697 the root of the subtree to be processed. GSI is the statement iterator used
2698 for inserting statements which are added after the current statement if
2699 INSERT_AFTER is true or before it otherwise. */
2702 init_subtree_with_zero (struct access
*access
, gimple_stmt_iterator
*gsi
,
2703 bool insert_after
, location_t loc
)
2706 struct access
*child
;
2708 if (access
->grp_to_be_replaced
)
2712 stmt
= gimple_build_assign (get_access_replacement (access
),
2713 build_zero_cst (access
->type
));
2715 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2717 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2719 gimple_set_location (stmt
, loc
);
2721 else if (access
->grp_to_be_debug_replaced
)
2723 gimple ds
= gimple_build_debug_bind (get_access_replacement (access
),
2724 build_zero_cst (access
->type
),
2727 gsi_insert_after (gsi
, ds
, GSI_NEW_STMT
);
2729 gsi_insert_before (gsi
, ds
, GSI_SAME_STMT
);
2732 for (child
= access
->first_child
; child
; child
= child
->next_sibling
)
2733 init_subtree_with_zero (child
, gsi
, insert_after
, loc
);
2736 /* Search for an access representative for the given expression EXPR and
2737 return it or NULL if it cannot be found. */
2739 static struct access
*
2740 get_access_for_expr (tree expr
)
2742 HOST_WIDE_INT offset
, size
, max_size
;
2745 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2746 a different size than the size of its argument and we need the latter
2748 if (TREE_CODE (expr
) == VIEW_CONVERT_EXPR
)
2749 expr
= TREE_OPERAND (expr
, 0);
2751 base
= get_ref_base_and_extent (expr
, &offset
, &size
, &max_size
);
2752 if (max_size
== -1 || !DECL_P (base
))
2755 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (base
)))
2758 return get_var_base_offset_size_access (base
, offset
, max_size
);
2761 /* Replace the expression EXPR with a scalar replacement if there is one and
2762 generate other statements to do type conversion or subtree copying if
2763 necessary. GSI is used to place newly created statements, WRITE is true if
2764 the expression is being written to (it is on a LHS of a statement or output
2765 in an assembly statement). */
2768 sra_modify_expr (tree
*expr
, gimple_stmt_iterator
*gsi
, bool write
)
2771 struct access
*access
;
2774 if (TREE_CODE (*expr
) == BIT_FIELD_REF
)
2777 expr
= &TREE_OPERAND (*expr
, 0);
2782 if (TREE_CODE (*expr
) == REALPART_EXPR
|| TREE_CODE (*expr
) == IMAGPART_EXPR
)
2783 expr
= &TREE_OPERAND (*expr
, 0);
2784 access
= get_access_for_expr (*expr
);
2787 type
= TREE_TYPE (*expr
);
2789 loc
= gimple_location (gsi_stmt (*gsi
));
2790 gimple_stmt_iterator alt_gsi
= gsi_none ();
2791 if (write
&& stmt_ends_bb_p (gsi_stmt (*gsi
)))
2793 alt_gsi
= gsi_start_edge (single_non_eh_succ (gsi_bb (*gsi
)));
2797 if (access
->grp_to_be_replaced
)
2799 tree repl
= get_access_replacement (access
);
2800 /* If we replace a non-register typed access simply use the original
2801 access expression to extract the scalar component afterwards.
2802 This happens if scalarizing a function return value or parameter
2803 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2804 gcc.c-torture/compile/20011217-1.c.
2806 We also want to use this when accessing a complex or vector which can
2807 be accessed as a different type too, potentially creating a need for
2808 type conversion (see PR42196) and when scalarized unions are involved
2809 in assembler statements (see PR42398). */
2810 if (!useless_type_conversion_p (type
, access
->type
))
2814 ref
= build_ref_for_model (loc
, access
->base
, access
->offset
, access
,
2821 if (access
->grp_partial_lhs
)
2822 ref
= force_gimple_operand_gsi (gsi
, ref
, true, NULL_TREE
,
2823 false, GSI_NEW_STMT
);
2824 stmt
= gimple_build_assign (repl
, ref
);
2825 gimple_set_location (stmt
, loc
);
2826 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2832 if (access
->grp_partial_lhs
)
2833 repl
= force_gimple_operand_gsi (gsi
, repl
, true, NULL_TREE
,
2834 true, GSI_SAME_STMT
);
2835 stmt
= gimple_build_assign (ref
, repl
);
2836 gimple_set_location (stmt
, loc
);
2837 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2844 else if (write
&& access
->grp_to_be_debug_replaced
)
2846 gimple ds
= gimple_build_debug_bind (get_access_replacement (access
),
2849 gsi_insert_after (gsi
, ds
, GSI_NEW_STMT
);
2852 if (access
->first_child
)
2854 HOST_WIDE_INT start_offset
, chunk_size
;
2856 && tree_fits_uhwi_p (TREE_OPERAND (bfr
, 1))
2857 && tree_fits_uhwi_p (TREE_OPERAND (bfr
, 2)))
2859 chunk_size
= tree_to_uhwi (TREE_OPERAND (bfr
, 1));
2860 start_offset
= access
->offset
2861 + tree_to_uhwi (TREE_OPERAND (bfr
, 2));
2864 start_offset
= chunk_size
= 0;
2866 generate_subtree_copies (access
->first_child
, access
->base
, 0,
2867 start_offset
, chunk_size
, gsi
, write
, write
,
2873 /* Where scalar replacements of the RHS have been written to when a replacement
2874 of a LHS of an assigments cannot be direclty loaded from a replacement of
2876 enum unscalarized_data_handling
{ SRA_UDH_NONE
, /* Nothing done so far. */
2877 SRA_UDH_RIGHT
, /* Data flushed to the RHS. */
2878 SRA_UDH_LEFT
}; /* Data flushed to the LHS. */
2880 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2881 base aggregate if there are unscalarized data or directly to LHS of the
2882 statement that is pointed to by GSI otherwise. */
2884 static enum unscalarized_data_handling
2885 handle_unscalarized_data_in_subtree (struct access
*top_racc
,
2886 gimple_stmt_iterator
*gsi
)
2888 if (top_racc
->grp_unscalarized_data
)
2890 generate_subtree_copies (top_racc
->first_child
, top_racc
->base
, 0, 0, 0,
2892 gimple_location (gsi_stmt (*gsi
)));
2893 return SRA_UDH_RIGHT
;
2897 tree lhs
= gimple_assign_lhs (gsi_stmt (*gsi
));
2898 generate_subtree_copies (top_racc
->first_child
, lhs
, top_racc
->offset
,
2899 0, 0, gsi
, false, false,
2900 gimple_location (gsi_stmt (*gsi
)));
2901 return SRA_UDH_LEFT
;
2906 /* Try to generate statements to load all sub-replacements in an access subtree
2907 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2908 If that is not possible, refresh the TOP_RACC base aggregate and load the
2909 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2910 copied. NEW_GSI is stmt iterator used for statement insertions after the
2911 original assignment, OLD_GSI is used to insert statements before the
2912 assignment. *REFRESHED keeps the information whether we have needed to
2913 refresh replacements of the LHS and from which side of the assignments this
2917 load_assign_lhs_subreplacements (struct access
*lacc
, struct access
*top_racc
,
2918 HOST_WIDE_INT left_offset
,
2919 gimple_stmt_iterator
*old_gsi
,
2920 gimple_stmt_iterator
*new_gsi
,
2921 enum unscalarized_data_handling
*refreshed
)
2923 location_t loc
= gimple_location (gsi_stmt (*old_gsi
));
2924 for (lacc
= lacc
->first_child
; lacc
; lacc
= lacc
->next_sibling
)
2926 HOST_WIDE_INT offset
= lacc
->offset
- left_offset
+ top_racc
->offset
;
2928 if (lacc
->grp_to_be_replaced
)
2930 struct access
*racc
;
2934 racc
= find_access_in_subtree (top_racc
, offset
, lacc
->size
);
2935 if (racc
&& racc
->grp_to_be_replaced
)
2937 rhs
= get_access_replacement (racc
);
2938 if (!useless_type_conversion_p (lacc
->type
, racc
->type
))
2939 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
, lacc
->type
, rhs
);
2941 if (racc
->grp_partial_lhs
&& lacc
->grp_partial_lhs
)
2942 rhs
= force_gimple_operand_gsi (old_gsi
, rhs
, true, NULL_TREE
,
2943 true, GSI_SAME_STMT
);
2947 /* No suitable access on the right hand side, need to load from
2948 the aggregate. See if we have to update it first... */
2949 if (*refreshed
== SRA_UDH_NONE
)
2950 *refreshed
= handle_unscalarized_data_in_subtree (top_racc
,
2953 if (*refreshed
== SRA_UDH_LEFT
)
2954 rhs
= build_ref_for_model (loc
, lacc
->base
, lacc
->offset
, lacc
,
2957 rhs
= build_ref_for_model (loc
, top_racc
->base
, offset
, lacc
,
2959 if (lacc
->grp_partial_lhs
)
2960 rhs
= force_gimple_operand_gsi (new_gsi
, rhs
, true, NULL_TREE
,
2961 false, GSI_NEW_STMT
);
2964 stmt
= gimple_build_assign (get_access_replacement (lacc
), rhs
);
2965 gsi_insert_after (new_gsi
, stmt
, GSI_NEW_STMT
);
2966 gimple_set_location (stmt
, loc
);
2968 sra_stats
.subreplacements
++;
2972 if (*refreshed
== SRA_UDH_NONE
2973 && lacc
->grp_read
&& !lacc
->grp_covered
)
2974 *refreshed
= handle_unscalarized_data_in_subtree (top_racc
,
2976 if (lacc
&& lacc
->grp_to_be_debug_replaced
)
2980 struct access
*racc
= find_access_in_subtree (top_racc
, offset
,
2983 if (racc
&& racc
->grp_to_be_replaced
)
2985 if (racc
->grp_write
)
2986 drhs
= get_access_replacement (racc
);
2990 else if (*refreshed
== SRA_UDH_LEFT
)
2991 drhs
= build_debug_ref_for_model (loc
, lacc
->base
, lacc
->offset
,
2993 else if (*refreshed
== SRA_UDH_RIGHT
)
2994 drhs
= build_debug_ref_for_model (loc
, top_racc
->base
, offset
,
2999 && !useless_type_conversion_p (lacc
->type
, TREE_TYPE (drhs
)))
3000 drhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
,
3002 ds
= gimple_build_debug_bind (get_access_replacement (lacc
),
3003 drhs
, gsi_stmt (*old_gsi
));
3004 gsi_insert_after (new_gsi
, ds
, GSI_NEW_STMT
);
3008 if (lacc
->first_child
)
3009 load_assign_lhs_subreplacements (lacc
, top_racc
, left_offset
,
3010 old_gsi
, new_gsi
, refreshed
);
3014 /* Result code for SRA assignment modification. */
3015 enum assignment_mod_result
{ SRA_AM_NONE
, /* nothing done for the stmt */
3016 SRA_AM_MODIFIED
, /* stmt changed but not
3018 SRA_AM_REMOVED
}; /* stmt eliminated */
3020 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
3021 to the assignment and GSI is the statement iterator pointing at it. Returns
3022 the same values as sra_modify_assign. */
3024 static enum assignment_mod_result
3025 sra_modify_constructor_assign (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
3027 tree lhs
= gimple_assign_lhs (*stmt
);
3031 acc
= get_access_for_expr (lhs
);
3035 if (gimple_clobber_p (*stmt
))
3037 /* Remove clobbers of fully scalarized variables, otherwise
3039 if (acc
->grp_covered
)
3041 unlink_stmt_vdef (*stmt
);
3042 gsi_remove (gsi
, true);
3043 release_defs (*stmt
);
3044 return SRA_AM_REMOVED
;
3050 loc
= gimple_location (*stmt
);
3051 if (vec_safe_length (CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt
))) > 0)
3053 /* I have never seen this code path trigger but if it can happen the
3054 following should handle it gracefully. */
3055 if (access_has_children_p (acc
))
3056 generate_subtree_copies (acc
->first_child
, acc
->base
, 0, 0, 0, gsi
,
3058 return SRA_AM_MODIFIED
;
3061 if (acc
->grp_covered
)
3063 init_subtree_with_zero (acc
, gsi
, false, loc
);
3064 unlink_stmt_vdef (*stmt
);
3065 gsi_remove (gsi
, true);
3066 release_defs (*stmt
);
3067 return SRA_AM_REMOVED
;
3071 init_subtree_with_zero (acc
, gsi
, true, loc
);
3072 return SRA_AM_MODIFIED
;
3076 /* Create and return a new suitable default definition SSA_NAME for RACC which
3077 is an access describing an uninitialized part of an aggregate that is being
3081 get_repl_default_def_ssa_name (struct access
*racc
)
3083 gcc_checking_assert (!racc
->grp_to_be_replaced
3084 && !racc
->grp_to_be_debug_replaced
);
3085 if (!racc
->replacement_decl
)
3086 racc
->replacement_decl
= create_access_replacement (racc
);
3087 return get_or_create_ssa_default_def (cfun
, racc
->replacement_decl
);
3090 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
3091 bit-field field declaration somewhere in it. */
3094 contains_vce_or_bfcref_p (const_tree ref
)
3096 while (handled_component_p (ref
))
3098 if (TREE_CODE (ref
) == VIEW_CONVERT_EXPR
3099 || (TREE_CODE (ref
) == COMPONENT_REF
3100 && DECL_BIT_FIELD (TREE_OPERAND (ref
, 1))))
3102 ref
= TREE_OPERAND (ref
, 0);
3108 /* Examine both sides of the assignment statement pointed to by STMT, replace
3109 them with a scalare replacement if there is one and generate copying of
3110 replacements if scalarized aggregates have been used in the assignment. GSI
3111 is used to hold generated statements for type conversions and subtree
3114 static enum assignment_mod_result
3115 sra_modify_assign (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
3117 struct access
*lacc
, *racc
;
3119 bool modify_this_stmt
= false;
3120 bool force_gimple_rhs
= false;
3122 gimple_stmt_iterator orig_gsi
= *gsi
;
3124 if (!gimple_assign_single_p (*stmt
))
3126 lhs
= gimple_assign_lhs (*stmt
);
3127 rhs
= gimple_assign_rhs1 (*stmt
);
3129 if (TREE_CODE (rhs
) == CONSTRUCTOR
)
3130 return sra_modify_constructor_assign (stmt
, gsi
);
3132 if (TREE_CODE (rhs
) == REALPART_EXPR
|| TREE_CODE (lhs
) == REALPART_EXPR
3133 || TREE_CODE (rhs
) == IMAGPART_EXPR
|| TREE_CODE (lhs
) == IMAGPART_EXPR
3134 || TREE_CODE (rhs
) == BIT_FIELD_REF
|| TREE_CODE (lhs
) == BIT_FIELD_REF
)
3136 modify_this_stmt
= sra_modify_expr (gimple_assign_rhs1_ptr (*stmt
),
3138 modify_this_stmt
|= sra_modify_expr (gimple_assign_lhs_ptr (*stmt
),
3140 return modify_this_stmt
? SRA_AM_MODIFIED
: SRA_AM_NONE
;
3143 lacc
= get_access_for_expr (lhs
);
3144 racc
= get_access_for_expr (rhs
);
3148 loc
= gimple_location (*stmt
);
3149 if (lacc
&& lacc
->grp_to_be_replaced
)
3151 lhs
= get_access_replacement (lacc
);
3152 gimple_assign_set_lhs (*stmt
, lhs
);
3153 modify_this_stmt
= true;
3154 if (lacc
->grp_partial_lhs
)
3155 force_gimple_rhs
= true;
3159 if (racc
&& racc
->grp_to_be_replaced
)
3161 rhs
= get_access_replacement (racc
);
3162 modify_this_stmt
= true;
3163 if (racc
->grp_partial_lhs
)
3164 force_gimple_rhs
= true;
3168 && !racc
->grp_unscalarized_data
3169 && TREE_CODE (lhs
) == SSA_NAME
3170 && !access_has_replacements_p (racc
))
3172 rhs
= get_repl_default_def_ssa_name (racc
);
3173 modify_this_stmt
= true;
3177 if (modify_this_stmt
)
3179 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
3181 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
3182 ??? This should move to fold_stmt which we simply should
3183 call after building a VIEW_CONVERT_EXPR here. */
3184 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs
))
3185 && !contains_bitfld_component_ref_p (lhs
))
3187 lhs
= build_ref_for_model (loc
, lhs
, 0, racc
, gsi
, false);
3188 gimple_assign_set_lhs (*stmt
, lhs
);
3190 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs
))
3191 && !contains_vce_or_bfcref_p (rhs
))
3192 rhs
= build_ref_for_model (loc
, rhs
, 0, lacc
, gsi
, false);
3194 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
3196 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
, TREE_TYPE (lhs
),
3198 if (is_gimple_reg_type (TREE_TYPE (lhs
))
3199 && TREE_CODE (lhs
) != SSA_NAME
)
3200 force_gimple_rhs
= true;
3205 if (lacc
&& lacc
->grp_to_be_debug_replaced
)
3207 tree dlhs
= get_access_replacement (lacc
);
3208 tree drhs
= unshare_expr (rhs
);
3209 if (!useless_type_conversion_p (TREE_TYPE (dlhs
), TREE_TYPE (drhs
)))
3211 if (AGGREGATE_TYPE_P (TREE_TYPE (drhs
))
3212 && !contains_vce_or_bfcref_p (drhs
))
3213 drhs
= build_debug_ref_for_model (loc
, drhs
, 0, lacc
);
3215 && !useless_type_conversion_p (TREE_TYPE (dlhs
),
3217 drhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
,
3218 TREE_TYPE (dlhs
), drhs
);
3220 gimple ds
= gimple_build_debug_bind (dlhs
, drhs
, *stmt
);
3221 gsi_insert_before (gsi
, ds
, GSI_SAME_STMT
);
3224 /* From this point on, the function deals with assignments in between
3225 aggregates when at least one has scalar reductions of some of its
3226 components. There are three possible scenarios: Both the LHS and RHS have
3227 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
3229 In the first case, we would like to load the LHS components from RHS
3230 components whenever possible. If that is not possible, we would like to
3231 read it directly from the RHS (after updating it by storing in it its own
3232 components). If there are some necessary unscalarized data in the LHS,
3233 those will be loaded by the original assignment too. If neither of these
3234 cases happen, the original statement can be removed. Most of this is done
3235 by load_assign_lhs_subreplacements.
3237 In the second case, we would like to store all RHS scalarized components
3238 directly into LHS and if they cover the aggregate completely, remove the
3239 statement too. In the third case, we want the LHS components to be loaded
3240 directly from the RHS (DSE will remove the original statement if it
3243 This is a bit complex but manageable when types match and when unions do
3244 not cause confusion in a way that we cannot really load a component of LHS
3245 from the RHS or vice versa (the access representing this level can have
3246 subaccesses that are accessible only through a different union field at a
3247 higher level - different from the one used in the examined expression).
3250 Therefore, I specially handle a fourth case, happening when there is a
3251 specific type cast or it is impossible to locate a scalarized subaccess on
3252 the other side of the expression. If that happens, I simply "refresh" the
3253 RHS by storing in it is scalarized components leave the original statement
3254 there to do the copying and then load the scalar replacements of the LHS.
3255 This is what the first branch does. */
3257 if (modify_this_stmt
3258 || gimple_has_volatile_ops (*stmt
)
3259 || contains_vce_or_bfcref_p (rhs
)
3260 || contains_vce_or_bfcref_p (lhs
)
3261 || stmt_ends_bb_p (*stmt
))
3263 if (access_has_children_p (racc
))
3264 generate_subtree_copies (racc
->first_child
, racc
->base
, 0, 0, 0,
3265 gsi
, false, false, loc
);
3266 if (access_has_children_p (lacc
))
3268 gimple_stmt_iterator alt_gsi
= gsi_none ();
3269 if (stmt_ends_bb_p (*stmt
))
3271 alt_gsi
= gsi_start_edge (single_non_eh_succ (gsi_bb (*gsi
)));
3274 generate_subtree_copies (lacc
->first_child
, lacc
->base
, 0, 0, 0,
3275 gsi
, true, true, loc
);
3277 sra_stats
.separate_lhs_rhs_handling
++;
3279 /* This gimplification must be done after generate_subtree_copies,
3280 lest we insert the subtree copies in the middle of the gimplified
3282 if (force_gimple_rhs
)
3283 rhs
= force_gimple_operand_gsi (&orig_gsi
, rhs
, true, NULL_TREE
,
3284 true, GSI_SAME_STMT
);
3285 if (gimple_assign_rhs1 (*stmt
) != rhs
)
3287 modify_this_stmt
= true;
3288 gimple_assign_set_rhs_from_tree (&orig_gsi
, rhs
);
3289 gcc_assert (*stmt
== gsi_stmt (orig_gsi
));
3292 return modify_this_stmt
? SRA_AM_MODIFIED
: SRA_AM_NONE
;
3296 if (access_has_children_p (lacc
)
3297 && access_has_children_p (racc
)
3298 /* When an access represents an unscalarizable region, it usually
3299 represents accesses with variable offset and thus must not be used
3300 to generate new memory accesses. */
3301 && !lacc
->grp_unscalarizable_region
3302 && !racc
->grp_unscalarizable_region
)
3304 gimple_stmt_iterator orig_gsi
= *gsi
;
3305 enum unscalarized_data_handling refreshed
;
3307 if (lacc
->grp_read
&& !lacc
->grp_covered
)
3308 refreshed
= handle_unscalarized_data_in_subtree (racc
, gsi
);
3310 refreshed
= SRA_UDH_NONE
;
3312 load_assign_lhs_subreplacements (lacc
, racc
, lacc
->offset
,
3313 &orig_gsi
, gsi
, &refreshed
);
3314 if (refreshed
!= SRA_UDH_RIGHT
)
3317 unlink_stmt_vdef (*stmt
);
3318 gsi_remove (&orig_gsi
, true);
3319 release_defs (*stmt
);
3320 sra_stats
.deleted
++;
3321 return SRA_AM_REMOVED
;
3326 if (access_has_children_p (racc
)
3327 && !racc
->grp_unscalarized_data
)
3331 fprintf (dump_file
, "Removing load: ");
3332 print_gimple_stmt (dump_file
, *stmt
, 0, 0);
3334 generate_subtree_copies (racc
->first_child
, lhs
,
3335 racc
->offset
, 0, 0, gsi
,
3337 gcc_assert (*stmt
== gsi_stmt (*gsi
));
3338 unlink_stmt_vdef (*stmt
);
3339 gsi_remove (gsi
, true);
3340 release_defs (*stmt
);
3341 sra_stats
.deleted
++;
3342 return SRA_AM_REMOVED
;
3344 /* Restore the aggregate RHS from its components so the
3345 prevailing aggregate copy does the right thing. */
3346 if (access_has_children_p (racc
))
3347 generate_subtree_copies (racc
->first_child
, racc
->base
, 0, 0, 0,
3348 gsi
, false, false, loc
);
3349 /* Re-load the components of the aggregate copy destination.
3350 But use the RHS aggregate to load from to expose more
3351 optimization opportunities. */
3352 if (access_has_children_p (lacc
))
3353 generate_subtree_copies (lacc
->first_child
, rhs
, lacc
->offset
,
3354 0, 0, gsi
, true, true, loc
);
3361 /* Traverse the function body and all modifications as decided in
3362 analyze_all_variable_accesses. Return true iff the CFG has been
3366 sra_modify_function_body (void)
3368 bool cfg_changed
= false;
3371 FOR_EACH_BB_FN (bb
, cfun
)
3373 gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
3374 while (!gsi_end_p (gsi
))
3376 gimple stmt
= gsi_stmt (gsi
);
3377 enum assignment_mod_result assign_result
;
3378 bool modified
= false, deleted
= false;
3382 switch (gimple_code (stmt
))
3385 t
= gimple_return_retval_ptr (stmt
);
3386 if (*t
!= NULL_TREE
)
3387 modified
|= sra_modify_expr (t
, &gsi
, false);
3391 assign_result
= sra_modify_assign (&stmt
, &gsi
);
3392 modified
|= assign_result
== SRA_AM_MODIFIED
;
3393 deleted
= assign_result
== SRA_AM_REMOVED
;
3397 /* Operands must be processed before the lhs. */
3398 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
3400 t
= gimple_call_arg_ptr (stmt
, i
);
3401 modified
|= sra_modify_expr (t
, &gsi
, false);
3404 if (gimple_call_lhs (stmt
))
3406 t
= gimple_call_lhs_ptr (stmt
);
3407 modified
|= sra_modify_expr (t
, &gsi
, true);
3412 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
3414 t
= &TREE_VALUE (gimple_asm_input_op (stmt
, i
));
3415 modified
|= sra_modify_expr (t
, &gsi
, false);
3417 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
3419 t
= &TREE_VALUE (gimple_asm_output_op (stmt
, i
));
3420 modified
|= sra_modify_expr (t
, &gsi
, true);
3431 if (maybe_clean_eh_stmt (stmt
)
3432 && gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
3440 gsi_commit_edge_inserts ();
3444 /* Generate statements initializing scalar replacements of parts of function
3448 initialize_parameter_reductions (void)
3450 gimple_stmt_iterator gsi
;
3451 gimple_seq seq
= NULL
;
3454 gsi
= gsi_start (seq
);
3455 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3457 parm
= DECL_CHAIN (parm
))
3459 vec
<access_p
> *access_vec
;
3460 struct access
*access
;
3462 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
3464 access_vec
= get_base_access_vector (parm
);
3468 for (access
= (*access_vec
)[0];
3470 access
= access
->next_grp
)
3471 generate_subtree_copies (access
, parm
, 0, 0, 0, &gsi
, true, true,
3472 EXPR_LOCATION (parm
));
3475 seq
= gsi_seq (gsi
);
3477 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun
)), seq
);
3480 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3481 it reveals there are components of some aggregates to be scalarized, it runs
3482 the required transformations. */
3484 perform_intra_sra (void)
3489 if (!find_var_candidates ())
3492 if (!scan_function ())
3495 if (!analyze_all_variable_accesses ())
3498 if (sra_modify_function_body ())
3499 ret
= TODO_update_ssa
| TODO_cleanup_cfg
;
3501 ret
= TODO_update_ssa
;
3502 initialize_parameter_reductions ();
3504 statistics_counter_event (cfun
, "Scalar replacements created",
3505 sra_stats
.replacements
);
3506 statistics_counter_event (cfun
, "Modified expressions", sra_stats
.exprs
);
3507 statistics_counter_event (cfun
, "Subtree copy stmts",
3508 sra_stats
.subtree_copies
);
3509 statistics_counter_event (cfun
, "Subreplacement stmts",
3510 sra_stats
.subreplacements
);
3511 statistics_counter_event (cfun
, "Deleted stmts", sra_stats
.deleted
);
3512 statistics_counter_event (cfun
, "Separate LHS and RHS handling",
3513 sra_stats
.separate_lhs_rhs_handling
);
3516 sra_deinitialize ();
3520 /* Perform early intraprocedural SRA. */
3522 early_intra_sra (void)
3524 sra_mode
= SRA_MODE_EARLY_INTRA
;
3525 return perform_intra_sra ();
3528 /* Perform "late" intraprocedural SRA. */
3530 late_intra_sra (void)
3532 sra_mode
= SRA_MODE_INTRA
;
3533 return perform_intra_sra ();
3538 gate_intra_sra (void)
3540 return flag_tree_sra
!= 0 && dbg_cnt (tree_sra
);
3546 const pass_data pass_data_sra_early
=
3548 GIMPLE_PASS
, /* type */
3550 OPTGROUP_NONE
, /* optinfo_flags */
3551 true, /* has_execute */
3552 TV_TREE_SRA
, /* tv_id */
3553 ( PROP_cfg
| PROP_ssa
), /* properties_required */
3554 0, /* properties_provided */
3555 0, /* properties_destroyed */
3556 0, /* todo_flags_start */
3557 ( TODO_update_ssa
| TODO_verify_ssa
), /* todo_flags_finish */
3560 class pass_sra_early
: public gimple_opt_pass
3563 pass_sra_early (gcc::context
*ctxt
)
3564 : gimple_opt_pass (pass_data_sra_early
, ctxt
)
3567 /* opt_pass methods: */
3568 virtual bool gate (function
*) { return gate_intra_sra (); }
3569 virtual unsigned int execute (function
*) { return early_intra_sra (); }
3571 }; // class pass_sra_early
3576 make_pass_sra_early (gcc::context
*ctxt
)
3578 return new pass_sra_early (ctxt
);
3583 const pass_data pass_data_sra
=
3585 GIMPLE_PASS
, /* type */
3587 OPTGROUP_NONE
, /* optinfo_flags */
3588 true, /* has_execute */
3589 TV_TREE_SRA
, /* tv_id */
3590 ( PROP_cfg
| PROP_ssa
), /* properties_required */
3591 0, /* properties_provided */
3592 0, /* properties_destroyed */
3593 TODO_update_address_taken
, /* todo_flags_start */
3594 ( TODO_update_ssa
| TODO_verify_ssa
), /* todo_flags_finish */
3597 class pass_sra
: public gimple_opt_pass
3600 pass_sra (gcc::context
*ctxt
)
3601 : gimple_opt_pass (pass_data_sra
, ctxt
)
3604 /* opt_pass methods: */
3605 virtual bool gate (function
*) { return gate_intra_sra (); }
3606 virtual unsigned int execute (function
*) { return late_intra_sra (); }
3608 }; // class pass_sra
3613 make_pass_sra (gcc::context
*ctxt
)
3615 return new pass_sra (ctxt
);
3619 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3623 is_unused_scalar_param (tree parm
)
3626 return (is_gimple_reg (parm
)
3627 && (!(name
= ssa_default_def (cfun
, parm
))
3628 || has_zero_uses (name
)));
3631 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3632 examine whether there are any direct or otherwise infeasible ones. If so,
3633 return true, otherwise return false. PARM must be a gimple register with a
3634 non-NULL default definition. */
3637 ptr_parm_has_direct_uses (tree parm
)
3639 imm_use_iterator ui
;
3641 tree name
= ssa_default_def (cfun
, parm
);
3644 FOR_EACH_IMM_USE_STMT (stmt
, ui
, name
)
3647 use_operand_p use_p
;
3649 if (is_gimple_debug (stmt
))
3652 /* Valid uses include dereferences on the lhs and the rhs. */
3653 if (gimple_has_lhs (stmt
))
3655 tree lhs
= gimple_get_lhs (stmt
);
3656 while (handled_component_p (lhs
))
3657 lhs
= TREE_OPERAND (lhs
, 0);
3658 if (TREE_CODE (lhs
) == MEM_REF
3659 && TREE_OPERAND (lhs
, 0) == name
3660 && integer_zerop (TREE_OPERAND (lhs
, 1))
3661 && types_compatible_p (TREE_TYPE (lhs
),
3662 TREE_TYPE (TREE_TYPE (name
)))
3663 && !TREE_THIS_VOLATILE (lhs
))
3666 if (gimple_assign_single_p (stmt
))
3668 tree rhs
= gimple_assign_rhs1 (stmt
);
3669 while (handled_component_p (rhs
))
3670 rhs
= TREE_OPERAND (rhs
, 0);
3671 if (TREE_CODE (rhs
) == MEM_REF
3672 && TREE_OPERAND (rhs
, 0) == name
3673 && integer_zerop (TREE_OPERAND (rhs
, 1))
3674 && types_compatible_p (TREE_TYPE (rhs
),
3675 TREE_TYPE (TREE_TYPE (name
)))
3676 && !TREE_THIS_VOLATILE (rhs
))
3679 else if (is_gimple_call (stmt
))
3682 for (i
= 0; i
< gimple_call_num_args (stmt
); ++i
)
3684 tree arg
= gimple_call_arg (stmt
, i
);
3685 while (handled_component_p (arg
))
3686 arg
= TREE_OPERAND (arg
, 0);
3687 if (TREE_CODE (arg
) == MEM_REF
3688 && TREE_OPERAND (arg
, 0) == name
3689 && integer_zerop (TREE_OPERAND (arg
, 1))
3690 && types_compatible_p (TREE_TYPE (arg
),
3691 TREE_TYPE (TREE_TYPE (name
)))
3692 && !TREE_THIS_VOLATILE (arg
))
3697 /* If the number of valid uses does not match the number of
3698 uses in this stmt there is an unhandled use. */
3699 FOR_EACH_IMM_USE_ON_STMT (use_p
, ui
)
3706 BREAK_FROM_IMM_USE_STMT (ui
);
3712 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3713 them in candidate_bitmap. Note that these do not necessarily include
3714 parameter which are unused and thus can be removed. Return true iff any
3715 such candidate has been found. */
3718 find_param_candidates (void)
3725 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3727 parm
= DECL_CHAIN (parm
))
3729 tree type
= TREE_TYPE (parm
);
3734 if (TREE_THIS_VOLATILE (parm
)
3735 || TREE_ADDRESSABLE (parm
)
3736 || (!is_gimple_reg_type (type
) && is_va_list_type (type
)))
3739 if (is_unused_scalar_param (parm
))
3745 if (POINTER_TYPE_P (type
))
3747 type
= TREE_TYPE (type
);
3749 if (TREE_CODE (type
) == FUNCTION_TYPE
3750 || TYPE_VOLATILE (type
)
3751 || (TREE_CODE (type
) == ARRAY_TYPE
3752 && TYPE_NONALIASED_COMPONENT (type
))
3753 || !is_gimple_reg (parm
)
3754 || is_va_list_type (type
)
3755 || ptr_parm_has_direct_uses (parm
))
3758 else if (!AGGREGATE_TYPE_P (type
))
3761 if (!COMPLETE_TYPE_P (type
)
3762 || !tree_fits_uhwi_p (TYPE_SIZE (type
))
3763 || tree_to_uhwi (TYPE_SIZE (type
)) == 0
3764 || (AGGREGATE_TYPE_P (type
)
3765 && type_internals_preclude_sra_p (type
, &msg
)))
3768 bitmap_set_bit (candidate_bitmap
, DECL_UID (parm
));
3769 slot
= candidates
.find_slot_with_hash (parm
, DECL_UID (parm
), INSERT
);
3773 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3775 fprintf (dump_file
, "Candidate (%d): ", DECL_UID (parm
));
3776 print_generic_expr (dump_file
, parm
, 0);
3777 fprintf (dump_file
, "\n");
3781 func_param_count
= count
;
3785 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3789 mark_maybe_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
3792 struct access
*repr
= (struct access
*) data
;
3794 repr
->grp_maybe_modified
= 1;
3798 /* Analyze what representatives (in linked lists accessible from
3799 REPRESENTATIVES) can be modified by side effects of statements in the
3800 current function. */
3803 analyze_modified_params (vec
<access_p
> representatives
)
3807 for (i
= 0; i
< func_param_count
; i
++)
3809 struct access
*repr
;
3811 for (repr
= representatives
[i
];
3813 repr
= repr
->next_grp
)
3815 struct access
*access
;
3819 if (no_accesses_p (repr
))
3821 if (!POINTER_TYPE_P (TREE_TYPE (repr
->base
))
3822 || repr
->grp_maybe_modified
)
3825 ao_ref_init (&ar
, repr
->expr
);
3826 visited
= BITMAP_ALLOC (NULL
);
3827 for (access
= repr
; access
; access
= access
->next_sibling
)
3829 /* All accesses are read ones, otherwise grp_maybe_modified would
3830 be trivially set. */
3831 walk_aliased_vdefs (&ar
, gimple_vuse (access
->stmt
),
3832 mark_maybe_modified
, repr
, &visited
);
3833 if (repr
->grp_maybe_modified
)
3836 BITMAP_FREE (visited
);
3841 /* Propagate distances in bb_dereferences in the opposite direction than the
3842 control flow edges, in each step storing the maximum of the current value
3843 and the minimum of all successors. These steps are repeated until the table
3844 stabilizes. Note that BBs which might terminate the functions (according to
3845 final_bbs bitmap) never updated in this way. */
3848 propagate_dereference_distances (void)
3852 auto_vec
<basic_block
> queue (last_basic_block_for_fn (cfun
));
3853 queue
.quick_push (ENTRY_BLOCK_PTR_FOR_FN (cfun
));
3854 FOR_EACH_BB_FN (bb
, cfun
)
3856 queue
.quick_push (bb
);
3860 while (!queue
.is_empty ())
3864 bool change
= false;
3870 if (bitmap_bit_p (final_bbs
, bb
->index
))
3873 for (i
= 0; i
< func_param_count
; i
++)
3875 int idx
= bb
->index
* func_param_count
+ i
;
3877 HOST_WIDE_INT inh
= 0;
3879 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3881 int succ_idx
= e
->dest
->index
* func_param_count
+ i
;
3883 if (e
->src
== EXIT_BLOCK_PTR_FOR_FN (cfun
))
3889 inh
= bb_dereferences
[succ_idx
];
3891 else if (bb_dereferences
[succ_idx
] < inh
)
3892 inh
= bb_dereferences
[succ_idx
];
3895 if (!first
&& bb_dereferences
[idx
] < inh
)
3897 bb_dereferences
[idx
] = inh
;
3902 if (change
&& !bitmap_bit_p (final_bbs
, bb
->index
))
3903 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3908 e
->src
->aux
= e
->src
;
3909 queue
.quick_push (e
->src
);
3914 /* Dump a dereferences TABLE with heading STR to file F. */
3917 dump_dereferences_table (FILE *f
, const char *str
, HOST_WIDE_INT
*table
)
3921 fprintf (dump_file
, str
);
3922 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR_FOR_FN (cfun
),
3923 EXIT_BLOCK_PTR_FOR_FN (cfun
), next_bb
)
3925 fprintf (f
, "%4i %i ", bb
->index
, bitmap_bit_p (final_bbs
, bb
->index
));
3926 if (bb
!= EXIT_BLOCK_PTR_FOR_FN (cfun
))
3929 for (i
= 0; i
< func_param_count
; i
++)
3931 int idx
= bb
->index
* func_param_count
+ i
;
3932 fprintf (f
, " %4" HOST_WIDE_INT_PRINT
"d", table
[idx
]);
3937 fprintf (dump_file
, "\n");
3940 /* Determine what (parts of) parameters passed by reference that are not
3941 assigned to are not certainly dereferenced in this function and thus the
3942 dereferencing cannot be safely moved to the caller without potentially
3943 introducing a segfault. Mark such REPRESENTATIVES as
3944 grp_not_necessarilly_dereferenced.
3946 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3947 part is calculated rather than simple booleans are calculated for each
3948 pointer parameter to handle cases when only a fraction of the whole
3949 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3952 The maximum dereference distances for each pointer parameter and BB are
3953 already stored in bb_dereference. This routine simply propagates these
3954 values upwards by propagate_dereference_distances and then compares the
3955 distances of individual parameters in the ENTRY BB to the equivalent
3956 distances of each representative of a (fraction of a) parameter. */
3959 analyze_caller_dereference_legality (vec
<access_p
> representatives
)
3963 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3964 dump_dereferences_table (dump_file
,
3965 "Dereference table before propagation:\n",
3968 propagate_dereference_distances ();
3970 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3971 dump_dereferences_table (dump_file
,
3972 "Dereference table after propagation:\n",
3975 for (i
= 0; i
< func_param_count
; i
++)
3977 struct access
*repr
= representatives
[i
];
3978 int idx
= ENTRY_BLOCK_PTR_FOR_FN (cfun
)->index
* func_param_count
+ i
;
3980 if (!repr
|| no_accesses_p (repr
))
3985 if ((repr
->offset
+ repr
->size
) > bb_dereferences
[idx
])
3986 repr
->grp_not_necessarilly_dereferenced
= 1;
3987 repr
= repr
->next_grp
;
3993 /* Return the representative access for the parameter declaration PARM if it is
3994 a scalar passed by reference which is not written to and the pointer value
3995 is not used directly. Thus, if it is legal to dereference it in the caller
3996 and we can rule out modifications through aliases, such parameter should be
3997 turned into one passed by value. Return NULL otherwise. */
3999 static struct access
*
4000 unmodified_by_ref_scalar_representative (tree parm
)
4002 int i
, access_count
;
4003 struct access
*repr
;
4004 vec
<access_p
> *access_vec
;
4006 access_vec
= get_base_access_vector (parm
);
4007 gcc_assert (access_vec
);
4008 repr
= (*access_vec
)[0];
4011 repr
->group_representative
= repr
;
4013 access_count
= access_vec
->length ();
4014 for (i
= 1; i
< access_count
; i
++)
4016 struct access
*access
= (*access_vec
)[i
];
4019 access
->group_representative
= repr
;
4020 access
->next_sibling
= repr
->next_sibling
;
4021 repr
->next_sibling
= access
;
4025 repr
->grp_scalar_ptr
= 1;
4029 /* Return true iff this ACCESS precludes IPA-SRA of the parameter it is
4030 associated with. REQ_ALIGN is the minimum required alignment. */
4033 access_precludes_ipa_sra_p (struct access
*access
, unsigned int req_align
)
4035 unsigned int exp_align
;
4036 /* Avoid issues such as the second simple testcase in PR 42025. The problem
4037 is incompatible assign in a call statement (and possibly even in asm
4038 statements). This can be relaxed by using a new temporary but only for
4039 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
4040 intraprocedural SRA we deal with this by keeping the old aggregate around,
4041 something we cannot do in IPA-SRA.) */
4043 && (is_gimple_call (access
->stmt
)
4044 || gimple_code (access
->stmt
) == GIMPLE_ASM
))
4047 exp_align
= get_object_alignment (access
->expr
);
4048 if (exp_align
< req_align
)
4055 /* Sort collected accesses for parameter PARM, identify representatives for
4056 each accessed region and link them together. Return NULL if there are
4057 different but overlapping accesses, return the special ptr value meaning
4058 there are no accesses for this parameter if that is the case and return the
4059 first representative otherwise. Set *RO_GRP if there is a group of accesses
4060 with only read (i.e. no write) accesses. */
4062 static struct access
*
4063 splice_param_accesses (tree parm
, bool *ro_grp
)
4065 int i
, j
, access_count
, group_count
;
4066 int agg_size
, total_size
= 0;
4067 struct access
*access
, *res
, **prev_acc_ptr
= &res
;
4068 vec
<access_p
> *access_vec
;
4070 access_vec
= get_base_access_vector (parm
);
4072 return &no_accesses_representant
;
4073 access_count
= access_vec
->length ();
4075 access_vec
->qsort (compare_access_positions
);
4080 while (i
< access_count
)
4084 access
= (*access_vec
)[i
];
4085 modification
= access
->write
;
4086 if (access_precludes_ipa_sra_p (access
, TYPE_ALIGN (access
->type
)))
4088 a1_alias_type
= reference_alias_ptr_type (access
->expr
);
4090 /* Access is about to become group representative unless we find some
4091 nasty overlap which would preclude us from breaking this parameter
4095 while (j
< access_count
)
4097 struct access
*ac2
= (*access_vec
)[j
];
4098 if (ac2
->offset
!= access
->offset
)
4100 /* All or nothing law for parameters. */
4101 if (access
->offset
+ access
->size
> ac2
->offset
)
4106 else if (ac2
->size
!= access
->size
)
4109 if (access_precludes_ipa_sra_p (ac2
, TYPE_ALIGN (access
->type
))
4110 || (ac2
->type
!= access
->type
4111 && (TREE_ADDRESSABLE (ac2
->type
)
4112 || TREE_ADDRESSABLE (access
->type
)))
4113 || (reference_alias_ptr_type (ac2
->expr
) != a1_alias_type
))
4116 modification
|= ac2
->write
;
4117 ac2
->group_representative
= access
;
4118 ac2
->next_sibling
= access
->next_sibling
;
4119 access
->next_sibling
= ac2
;
4124 access
->grp_maybe_modified
= modification
;
4127 *prev_acc_ptr
= access
;
4128 prev_acc_ptr
= &access
->next_grp
;
4129 total_size
+= access
->size
;
4133 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
4134 agg_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm
))));
4136 agg_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (parm
)));
4137 if (total_size
>= agg_size
)
4140 gcc_assert (group_count
> 0);
4144 /* Decide whether parameters with representative accesses given by REPR should
4145 be reduced into components. */
4148 decide_one_param_reduction (struct access
*repr
)
4150 int total_size
, cur_parm_size
, agg_size
, new_param_count
, parm_size_limit
;
4155 cur_parm_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (parm
)));
4156 gcc_assert (cur_parm_size
> 0);
4158 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
4161 agg_size
= tree_to_uhwi (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm
))));
4166 agg_size
= cur_parm_size
;
4172 fprintf (dump_file
, "Evaluating PARAM group sizes for ");
4173 print_generic_expr (dump_file
, parm
, 0);
4174 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (parm
));
4175 for (acc
= repr
; acc
; acc
= acc
->next_grp
)
4176 dump_access (dump_file
, acc
, true);
4180 new_param_count
= 0;
4182 for (; repr
; repr
= repr
->next_grp
)
4184 gcc_assert (parm
== repr
->base
);
4186 /* Taking the address of a non-addressable field is verboten. */
4187 if (by_ref
&& repr
->non_addressable
)
4190 /* Do not decompose a non-BLKmode param in a way that would
4191 create BLKmode params. Especially for by-reference passing
4192 (thus, pointer-type param) this is hardly worthwhile. */
4193 if (DECL_MODE (parm
) != BLKmode
4194 && TYPE_MODE (repr
->type
) == BLKmode
)
4197 if (!by_ref
|| (!repr
->grp_maybe_modified
4198 && !repr
->grp_not_necessarilly_dereferenced
))
4199 total_size
+= repr
->size
;
4201 total_size
+= cur_parm_size
;
4206 gcc_assert (new_param_count
> 0);
4208 if (optimize_function_for_size_p (cfun
))
4209 parm_size_limit
= cur_parm_size
;
4211 parm_size_limit
= (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR
)
4214 if (total_size
< agg_size
4215 && total_size
<= parm_size_limit
)
4218 fprintf (dump_file
, " ....will be split into %i components\n",
4220 return new_param_count
;
4226 /* The order of the following enums is important, we need to do extra work for
4227 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
4228 enum ipa_splicing_result
{ NO_GOOD_ACCESS
, UNUSED_PARAMS
, BY_VAL_ACCESSES
,
4229 MODIF_BY_REF_ACCESSES
, UNMODIF_BY_REF_ACCESSES
};
4231 /* Identify representatives of all accesses to all candidate parameters for
4232 IPA-SRA. Return result based on what representatives have been found. */
4234 static enum ipa_splicing_result
4235 splice_all_param_accesses (vec
<access_p
> &representatives
)
4237 enum ipa_splicing_result result
= NO_GOOD_ACCESS
;
4239 struct access
*repr
;
4241 representatives
.create (func_param_count
);
4243 for (parm
= DECL_ARGUMENTS (current_function_decl
);
4245 parm
= DECL_CHAIN (parm
))
4247 if (is_unused_scalar_param (parm
))
4249 representatives
.quick_push (&no_accesses_representant
);
4250 if (result
== NO_GOOD_ACCESS
)
4251 result
= UNUSED_PARAMS
;
4253 else if (POINTER_TYPE_P (TREE_TYPE (parm
))
4254 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm
)))
4255 && bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
4257 repr
= unmodified_by_ref_scalar_representative (parm
);
4258 representatives
.quick_push (repr
);
4260 result
= UNMODIF_BY_REF_ACCESSES
;
4262 else if (bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
4264 bool ro_grp
= false;
4265 repr
= splice_param_accesses (parm
, &ro_grp
);
4266 representatives
.quick_push (repr
);
4268 if (repr
&& !no_accesses_p (repr
))
4270 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
4273 result
= UNMODIF_BY_REF_ACCESSES
;
4274 else if (result
< MODIF_BY_REF_ACCESSES
)
4275 result
= MODIF_BY_REF_ACCESSES
;
4277 else if (result
< BY_VAL_ACCESSES
)
4278 result
= BY_VAL_ACCESSES
;
4280 else if (no_accesses_p (repr
) && (result
== NO_GOOD_ACCESS
))
4281 result
= UNUSED_PARAMS
;
4284 representatives
.quick_push (NULL
);
4287 if (result
== NO_GOOD_ACCESS
)
4289 representatives
.release ();
4290 return NO_GOOD_ACCESS
;
4296 /* Return the index of BASE in PARMS. Abort if it is not found. */
4299 get_param_index (tree base
, vec
<tree
> parms
)
4303 len
= parms
.length ();
4304 for (i
= 0; i
< len
; i
++)
4305 if (parms
[i
] == base
)
4310 /* Convert the decisions made at the representative level into compact
4311 parameter adjustments. REPRESENTATIVES are pointers to first
4312 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4313 final number of adjustments. */
4315 static ipa_parm_adjustment_vec
4316 turn_representatives_into_adjustments (vec
<access_p
> representatives
,
4317 int adjustments_count
)
4320 ipa_parm_adjustment_vec adjustments
;
4324 gcc_assert (adjustments_count
> 0);
4325 parms
= ipa_get_vector_of_formal_parms (current_function_decl
);
4326 adjustments
.create (adjustments_count
);
4327 parm
= DECL_ARGUMENTS (current_function_decl
);
4328 for (i
= 0; i
< func_param_count
; i
++, parm
= DECL_CHAIN (parm
))
4330 struct access
*repr
= representatives
[i
];
4332 if (!repr
|| no_accesses_p (repr
))
4334 struct ipa_parm_adjustment adj
;
4336 memset (&adj
, 0, sizeof (adj
));
4337 adj
.base_index
= get_param_index (parm
, parms
);
4340 adj
.op
= IPA_PARM_OP_COPY
;
4342 adj
.op
= IPA_PARM_OP_REMOVE
;
4343 adj
.arg_prefix
= "ISRA";
4344 adjustments
.quick_push (adj
);
4348 struct ipa_parm_adjustment adj
;
4349 int index
= get_param_index (parm
, parms
);
4351 for (; repr
; repr
= repr
->next_grp
)
4353 memset (&adj
, 0, sizeof (adj
));
4354 gcc_assert (repr
->base
== parm
);
4355 adj
.base_index
= index
;
4356 adj
.base
= repr
->base
;
4357 adj
.type
= repr
->type
;
4358 adj
.alias_ptr_type
= reference_alias_ptr_type (repr
->expr
);
4359 adj
.offset
= repr
->offset
;
4360 adj
.by_ref
= (POINTER_TYPE_P (TREE_TYPE (repr
->base
))
4361 && (repr
->grp_maybe_modified
4362 || repr
->grp_not_necessarilly_dereferenced
));
4363 adj
.arg_prefix
= "ISRA";
4364 adjustments
.quick_push (adj
);
4372 /* Analyze the collected accesses and produce a plan what to do with the
4373 parameters in the form of adjustments, NULL meaning nothing. */
4375 static ipa_parm_adjustment_vec
4376 analyze_all_param_acesses (void)
4378 enum ipa_splicing_result repr_state
;
4379 bool proceed
= false;
4380 int i
, adjustments_count
= 0;
4381 vec
<access_p
> representatives
;
4382 ipa_parm_adjustment_vec adjustments
;
4384 repr_state
= splice_all_param_accesses (representatives
);
4385 if (repr_state
== NO_GOOD_ACCESS
)
4386 return ipa_parm_adjustment_vec ();
4388 /* If there are any parameters passed by reference which are not modified
4389 directly, we need to check whether they can be modified indirectly. */
4390 if (repr_state
== UNMODIF_BY_REF_ACCESSES
)
4392 analyze_caller_dereference_legality (representatives
);
4393 analyze_modified_params (representatives
);
4396 for (i
= 0; i
< func_param_count
; i
++)
4398 struct access
*repr
= representatives
[i
];
4400 if (repr
&& !no_accesses_p (repr
))
4402 if (repr
->grp_scalar_ptr
)
4404 adjustments_count
++;
4405 if (repr
->grp_not_necessarilly_dereferenced
4406 || repr
->grp_maybe_modified
)
4407 representatives
[i
] = NULL
;
4411 sra_stats
.scalar_by_ref_to_by_val
++;
4416 int new_components
= decide_one_param_reduction (repr
);
4418 if (new_components
== 0)
4420 representatives
[i
] = NULL
;
4421 adjustments_count
++;
4425 adjustments_count
+= new_components
;
4426 sra_stats
.aggregate_params_reduced
++;
4427 sra_stats
.param_reductions_created
+= new_components
;
4434 if (no_accesses_p (repr
))
4437 sra_stats
.deleted_unused_parameters
++;
4439 adjustments_count
++;
4443 if (!proceed
&& dump_file
)
4444 fprintf (dump_file
, "NOT proceeding to change params.\n");
4447 adjustments
= turn_representatives_into_adjustments (representatives
,
4450 adjustments
= ipa_parm_adjustment_vec ();
4452 representatives
.release ();
4456 /* If a parameter replacement identified by ADJ does not yet exist in the form
4457 of declaration, create it and record it, otherwise return the previously
4461 get_replaced_param_substitute (struct ipa_parm_adjustment
*adj
)
4464 if (!adj
->new_ssa_base
)
4466 char *pretty_name
= make_fancy_name (adj
->base
);
4468 repl
= create_tmp_reg (TREE_TYPE (adj
->base
), "ISR");
4469 DECL_NAME (repl
) = get_identifier (pretty_name
);
4470 obstack_free (&name_obstack
, pretty_name
);
4472 adj
->new_ssa_base
= repl
;
4475 repl
= adj
->new_ssa_base
;
4479 /* Find the first adjustment for a particular parameter BASE in a vector of
4480 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4483 static struct ipa_parm_adjustment
*
4484 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments
, tree base
)
4488 len
= adjustments
.length ();
4489 for (i
= 0; i
< len
; i
++)
4491 struct ipa_parm_adjustment
*adj
;
4493 adj
= &adjustments
[i
];
4494 if (adj
->op
!= IPA_PARM_OP_COPY
&& adj
->base
== base
)
4501 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4502 removed because its value is not used, replace the SSA_NAME with a one
4503 relating to a created VAR_DECL together all of its uses and return true.
4504 ADJUSTMENTS is a pointer to an adjustments vector. */
4507 replace_removed_params_ssa_names (gimple stmt
,
4508 ipa_parm_adjustment_vec adjustments
)
4510 struct ipa_parm_adjustment
*adj
;
4511 tree lhs
, decl
, repl
, name
;
4513 if (gimple_code (stmt
) == GIMPLE_PHI
)
4514 lhs
= gimple_phi_result (stmt
);
4515 else if (is_gimple_assign (stmt
))
4516 lhs
= gimple_assign_lhs (stmt
);
4517 else if (is_gimple_call (stmt
))
4518 lhs
= gimple_call_lhs (stmt
);
4522 if (TREE_CODE (lhs
) != SSA_NAME
)
4525 decl
= SSA_NAME_VAR (lhs
);
4526 if (decl
== NULL_TREE
4527 || TREE_CODE (decl
) != PARM_DECL
)
4530 adj
= get_adjustment_for_base (adjustments
, decl
);
4534 repl
= get_replaced_param_substitute (adj
);
4535 name
= make_ssa_name (repl
, stmt
);
4539 fprintf (dump_file
, "replacing an SSA name of a removed param ");
4540 print_generic_expr (dump_file
, lhs
, 0);
4541 fprintf (dump_file
, " with ");
4542 print_generic_expr (dump_file
, name
, 0);
4543 fprintf (dump_file
, "\n");
4546 if (is_gimple_assign (stmt
))
4547 gimple_assign_set_lhs (stmt
, name
);
4548 else if (is_gimple_call (stmt
))
4549 gimple_call_set_lhs (stmt
, name
);
4551 gimple_phi_set_result (stmt
, name
);
4553 replace_uses_by (lhs
, name
);
4554 release_ssa_name (lhs
);
4558 /* If the statement pointed to by STMT_PTR contains any expressions that need
4559 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4560 potential type incompatibilities (GSI is used to accommodate conversion
4561 statements and must point to the statement). Return true iff the statement
4565 sra_ipa_modify_assign (gimple
*stmt_ptr
, gimple_stmt_iterator
*gsi
,
4566 ipa_parm_adjustment_vec adjustments
)
4568 gimple stmt
= *stmt_ptr
;
4569 tree
*lhs_p
, *rhs_p
;
4572 if (!gimple_assign_single_p (stmt
))
4575 rhs_p
= gimple_assign_rhs1_ptr (stmt
);
4576 lhs_p
= gimple_assign_lhs_ptr (stmt
);
4578 any
= ipa_modify_expr (rhs_p
, false, adjustments
);
4579 any
|= ipa_modify_expr (lhs_p
, false, adjustments
);
4582 tree new_rhs
= NULL_TREE
;
4584 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p
), TREE_TYPE (*rhs_p
)))
4586 if (TREE_CODE (*rhs_p
) == CONSTRUCTOR
)
4588 /* V_C_Es of constructors can cause trouble (PR 42714). */
4589 if (is_gimple_reg_type (TREE_TYPE (*lhs_p
)))
4590 *rhs_p
= build_zero_cst (TREE_TYPE (*lhs_p
));
4592 *rhs_p
= build_constructor (TREE_TYPE (*lhs_p
),
4596 new_rhs
= fold_build1_loc (gimple_location (stmt
),
4597 VIEW_CONVERT_EXPR
, TREE_TYPE (*lhs_p
),
4600 else if (REFERENCE_CLASS_P (*rhs_p
)
4601 && is_gimple_reg_type (TREE_TYPE (*lhs_p
))
4602 && !is_gimple_reg (*lhs_p
))
4603 /* This can happen when an assignment in between two single field
4604 structures is turned into an assignment in between two pointers to
4605 scalars (PR 42237). */
4610 tree tmp
= force_gimple_operand_gsi (gsi
, new_rhs
, true, NULL_TREE
,
4611 true, GSI_SAME_STMT
);
4613 gimple_assign_set_rhs_from_tree (gsi
, tmp
);
4622 /* Traverse the function body and all modifications as described in
4623 ADJUSTMENTS. Return true iff the CFG has been changed. */
4626 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments
)
4628 bool cfg_changed
= false;
4631 FOR_EACH_BB_FN (bb
, cfun
)
4633 gimple_stmt_iterator gsi
;
4635 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4636 replace_removed_params_ssa_names (gsi_stmt (gsi
), adjustments
);
4638 gsi
= gsi_start_bb (bb
);
4639 while (!gsi_end_p (gsi
))
4641 gimple stmt
= gsi_stmt (gsi
);
4642 bool modified
= false;
4646 switch (gimple_code (stmt
))
4649 t
= gimple_return_retval_ptr (stmt
);
4650 if (*t
!= NULL_TREE
)
4651 modified
|= ipa_modify_expr (t
, true, adjustments
);
4655 modified
|= sra_ipa_modify_assign (&stmt
, &gsi
, adjustments
);
4656 modified
|= replace_removed_params_ssa_names (stmt
, adjustments
);
4660 /* Operands must be processed before the lhs. */
4661 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
4663 t
= gimple_call_arg_ptr (stmt
, i
);
4664 modified
|= ipa_modify_expr (t
, true, adjustments
);
4667 if (gimple_call_lhs (stmt
))
4669 t
= gimple_call_lhs_ptr (stmt
);
4670 modified
|= ipa_modify_expr (t
, false, adjustments
);
4671 modified
|= replace_removed_params_ssa_names (stmt
,
4677 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
4679 t
= &TREE_VALUE (gimple_asm_input_op (stmt
, i
));
4680 modified
|= ipa_modify_expr (t
, true, adjustments
);
4682 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
4684 t
= &TREE_VALUE (gimple_asm_output_op (stmt
, i
));
4685 modified
|= ipa_modify_expr (t
, false, adjustments
);
4696 if (maybe_clean_eh_stmt (stmt
)
4697 && gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
4707 /* Call gimple_debug_bind_reset_value on all debug statements describing
4708 gimple register parameters that are being removed or replaced. */
4711 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments
)
4714 gimple_stmt_iterator
*gsip
= NULL
, gsi
;
4716 if (MAY_HAVE_DEBUG_STMTS
&& single_succ_p (ENTRY_BLOCK_PTR_FOR_FN (cfun
)))
4718 gsi
= gsi_after_labels (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun
)));
4721 len
= adjustments
.length ();
4722 for (i
= 0; i
< len
; i
++)
4724 struct ipa_parm_adjustment
*adj
;
4725 imm_use_iterator ui
;
4726 gimple stmt
, def_temp
;
4727 tree name
, vexpr
, copy
= NULL_TREE
;
4728 use_operand_p use_p
;
4730 adj
= &adjustments
[i
];
4731 if (adj
->op
== IPA_PARM_OP_COPY
|| !is_gimple_reg (adj
->base
))
4733 name
= ssa_default_def (cfun
, adj
->base
);
4736 FOR_EACH_IMM_USE_STMT (stmt
, ui
, name
)
4738 if (gimple_clobber_p (stmt
))
4740 gimple_stmt_iterator cgsi
= gsi_for_stmt (stmt
);
4741 unlink_stmt_vdef (stmt
);
4742 gsi_remove (&cgsi
, true);
4743 release_defs (stmt
);
4746 /* All other users must have been removed by
4747 ipa_sra_modify_function_body. */
4748 gcc_assert (is_gimple_debug (stmt
));
4749 if (vexpr
== NULL
&& gsip
!= NULL
)
4751 gcc_assert (TREE_CODE (adj
->base
) == PARM_DECL
);
4752 vexpr
= make_node (DEBUG_EXPR_DECL
);
4753 def_temp
= gimple_build_debug_source_bind (vexpr
, adj
->base
,
4755 DECL_ARTIFICIAL (vexpr
) = 1;
4756 TREE_TYPE (vexpr
) = TREE_TYPE (name
);
4757 DECL_MODE (vexpr
) = DECL_MODE (adj
->base
);
4758 gsi_insert_before (gsip
, def_temp
, GSI_SAME_STMT
);
4762 FOR_EACH_IMM_USE_ON_STMT (use_p
, ui
)
4763 SET_USE (use_p
, vexpr
);
4766 gimple_debug_bind_reset_value (stmt
);
4769 /* Create a VAR_DECL for debug info purposes. */
4770 if (!DECL_IGNORED_P (adj
->base
))
4772 copy
= build_decl (DECL_SOURCE_LOCATION (current_function_decl
),
4773 VAR_DECL
, DECL_NAME (adj
->base
),
4774 TREE_TYPE (adj
->base
));
4775 if (DECL_PT_UID_SET_P (adj
->base
))
4776 SET_DECL_PT_UID (copy
, DECL_PT_UID (adj
->base
));
4777 TREE_ADDRESSABLE (copy
) = TREE_ADDRESSABLE (adj
->base
);
4778 TREE_READONLY (copy
) = TREE_READONLY (adj
->base
);
4779 TREE_THIS_VOLATILE (copy
) = TREE_THIS_VOLATILE (adj
->base
);
4780 DECL_GIMPLE_REG_P (copy
) = DECL_GIMPLE_REG_P (adj
->base
);
4781 DECL_ARTIFICIAL (copy
) = DECL_ARTIFICIAL (adj
->base
);
4782 DECL_IGNORED_P (copy
) = DECL_IGNORED_P (adj
->base
);
4783 DECL_ABSTRACT_ORIGIN (copy
) = DECL_ORIGIN (adj
->base
);
4784 DECL_SEEN_IN_BIND_EXPR_P (copy
) = 1;
4785 SET_DECL_RTL (copy
, 0);
4786 TREE_USED (copy
) = 1;
4787 DECL_CONTEXT (copy
) = current_function_decl
;
4788 add_local_decl (cfun
, copy
);
4790 BLOCK_VARS (DECL_INITIAL (current_function_decl
));
4791 BLOCK_VARS (DECL_INITIAL (current_function_decl
)) = copy
;
4793 if (gsip
!= NULL
&& copy
&& target_for_debug_bind (adj
->base
))
4795 gcc_assert (TREE_CODE (adj
->base
) == PARM_DECL
);
4797 def_temp
= gimple_build_debug_bind (copy
, vexpr
, NULL
);
4799 def_temp
= gimple_build_debug_source_bind (copy
, adj
->base
,
4801 gsi_insert_before (gsip
, def_temp
, GSI_SAME_STMT
);
4806 /* Return false if all callers have at least as many actual arguments as there
4807 are formal parameters in the current function and that their types
4811 some_callers_have_mismatched_arguments_p (struct cgraph_node
*node
,
4812 void *data ATTRIBUTE_UNUSED
)
4814 struct cgraph_edge
*cs
;
4815 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4816 if (!callsite_arguments_match_p (cs
->call_stmt
))
4822 /* Convert all callers of NODE. */
4825 convert_callers_for_node (struct cgraph_node
*node
,
4828 ipa_parm_adjustment_vec
*adjustments
= (ipa_parm_adjustment_vec
*) data
;
4829 bitmap recomputed_callers
= BITMAP_ALLOC (NULL
);
4830 struct cgraph_edge
*cs
;
4832 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4834 push_cfun (DECL_STRUCT_FUNCTION (cs
->caller
->decl
));
4837 fprintf (dump_file
, "Adjusting call %s/%i -> %s/%i\n",
4838 xstrdup (cs
->caller
->name ()),
4840 xstrdup (cs
->callee
->name ()),
4843 ipa_modify_call_arguments (cs
, cs
->call_stmt
, *adjustments
);
4848 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4849 if (bitmap_set_bit (recomputed_callers
, cs
->caller
->uid
)
4850 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs
->caller
->decl
)))
4851 compute_inline_parameters (cs
->caller
, true);
4852 BITMAP_FREE (recomputed_callers
);
4857 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4860 convert_callers (struct cgraph_node
*node
, tree old_decl
,
4861 ipa_parm_adjustment_vec adjustments
)
4863 basic_block this_block
;
4865 cgraph_for_node_and_aliases (node
, convert_callers_for_node
,
4866 &adjustments
, false);
4868 if (!encountered_recursive_call
)
4871 FOR_EACH_BB_FN (this_block
, cfun
)
4873 gimple_stmt_iterator gsi
;
4875 for (gsi
= gsi_start_bb (this_block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4877 gimple stmt
= gsi_stmt (gsi
);
4879 if (gimple_code (stmt
) != GIMPLE_CALL
)
4881 call_fndecl
= gimple_call_fndecl (stmt
);
4882 if (call_fndecl
== old_decl
)
4885 fprintf (dump_file
, "Adjusting recursive call");
4886 gimple_call_set_fndecl (stmt
, node
->decl
);
4887 ipa_modify_call_arguments (NULL
, stmt
, adjustments
);
4895 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4896 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4899 modify_function (struct cgraph_node
*node
, ipa_parm_adjustment_vec adjustments
)
4901 struct cgraph_node
*new_node
;
4903 vec
<cgraph_edge_p
> redirect_callers
= collect_callers_of_node (node
);
4905 rebuild_cgraph_edges ();
4906 free_dominance_info (CDI_DOMINATORS
);
4909 new_node
= cgraph_function_versioning (node
, redirect_callers
,
4911 NULL
, false, NULL
, NULL
, "isra");
4912 redirect_callers
.release ();
4914 push_cfun (DECL_STRUCT_FUNCTION (new_node
->decl
));
4915 ipa_modify_formal_parameters (current_function_decl
, adjustments
);
4916 cfg_changed
= ipa_sra_modify_function_body (adjustments
);
4917 sra_ipa_reset_debug_stmts (adjustments
);
4918 convert_callers (new_node
, node
->decl
, adjustments
);
4919 cgraph_make_node_local (new_node
);
4923 /* If NODE has a caller, return true. */
4926 has_caller_p (struct cgraph_node
*node
, void *data ATTRIBUTE_UNUSED
)
4933 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4934 attributes, return true otherwise. NODE is the cgraph node of the current
4938 ipa_sra_preliminary_function_checks (struct cgraph_node
*node
)
4940 if (!cgraph_node_can_be_local_p (node
))
4943 fprintf (dump_file
, "Function not local to this compilation unit.\n");
4947 if (!node
->local
.can_change_signature
)
4950 fprintf (dump_file
, "Function can not change signature.\n");
4954 if (!tree_versionable_function_p (node
->decl
))
4957 fprintf (dump_file
, "Function is not versionable.\n");
4961 if (!opt_for_fn (node
->decl
, optimize
)
4962 || !opt_for_fn (node
->decl
, flag_ipa_sra
))
4965 fprintf (dump_file
, "Function not optimized.\n");
4969 if (DECL_VIRTUAL_P (current_function_decl
))
4972 fprintf (dump_file
, "Function is a virtual method.\n");
4976 if ((DECL_COMDAT (node
->decl
) || DECL_EXTERNAL (node
->decl
))
4977 && inline_summary (node
)->size
>= MAX_INLINE_INSNS_AUTO
)
4980 fprintf (dump_file
, "Function too big to be made truly local.\n");
4984 if (!cgraph_for_node_and_aliases (node
, has_caller_p
, NULL
, true))
4988 "Function has no callers in this compilation unit.\n");
4995 fprintf (dump_file
, "Function uses stdarg. \n");
4999 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
5002 if (DECL_DISREGARD_INLINE_LIMITS (node
->decl
))
5005 fprintf (dump_file
, "Always inline function will be inlined "
5013 /* Perform early interprocedural SRA. */
5016 ipa_early_sra (void)
5018 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
5019 ipa_parm_adjustment_vec adjustments
;
5022 if (!ipa_sra_preliminary_function_checks (node
))
5026 sra_mode
= SRA_MODE_EARLY_IPA
;
5028 if (!find_param_candidates ())
5031 fprintf (dump_file
, "Function has no IPA-SRA candidates.\n");
5035 if (cgraph_for_node_and_aliases (node
,
5036 some_callers_have_mismatched_arguments_p
,
5040 fprintf (dump_file
, "There are callers with insufficient number of "
5041 "arguments or arguments with type mismatches.\n");
5045 bb_dereferences
= XCNEWVEC (HOST_WIDE_INT
,
5047 * last_basic_block_for_fn (cfun
));
5048 final_bbs
= BITMAP_ALLOC (NULL
);
5051 if (encountered_apply_args
)
5054 fprintf (dump_file
, "Function calls __builtin_apply_args().\n");
5058 if (encountered_unchangable_recursive_call
)
5061 fprintf (dump_file
, "Function calls itself with insufficient "
5062 "number of arguments.\n");
5066 adjustments
= analyze_all_param_acesses ();
5067 if (!adjustments
.exists ())
5070 ipa_dump_param_adjustments (dump_file
, adjustments
, current_function_decl
);
5072 if (modify_function (node
, adjustments
))
5073 ret
= TODO_update_ssa
| TODO_cleanup_cfg
;
5075 ret
= TODO_update_ssa
;
5076 adjustments
.release ();
5078 statistics_counter_event (cfun
, "Unused parameters deleted",
5079 sra_stats
.deleted_unused_parameters
);
5080 statistics_counter_event (cfun
, "Scalar parameters converted to by-value",
5081 sra_stats
.scalar_by_ref_to_by_val
);
5082 statistics_counter_event (cfun
, "Aggregate parameters broken up",
5083 sra_stats
.aggregate_params_reduced
);
5084 statistics_counter_event (cfun
, "Aggregate parameter components created",
5085 sra_stats
.param_reductions_created
);
5088 BITMAP_FREE (final_bbs
);
5089 free (bb_dereferences
);
5091 sra_deinitialize ();
5097 const pass_data pass_data_early_ipa_sra
=
5099 GIMPLE_PASS
, /* type */
5100 "eipa_sra", /* name */
5101 OPTGROUP_NONE
, /* optinfo_flags */
5102 true, /* has_execute */
5103 TV_IPA_SRA
, /* tv_id */
5104 0, /* properties_required */
5105 0, /* properties_provided */
5106 0, /* properties_destroyed */
5107 0, /* todo_flags_start */
5108 TODO_dump_symtab
, /* todo_flags_finish */
5111 class pass_early_ipa_sra
: public gimple_opt_pass
5114 pass_early_ipa_sra (gcc::context
*ctxt
)
5115 : gimple_opt_pass (pass_data_early_ipa_sra
, ctxt
)
5118 /* opt_pass methods: */
5119 virtual bool gate (function
*) { return flag_ipa_sra
&& dbg_cnt (eipa_sra
); }
5120 virtual unsigned int execute (function
*) { return ipa_early_sra (); }
5122 }; // class pass_early_ipa_sra
5127 make_pass_early_ipa_sra (gcc::context
*ctxt
)
5129 return new pass_early_ipa_sra (ctxt
);