1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
4 Copyright (C) 2008, 2009, 2010, 2011 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 "alloc-pool.h"
82 #include "tree-flow.h"
83 #include "tree-pass.h"
85 #include "statistics.h"
90 #include "tree-inline.h"
91 #include "gimple-pretty-print.h"
92 #include "ipa-inline.h"
94 /* Enumeration of all aggregate reductions we can do. */
95 enum sra_mode
{ SRA_MODE_EARLY_IPA
, /* early call regularization */
96 SRA_MODE_EARLY_INTRA
, /* early intraprocedural SRA */
97 SRA_MODE_INTRA
}; /* late intraprocedural SRA */
99 /* Global variable describing which aggregate reduction we are performing at
101 static enum sra_mode sra_mode
;
105 /* ACCESS represents each access to an aggregate variable (as a whole or a
106 part). It can also represent a group of accesses that refer to exactly the
107 same fragment of an aggregate (i.e. those that have exactly the same offset
108 and size). Such representatives for a single aggregate, once determined,
109 are linked in a linked list and have the group fields set.
111 Moreover, when doing intraprocedural SRA, a tree is built from those
112 representatives (by the means of first_child and next_sibling pointers), in
113 which all items in a subtree are "within" the root, i.e. their offset is
114 greater or equal to offset of the root and offset+size is smaller or equal
115 to offset+size of the root. Children of an access are sorted by offset.
117 Note that accesses to parts of vector and complex number types always
118 represented by an access to the whole complex number or a vector. It is a
119 duty of the modifying functions to replace them appropriately. */
123 /* Values returned by `get_ref_base_and_extent' for each component reference
124 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
125 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
126 HOST_WIDE_INT offset
;
130 /* Expression. It is context dependent so do not use it to create new
131 expressions to access the original aggregate. See PR 42154 for a
137 /* The statement this access belongs to. */
140 /* Next group representative for this aggregate. */
141 struct access
*next_grp
;
143 /* Pointer to the group representative. Pointer to itself if the struct is
144 the representative. */
145 struct access
*group_representative
;
147 /* If this access has any children (in terms of the definition above), this
148 points to the first one. */
149 struct access
*first_child
;
151 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
152 described above. In IPA-SRA this is a pointer to the next access
153 belonging to the same group (having the same representative). */
154 struct access
*next_sibling
;
156 /* Pointers to the first and last element in the linked list of assign
158 struct assign_link
*first_link
, *last_link
;
160 /* Pointer to the next access in the work queue. */
161 struct access
*next_queued
;
163 /* Replacement variable for this access "region." Never to be accessed
164 directly, always only by the means of get_access_replacement() and only
165 when grp_to_be_replaced flag is set. */
166 tree replacement_decl
;
168 /* Is this particular access write access? */
171 /* Is this access an access to a non-addressable field? */
172 unsigned non_addressable
: 1;
174 /* Is this access currently in the work queue? */
175 unsigned grp_queued
: 1;
177 /* Does this group contain a write access? This flag is propagated down the
179 unsigned grp_write
: 1;
181 /* Does this group contain a read access? This flag is propagated down the
183 unsigned grp_read
: 1;
185 /* Does this group contain a read access that comes from an assignment
186 statement? This flag is propagated down the access tree. */
187 unsigned grp_assignment_read
: 1;
189 /* Does this group contain a write access that comes from an assignment
190 statement? This flag is propagated down the access tree. */
191 unsigned grp_assignment_write
: 1;
193 /* Does this group contain a read access through a scalar type? This flag is
194 not propagated in the access tree in any direction. */
195 unsigned grp_scalar_read
: 1;
197 /* Does this group contain a write access through a scalar type? This flag
198 is not propagated in the access tree in any direction. */
199 unsigned grp_scalar_write
: 1;
201 /* Is this access an artificial one created to scalarize some record
203 unsigned grp_total_scalarization
: 1;
205 /* Other passes of the analysis use this bit to make function
206 analyze_access_subtree create scalar replacements for this group if
208 unsigned grp_hint
: 1;
210 /* Is the subtree rooted in this access fully covered by scalar
212 unsigned grp_covered
: 1;
214 /* If set to true, this access and all below it in an access tree must not be
216 unsigned grp_unscalarizable_region
: 1;
218 /* Whether data have been written to parts of the aggregate covered by this
219 access which is not to be scalarized. This flag is propagated up in the
221 unsigned grp_unscalarized_data
: 1;
223 /* Does this access and/or group contain a write access through a
225 unsigned grp_partial_lhs
: 1;
227 /* Set when a scalar replacement should be created for this variable. */
228 unsigned grp_to_be_replaced
: 1;
230 /* Should TREE_NO_WARNING of a replacement be set? */
231 unsigned grp_no_warning
: 1;
233 /* Is it possible that the group refers to data which might be (directly or
234 otherwise) modified? */
235 unsigned grp_maybe_modified
: 1;
237 /* Set when this is a representative of a pointer to scalar (i.e. by
238 reference) parameter which we consider for turning into a plain scalar
239 (i.e. a by value parameter). */
240 unsigned grp_scalar_ptr
: 1;
242 /* Set when we discover that this pointer is not safe to dereference in the
244 unsigned grp_not_necessarilly_dereferenced
: 1;
247 typedef struct access
*access_p
;
249 DEF_VEC_P (access_p
);
250 DEF_VEC_ALLOC_P (access_p
, heap
);
252 /* Alloc pool for allocating access structures. */
253 static alloc_pool access_pool
;
255 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
256 are used to propagate subaccesses from rhs to lhs as long as they don't
257 conflict with what is already there. */
260 struct access
*lacc
, *racc
;
261 struct assign_link
*next
;
264 /* Alloc pool for allocating assign link structures. */
265 static alloc_pool link_pool
;
267 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
268 static struct pointer_map_t
*base_access_vec
;
270 /* Set of candidates. */
271 static bitmap candidate_bitmap
;
272 static htab_t candidates
;
274 /* For a candidate UID return the candidates decl. */
277 candidate (unsigned uid
)
279 struct tree_decl_minimal t
;
281 return (tree
) htab_find_with_hash (candidates
, &t
, uid
);
284 /* Bitmap of candidates which we should try to entirely scalarize away and
285 those which cannot be (because they are and need be used as a whole). */
286 static bitmap should_scalarize_away_bitmap
, cannot_scalarize_away_bitmap
;
288 /* Obstack for creation of fancy names. */
289 static struct obstack name_obstack
;
291 /* Head of a linked list of accesses that need to have its subaccesses
292 propagated to their assignment counterparts. */
293 static struct access
*work_queue_head
;
295 /* Number of parameters of the analyzed function when doing early ipa SRA. */
296 static int func_param_count
;
298 /* scan_function sets the following to true if it encounters a call to
299 __builtin_apply_args. */
300 static bool encountered_apply_args
;
302 /* Set by scan_function when it finds a recursive call. */
303 static bool encountered_recursive_call
;
305 /* Set by scan_function when it finds a recursive call with less actual
306 arguments than formal parameters.. */
307 static bool encountered_unchangable_recursive_call
;
309 /* This is a table in which for each basic block and parameter there is a
310 distance (offset + size) in that parameter which is dereferenced and
311 accessed in that BB. */
312 static HOST_WIDE_INT
*bb_dereferences
;
313 /* Bitmap of BBs that can cause the function to "stop" progressing by
314 returning, throwing externally, looping infinitely or calling a function
315 which might abort etc.. */
316 static bitmap final_bbs
;
318 /* Representative of no accesses at all. */
319 static struct access no_accesses_representant
;
321 /* Predicate to test the special value. */
324 no_accesses_p (struct access
*access
)
326 return access
== &no_accesses_representant
;
329 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
330 representative fields are dumped, otherwise those which only describe the
331 individual access are. */
335 /* Number of processed aggregates is readily available in
336 analyze_all_variable_accesses and so is not stored here. */
338 /* Number of created scalar replacements. */
341 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
345 /* Number of statements created by generate_subtree_copies. */
348 /* Number of statements created by load_assign_lhs_subreplacements. */
351 /* Number of times sra_modify_assign has deleted a statement. */
354 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
355 RHS reparately due to type conversions or nonexistent matching
357 int separate_lhs_rhs_handling
;
359 /* Number of parameters that were removed because they were unused. */
360 int deleted_unused_parameters
;
362 /* Number of scalars passed as parameters by reference that have been
363 converted to be passed by value. */
364 int scalar_by_ref_to_by_val
;
366 /* Number of aggregate parameters that were replaced by one or more of their
368 int aggregate_params_reduced
;
370 /* Numbber of components created when splitting aggregate parameters. */
371 int param_reductions_created
;
375 dump_access (FILE *f
, struct access
*access
, bool grp
)
377 fprintf (f
, "access { ");
378 fprintf (f
, "base = (%d)'", DECL_UID (access
->base
));
379 print_generic_expr (f
, access
->base
, 0);
380 fprintf (f
, "', offset = " HOST_WIDE_INT_PRINT_DEC
, access
->offset
);
381 fprintf (f
, ", size = " HOST_WIDE_INT_PRINT_DEC
, access
->size
);
382 fprintf (f
, ", expr = ");
383 print_generic_expr (f
, access
->expr
, 0);
384 fprintf (f
, ", type = ");
385 print_generic_expr (f
, access
->type
, 0);
387 fprintf (f
, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
388 "grp_assignment_write = %d, grp_scalar_read = %d, "
389 "grp_scalar_write = %d, grp_total_scalarization = %d, "
390 "grp_hint = %d, grp_covered = %d, "
391 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
392 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
393 "grp_maybe_modified = %d, "
394 "grp_not_necessarilly_dereferenced = %d\n",
395 access
->grp_read
, access
->grp_write
, access
->grp_assignment_read
,
396 access
->grp_assignment_write
, access
->grp_scalar_read
,
397 access
->grp_scalar_write
, access
->grp_total_scalarization
,
398 access
->grp_hint
, access
->grp_covered
,
399 access
->grp_unscalarizable_region
, access
->grp_unscalarized_data
,
400 access
->grp_partial_lhs
, access
->grp_to_be_replaced
,
401 access
->grp_maybe_modified
,
402 access
->grp_not_necessarilly_dereferenced
);
404 fprintf (f
, ", write = %d, grp_total_scalarization = %d, "
405 "grp_partial_lhs = %d\n",
406 access
->write
, access
->grp_total_scalarization
,
407 access
->grp_partial_lhs
);
410 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
413 dump_access_tree_1 (FILE *f
, struct access
*access
, int level
)
419 for (i
= 0; i
< level
; i
++)
420 fputs ("* ", dump_file
);
422 dump_access (f
, access
, true);
424 if (access
->first_child
)
425 dump_access_tree_1 (f
, access
->first_child
, level
+ 1);
427 access
= access
->next_sibling
;
432 /* Dump all access trees for a variable, given the pointer to the first root in
436 dump_access_tree (FILE *f
, struct access
*access
)
438 for (; access
; access
= access
->next_grp
)
439 dump_access_tree_1 (f
, access
, 0);
442 /* Return true iff ACC is non-NULL and has subaccesses. */
445 access_has_children_p (struct access
*acc
)
447 return acc
&& acc
->first_child
;
450 /* Return true iff ACC is (partly) covered by at least one replacement. */
453 access_has_replacements_p (struct access
*acc
)
455 struct access
*child
;
456 if (acc
->grp_to_be_replaced
)
458 for (child
= acc
->first_child
; child
; child
= child
->next_sibling
)
459 if (access_has_replacements_p (child
))
464 /* Return a vector of pointers to accesses for the variable given in BASE or
465 NULL if there is none. */
467 static VEC (access_p
, heap
) *
468 get_base_access_vector (tree base
)
472 slot
= pointer_map_contains (base_access_vec
, base
);
476 return *(VEC (access_p
, heap
) **) slot
;
479 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
480 in ACCESS. Return NULL if it cannot be found. */
482 static struct access
*
483 find_access_in_subtree (struct access
*access
, HOST_WIDE_INT offset
,
486 while (access
&& (access
->offset
!= offset
|| access
->size
!= size
))
488 struct access
*child
= access
->first_child
;
490 while (child
&& (child
->offset
+ child
->size
<= offset
))
491 child
= child
->next_sibling
;
498 /* Return the first group representative for DECL or NULL if none exists. */
500 static struct access
*
501 get_first_repr_for_decl (tree base
)
503 VEC (access_p
, heap
) *access_vec
;
505 access_vec
= get_base_access_vector (base
);
509 return VEC_index (access_p
, access_vec
, 0);
512 /* Find an access representative for the variable BASE and given OFFSET and
513 SIZE. Requires that access trees have already been built. Return NULL if
514 it cannot be found. */
516 static struct access
*
517 get_var_base_offset_size_access (tree base
, HOST_WIDE_INT offset
,
520 struct access
*access
;
522 access
= get_first_repr_for_decl (base
);
523 while (access
&& (access
->offset
+ access
->size
<= offset
))
524 access
= access
->next_grp
;
528 return find_access_in_subtree (access
, offset
, size
);
531 /* Add LINK to the linked list of assign links of RACC. */
533 add_link_to_rhs (struct access
*racc
, struct assign_link
*link
)
535 gcc_assert (link
->racc
== racc
);
537 if (!racc
->first_link
)
539 gcc_assert (!racc
->last_link
);
540 racc
->first_link
= link
;
543 racc
->last_link
->next
= link
;
545 racc
->last_link
= link
;
549 /* Move all link structures in their linked list in OLD_RACC to the linked list
552 relink_to_new_repr (struct access
*new_racc
, struct access
*old_racc
)
554 if (!old_racc
->first_link
)
556 gcc_assert (!old_racc
->last_link
);
560 if (new_racc
->first_link
)
562 gcc_assert (!new_racc
->last_link
->next
);
563 gcc_assert (!old_racc
->last_link
|| !old_racc
->last_link
->next
);
565 new_racc
->last_link
->next
= old_racc
->first_link
;
566 new_racc
->last_link
= old_racc
->last_link
;
570 gcc_assert (!new_racc
->last_link
);
572 new_racc
->first_link
= old_racc
->first_link
;
573 new_racc
->last_link
= old_racc
->last_link
;
575 old_racc
->first_link
= old_racc
->last_link
= NULL
;
578 /* Add ACCESS to the work queue (which is actually a stack). */
581 add_access_to_work_queue (struct access
*access
)
583 if (!access
->grp_queued
)
585 gcc_assert (!access
->next_queued
);
586 access
->next_queued
= work_queue_head
;
587 access
->grp_queued
= 1;
588 work_queue_head
= access
;
592 /* Pop an access from the work queue, and return it, assuming there is one. */
594 static struct access
*
595 pop_access_from_work_queue (void)
597 struct access
*access
= work_queue_head
;
599 work_queue_head
= access
->next_queued
;
600 access
->next_queued
= NULL
;
601 access
->grp_queued
= 0;
606 /* Allocate necessary structures. */
609 sra_initialize (void)
611 candidate_bitmap
= BITMAP_ALLOC (NULL
);
612 candidates
= htab_create (VEC_length (tree
, cfun
->local_decls
) / 2,
613 uid_decl_map_hash
, uid_decl_map_eq
, NULL
);
614 should_scalarize_away_bitmap
= BITMAP_ALLOC (NULL
);
615 cannot_scalarize_away_bitmap
= BITMAP_ALLOC (NULL
);
616 gcc_obstack_init (&name_obstack
);
617 access_pool
= create_alloc_pool ("SRA accesses", sizeof (struct access
), 16);
618 link_pool
= create_alloc_pool ("SRA links", sizeof (struct assign_link
), 16);
619 base_access_vec
= pointer_map_create ();
620 memset (&sra_stats
, 0, sizeof (sra_stats
));
621 encountered_apply_args
= false;
622 encountered_recursive_call
= false;
623 encountered_unchangable_recursive_call
= false;
626 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
629 delete_base_accesses (const void *key ATTRIBUTE_UNUSED
, void **value
,
630 void *data ATTRIBUTE_UNUSED
)
632 VEC (access_p
, heap
) *access_vec
;
633 access_vec
= (VEC (access_p
, heap
) *) *value
;
634 VEC_free (access_p
, heap
, access_vec
);
639 /* Deallocate all general structures. */
642 sra_deinitialize (void)
644 BITMAP_FREE (candidate_bitmap
);
645 htab_delete (candidates
);
646 BITMAP_FREE (should_scalarize_away_bitmap
);
647 BITMAP_FREE (cannot_scalarize_away_bitmap
);
648 free_alloc_pool (access_pool
);
649 free_alloc_pool (link_pool
);
650 obstack_free (&name_obstack
, NULL
);
652 pointer_map_traverse (base_access_vec
, delete_base_accesses
, NULL
);
653 pointer_map_destroy (base_access_vec
);
656 /* Remove DECL from candidates for SRA and write REASON to the dump file if
659 disqualify_candidate (tree decl
, const char *reason
)
661 if (bitmap_clear_bit (candidate_bitmap
, DECL_UID (decl
)))
662 htab_clear_slot (candidates
,
663 htab_find_slot_with_hash (candidates
, decl
,
664 DECL_UID (decl
), NO_INSERT
));
666 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
668 fprintf (dump_file
, "! Disqualifying ");
669 print_generic_expr (dump_file
, decl
, 0);
670 fprintf (dump_file
, " - %s\n", reason
);
674 /* Return true iff the type contains a field or an element which does not allow
678 type_internals_preclude_sra_p (tree type
, const char **msg
)
683 switch (TREE_CODE (type
))
687 case QUAL_UNION_TYPE
:
688 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
689 if (TREE_CODE (fld
) == FIELD_DECL
)
691 tree ft
= TREE_TYPE (fld
);
693 if (TREE_THIS_VOLATILE (fld
))
695 *msg
= "volatile structure field";
698 if (!DECL_FIELD_OFFSET (fld
))
700 *msg
= "no structure field offset";
703 if (!DECL_SIZE (fld
))
705 *msg
= "zero structure field size";
708 if (!host_integerp (DECL_FIELD_OFFSET (fld
), 1))
710 *msg
= "structure field offset not fixed";
713 if (!host_integerp (DECL_SIZE (fld
), 1))
715 *msg
= "structure field size not fixed";
718 if (AGGREGATE_TYPE_P (ft
)
719 && int_bit_position (fld
) % BITS_PER_UNIT
!= 0)
721 *msg
= "structure field is bit field";
725 if (AGGREGATE_TYPE_P (ft
) && type_internals_preclude_sra_p (ft
, msg
))
732 et
= TREE_TYPE (type
);
734 if (TYPE_VOLATILE (et
))
736 *msg
= "element type is volatile";
740 if (AGGREGATE_TYPE_P (et
) && type_internals_preclude_sra_p (et
, msg
))
750 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
751 base variable if it is. Return T if it is not an SSA_NAME. */
754 get_ssa_base_param (tree t
)
756 if (TREE_CODE (t
) == SSA_NAME
)
758 if (SSA_NAME_IS_DEFAULT_DEF (t
))
759 return SSA_NAME_VAR (t
);
766 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
767 belongs to, unless the BB has already been marked as a potentially
771 mark_parm_dereference (tree base
, HOST_WIDE_INT dist
, gimple stmt
)
773 basic_block bb
= gimple_bb (stmt
);
774 int idx
, parm_index
= 0;
777 if (bitmap_bit_p (final_bbs
, bb
->index
))
780 for (parm
= DECL_ARGUMENTS (current_function_decl
);
781 parm
&& parm
!= base
;
782 parm
= DECL_CHAIN (parm
))
785 gcc_assert (parm_index
< func_param_count
);
787 idx
= bb
->index
* func_param_count
+ parm_index
;
788 if (bb_dereferences
[idx
] < dist
)
789 bb_dereferences
[idx
] = dist
;
792 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
793 the three fields. Also add it to the vector of accesses corresponding to
794 the base. Finally, return the new access. */
796 static struct access
*
797 create_access_1 (tree base
, HOST_WIDE_INT offset
, HOST_WIDE_INT size
)
799 VEC (access_p
, heap
) *vec
;
800 struct access
*access
;
803 access
= (struct access
*) pool_alloc (access_pool
);
804 memset (access
, 0, sizeof (struct access
));
806 access
->offset
= offset
;
809 slot
= pointer_map_contains (base_access_vec
, base
);
811 vec
= (VEC (access_p
, heap
) *) *slot
;
813 vec
= VEC_alloc (access_p
, heap
, 32);
815 VEC_safe_push (access_p
, heap
, vec
, access
);
817 *((struct VEC (access_p
,heap
) **)
818 pointer_map_insert (base_access_vec
, base
)) = vec
;
823 /* Create and insert access for EXPR. Return created access, or NULL if it is
826 static struct access
*
827 create_access (tree expr
, gimple stmt
, bool write
)
829 struct access
*access
;
830 HOST_WIDE_INT offset
, size
, max_size
;
832 bool ptr
, unscalarizable_region
= false;
834 base
= get_ref_base_and_extent (expr
, &offset
, &size
, &max_size
);
836 if (sra_mode
== SRA_MODE_EARLY_IPA
837 && TREE_CODE (base
) == MEM_REF
)
839 base
= get_ssa_base_param (TREE_OPERAND (base
, 0));
847 if (!DECL_P (base
) || !bitmap_bit_p (candidate_bitmap
, DECL_UID (base
)))
850 if (sra_mode
== SRA_MODE_EARLY_IPA
)
852 if (size
< 0 || size
!= max_size
)
854 disqualify_candidate (base
, "Encountered a variable sized access.");
857 if (TREE_CODE (expr
) == COMPONENT_REF
858 && DECL_BIT_FIELD (TREE_OPERAND (expr
, 1)))
860 disqualify_candidate (base
, "Encountered a bit-field access.");
863 gcc_checking_assert ((offset
% BITS_PER_UNIT
) == 0);
866 mark_parm_dereference (base
, offset
+ size
, stmt
);
870 if (size
!= max_size
)
873 unscalarizable_region
= true;
877 disqualify_candidate (base
, "Encountered an unconstrained access.");
882 access
= create_access_1 (base
, offset
, size
);
884 access
->type
= TREE_TYPE (expr
);
885 access
->write
= write
;
886 access
->grp_unscalarizable_region
= unscalarizable_region
;
889 if (TREE_CODE (expr
) == COMPONENT_REF
890 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr
, 1)))
891 access
->non_addressable
= 1;
897 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
898 register types or (recursively) records with only these two kinds of fields.
899 It also returns false if any of these records contains a bit-field. */
902 type_consists_of_records_p (tree type
)
906 if (TREE_CODE (type
) != RECORD_TYPE
)
909 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
910 if (TREE_CODE (fld
) == FIELD_DECL
)
912 tree ft
= TREE_TYPE (fld
);
914 if (DECL_BIT_FIELD (fld
))
917 if (!is_gimple_reg_type (ft
)
918 && !type_consists_of_records_p (ft
))
925 /* Create total_scalarization accesses for all scalar type fields in DECL that
926 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
927 must be the top-most VAR_DECL representing the variable, OFFSET must be the
928 offset of DECL within BASE. REF must be the memory reference expression for
932 completely_scalarize_record (tree base
, tree decl
, HOST_WIDE_INT offset
,
935 tree fld
, decl_type
= TREE_TYPE (decl
);
937 for (fld
= TYPE_FIELDS (decl_type
); fld
; fld
= DECL_CHAIN (fld
))
938 if (TREE_CODE (fld
) == FIELD_DECL
)
940 HOST_WIDE_INT pos
= offset
+ int_bit_position (fld
);
941 tree ft
= TREE_TYPE (fld
);
942 tree nref
= build3 (COMPONENT_REF
, TREE_TYPE (fld
), ref
, fld
,
945 if (is_gimple_reg_type (ft
))
947 struct access
*access
;
950 size
= tree_low_cst (DECL_SIZE (fld
), 1);
951 access
= create_access_1 (base
, pos
, size
);
954 access
->grp_total_scalarization
= 1;
955 /* Accesses for intraprocedural SRA can have their stmt NULL. */
958 completely_scalarize_record (base
, fld
, pos
, nref
);
962 /* Create total_scalarization accesses for all scalar type fields in VAR and
963 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
964 type_consists_of_records_p. */
967 completely_scalarize_var (tree var
)
969 HOST_WIDE_INT size
= tree_low_cst (DECL_SIZE (var
), 1);
970 struct access
*access
;
972 access
= create_access_1 (var
, 0, size
);
974 access
->type
= TREE_TYPE (var
);
975 access
->grp_total_scalarization
= 1;
977 completely_scalarize_record (var
, var
, 0, var
);
980 /* Search the given tree for a declaration by skipping handled components and
981 exclude it from the candidates. */
984 disqualify_base_of_expr (tree t
, const char *reason
)
986 t
= get_base_address (t
);
987 if (sra_mode
== SRA_MODE_EARLY_IPA
988 && TREE_CODE (t
) == MEM_REF
)
989 t
= get_ssa_base_param (TREE_OPERAND (t
, 0));
992 disqualify_candidate (t
, reason
);
995 /* Scan expression EXPR and create access structures for all accesses to
996 candidates for scalarization. Return the created access or NULL if none is
999 static struct access
*
1000 build_access_from_expr_1 (tree expr
, gimple stmt
, bool write
)
1002 struct access
*ret
= NULL
;
1005 if (TREE_CODE (expr
) == BIT_FIELD_REF
1006 || TREE_CODE (expr
) == IMAGPART_EXPR
1007 || TREE_CODE (expr
) == REALPART_EXPR
)
1009 expr
= TREE_OPERAND (expr
, 0);
1013 partial_ref
= false;
1015 /* We need to dive through V_C_Es in order to get the size of its parameter
1016 and not the result type. Ada produces such statements. We are also
1017 capable of handling the topmost V_C_E but not any of those buried in other
1018 handled components. */
1019 if (TREE_CODE (expr
) == VIEW_CONVERT_EXPR
)
1020 expr
= TREE_OPERAND (expr
, 0);
1022 if (contains_view_convert_expr_p (expr
))
1024 disqualify_base_of_expr (expr
, "V_C_E under a different handled "
1029 switch (TREE_CODE (expr
))
1032 if (TREE_CODE (TREE_OPERAND (expr
, 0)) != ADDR_EXPR
1033 && sra_mode
!= SRA_MODE_EARLY_IPA
)
1041 case ARRAY_RANGE_REF
:
1042 ret
= create_access (expr
, stmt
, write
);
1049 if (write
&& partial_ref
&& ret
)
1050 ret
->grp_partial_lhs
= 1;
1055 /* Scan expression EXPR and create access structures for all accesses to
1056 candidates for scalarization. Return true if any access has been inserted.
1057 STMT must be the statement from which the expression is taken, WRITE must be
1058 true if the expression is a store and false otherwise. */
1061 build_access_from_expr (tree expr
, gimple stmt
, bool write
)
1063 struct access
*access
;
1065 access
= build_access_from_expr_1 (expr
, stmt
, write
);
1068 /* This means the aggregate is accesses as a whole in a way other than an
1069 assign statement and thus cannot be removed even if we had a scalar
1070 replacement for everything. */
1071 if (cannot_scalarize_away_bitmap
)
1072 bitmap_set_bit (cannot_scalarize_away_bitmap
, DECL_UID (access
->base
));
1078 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1079 modes in which it matters, return true iff they have been disqualified. RHS
1080 may be NULL, in that case ignore it. If we scalarize an aggregate in
1081 intra-SRA we may need to add statements after each statement. This is not
1082 possible if a statement unconditionally has to end the basic block. */
1084 disqualify_ops_if_throwing_stmt (gimple stmt
, tree lhs
, tree rhs
)
1086 if ((sra_mode
== SRA_MODE_EARLY_INTRA
|| sra_mode
== SRA_MODE_INTRA
)
1087 && (stmt_can_throw_internal (stmt
) || stmt_ends_bb_p (stmt
)))
1089 disqualify_base_of_expr (lhs
, "LHS of a throwing stmt.");
1091 disqualify_base_of_expr (rhs
, "RHS of a throwing stmt.");
1097 /* Scan expressions occurring in STMT, create access structures for all accesses
1098 to candidates for scalarization and remove those candidates which occur in
1099 statements or expressions that prevent them from being split apart. Return
1100 true if any access has been inserted. */
1103 build_accesses_from_assign (gimple stmt
)
1106 struct access
*lacc
, *racc
;
1108 if (!gimple_assign_single_p (stmt
)
1109 /* Scope clobbers don't influence scalarization. */
1110 || gimple_clobber_p (stmt
))
1113 lhs
= gimple_assign_lhs (stmt
);
1114 rhs
= gimple_assign_rhs1 (stmt
);
1116 if (disqualify_ops_if_throwing_stmt (stmt
, lhs
, rhs
))
1119 racc
= build_access_from_expr_1 (rhs
, stmt
, false);
1120 lacc
= build_access_from_expr_1 (lhs
, stmt
, true);
1123 lacc
->grp_assignment_write
= 1;
1127 racc
->grp_assignment_read
= 1;
1128 if (should_scalarize_away_bitmap
&& !gimple_has_volatile_ops (stmt
)
1129 && !is_gimple_reg_type (racc
->type
))
1130 bitmap_set_bit (should_scalarize_away_bitmap
, DECL_UID (racc
->base
));
1134 && (sra_mode
== SRA_MODE_EARLY_INTRA
|| sra_mode
== SRA_MODE_INTRA
)
1135 && !lacc
->grp_unscalarizable_region
1136 && !racc
->grp_unscalarizable_region
1137 && AGGREGATE_TYPE_P (TREE_TYPE (lhs
))
1138 && lacc
->size
== racc
->size
1139 && useless_type_conversion_p (lacc
->type
, racc
->type
))
1141 struct assign_link
*link
;
1143 link
= (struct assign_link
*) pool_alloc (link_pool
);
1144 memset (link
, 0, sizeof (struct assign_link
));
1149 add_link_to_rhs (racc
, link
);
1152 return lacc
|| racc
;
1155 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1156 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1159 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED
, tree op
,
1160 void *data ATTRIBUTE_UNUSED
)
1162 op
= get_base_address (op
);
1165 disqualify_candidate (op
, "Non-scalarizable GIMPLE_ASM operand.");
1170 /* Return true iff callsite CALL has at least as many actual arguments as there
1171 are formal parameters of the function currently processed by IPA-SRA. */
1174 callsite_has_enough_arguments_p (gimple call
)
1176 return gimple_call_num_args (call
) >= (unsigned) func_param_count
;
1179 /* Scan function and look for interesting expressions and create access
1180 structures for them. Return true iff any access is created. */
1183 scan_function (void)
1190 gimple_stmt_iterator gsi
;
1191 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1193 gimple stmt
= gsi_stmt (gsi
);
1197 if (final_bbs
&& stmt_can_throw_external (stmt
))
1198 bitmap_set_bit (final_bbs
, bb
->index
);
1199 switch (gimple_code (stmt
))
1202 t
= gimple_return_retval (stmt
);
1204 ret
|= build_access_from_expr (t
, stmt
, false);
1206 bitmap_set_bit (final_bbs
, bb
->index
);
1210 ret
|= build_accesses_from_assign (stmt
);
1214 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
1215 ret
|= build_access_from_expr (gimple_call_arg (stmt
, i
),
1218 if (sra_mode
== SRA_MODE_EARLY_IPA
)
1220 tree dest
= gimple_call_fndecl (stmt
);
1221 int flags
= gimple_call_flags (stmt
);
1225 if (DECL_BUILT_IN_CLASS (dest
) == BUILT_IN_NORMAL
1226 && DECL_FUNCTION_CODE (dest
) == BUILT_IN_APPLY_ARGS
)
1227 encountered_apply_args
= true;
1228 if (cgraph_get_node (dest
)
1229 == cgraph_get_node (current_function_decl
))
1231 encountered_recursive_call
= true;
1232 if (!callsite_has_enough_arguments_p (stmt
))
1233 encountered_unchangable_recursive_call
= true;
1238 && (flags
& (ECF_CONST
| ECF_PURE
)) == 0)
1239 bitmap_set_bit (final_bbs
, bb
->index
);
1242 t
= gimple_call_lhs (stmt
);
1243 if (t
&& !disqualify_ops_if_throwing_stmt (stmt
, t
, NULL
))
1244 ret
|= build_access_from_expr (t
, stmt
, true);
1248 walk_stmt_load_store_addr_ops (stmt
, NULL
, NULL
, NULL
,
1251 bitmap_set_bit (final_bbs
, bb
->index
);
1253 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
1255 t
= TREE_VALUE (gimple_asm_input_op (stmt
, i
));
1256 ret
|= build_access_from_expr (t
, stmt
, false);
1258 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
1260 t
= TREE_VALUE (gimple_asm_output_op (stmt
, i
));
1261 ret
|= build_access_from_expr (t
, stmt
, true);
1274 /* Helper of QSORT function. There are pointers to accesses in the array. An
1275 access is considered smaller than another if it has smaller offset or if the
1276 offsets are the same but is size is bigger. */
1279 compare_access_positions (const void *a
, const void *b
)
1281 const access_p
*fp1
= (const access_p
*) a
;
1282 const access_p
*fp2
= (const access_p
*) b
;
1283 const access_p f1
= *fp1
;
1284 const access_p f2
= *fp2
;
1286 if (f1
->offset
!= f2
->offset
)
1287 return f1
->offset
< f2
->offset
? -1 : 1;
1289 if (f1
->size
== f2
->size
)
1291 if (f1
->type
== f2
->type
)
1293 /* Put any non-aggregate type before any aggregate type. */
1294 else if (!is_gimple_reg_type (f1
->type
)
1295 && is_gimple_reg_type (f2
->type
))
1297 else if (is_gimple_reg_type (f1
->type
)
1298 && !is_gimple_reg_type (f2
->type
))
1300 /* Put any complex or vector type before any other scalar type. */
1301 else if (TREE_CODE (f1
->type
) != COMPLEX_TYPE
1302 && TREE_CODE (f1
->type
) != VECTOR_TYPE
1303 && (TREE_CODE (f2
->type
) == COMPLEX_TYPE
1304 || TREE_CODE (f2
->type
) == VECTOR_TYPE
))
1306 else if ((TREE_CODE (f1
->type
) == COMPLEX_TYPE
1307 || TREE_CODE (f1
->type
) == VECTOR_TYPE
)
1308 && TREE_CODE (f2
->type
) != COMPLEX_TYPE
1309 && TREE_CODE (f2
->type
) != VECTOR_TYPE
)
1311 /* Put the integral type with the bigger precision first. */
1312 else if (INTEGRAL_TYPE_P (f1
->type
)
1313 && INTEGRAL_TYPE_P (f2
->type
))
1314 return TYPE_PRECISION (f2
->type
) - TYPE_PRECISION (f1
->type
);
1315 /* Put any integral type with non-full precision last. */
1316 else if (INTEGRAL_TYPE_P (f1
->type
)
1317 && (TREE_INT_CST_LOW (TYPE_SIZE (f1
->type
))
1318 != TYPE_PRECISION (f1
->type
)))
1320 else if (INTEGRAL_TYPE_P (f2
->type
)
1321 && (TREE_INT_CST_LOW (TYPE_SIZE (f2
->type
))
1322 != TYPE_PRECISION (f2
->type
)))
1324 /* Stabilize the sort. */
1325 return TYPE_UID (f1
->type
) - TYPE_UID (f2
->type
);
1328 /* We want the bigger accesses first, thus the opposite operator in the next
1330 return f1
->size
> f2
->size
? -1 : 1;
1334 /* Append a name of the declaration to the name obstack. A helper function for
1338 make_fancy_decl_name (tree decl
)
1342 tree name
= DECL_NAME (decl
);
1344 obstack_grow (&name_obstack
, IDENTIFIER_POINTER (name
),
1345 IDENTIFIER_LENGTH (name
));
1348 sprintf (buffer
, "D%u", DECL_UID (decl
));
1349 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1353 /* Helper for make_fancy_name. */
1356 make_fancy_name_1 (tree expr
)
1363 make_fancy_decl_name (expr
);
1367 switch (TREE_CODE (expr
))
1370 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1371 obstack_1grow (&name_obstack
, '$');
1372 make_fancy_decl_name (TREE_OPERAND (expr
, 1));
1376 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1377 obstack_1grow (&name_obstack
, '$');
1378 /* Arrays with only one element may not have a constant as their
1380 index
= TREE_OPERAND (expr
, 1);
1381 if (TREE_CODE (index
) != INTEGER_CST
)
1383 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
, TREE_INT_CST_LOW (index
));
1384 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1388 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1392 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1393 if (!integer_zerop (TREE_OPERAND (expr
, 1)))
1395 obstack_1grow (&name_obstack
, '$');
1396 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
,
1397 TREE_INT_CST_LOW (TREE_OPERAND (expr
, 1)));
1398 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1405 gcc_unreachable (); /* we treat these as scalars. */
1412 /* Create a human readable name for replacement variable of ACCESS. */
1415 make_fancy_name (tree expr
)
1417 make_fancy_name_1 (expr
);
1418 obstack_1grow (&name_obstack
, '\0');
1419 return XOBFINISH (&name_obstack
, char *);
1422 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1423 EXP_TYPE at the given OFFSET. If BASE is something for which
1424 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1425 to insert new statements either before or below the current one as specified
1426 by INSERT_AFTER. This function is not capable of handling bitfields. */
1429 build_ref_for_offset (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1430 tree exp_type
, gimple_stmt_iterator
*gsi
,
1433 tree prev_base
= base
;
1435 HOST_WIDE_INT base_offset
;
1436 unsigned HOST_WIDE_INT misalign
;
1439 gcc_checking_assert (offset
% BITS_PER_UNIT
== 0);
1441 base
= get_addr_base_and_unit_offset (base
, &base_offset
);
1443 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1444 offset such as array[var_index]. */
1450 gcc_checking_assert (gsi
);
1451 tmp
= make_ssa_name (build_pointer_type (TREE_TYPE (prev_base
)), NULL
);
1452 addr
= build_fold_addr_expr (unshare_expr (prev_base
));
1453 STRIP_USELESS_TYPE_CONVERSION (addr
);
1454 stmt
= gimple_build_assign (tmp
, addr
);
1455 gimple_set_location (stmt
, loc
);
1457 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
1459 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1461 off
= build_int_cst (reference_alias_ptr_type (prev_base
),
1462 offset
/ BITS_PER_UNIT
);
1465 else if (TREE_CODE (base
) == MEM_REF
)
1467 off
= build_int_cst (TREE_TYPE (TREE_OPERAND (base
, 1)),
1468 base_offset
+ offset
/ BITS_PER_UNIT
);
1469 off
= int_const_binop (PLUS_EXPR
, TREE_OPERAND (base
, 1), off
);
1470 base
= unshare_expr (TREE_OPERAND (base
, 0));
1474 off
= build_int_cst (reference_alias_ptr_type (base
),
1475 base_offset
+ offset
/ BITS_PER_UNIT
);
1476 base
= build_fold_addr_expr (unshare_expr (base
));
1479 /* If prev_base were always an originally performed access
1480 we can extract more optimistic alignment information
1481 by looking at the access mode. That would constrain the
1482 alignment of base + base_offset which we would need to
1483 adjust according to offset. */
1484 if (!get_pointer_alignment_1 (base
, &align
, &misalign
))
1486 gcc_assert (misalign
== 0);
1487 if (TREE_CODE (prev_base
) == MEM_REF
1488 || TREE_CODE (prev_base
) == TARGET_MEM_REF
)
1489 align
= TYPE_ALIGN (TREE_TYPE (prev_base
));
1491 misalign
+= (double_int_sext (tree_to_double_int (off
),
1492 TYPE_PRECISION (TREE_TYPE (off
))).low
1494 misalign
= misalign
& (align
- 1);
1496 align
= (misalign
& -misalign
);
1497 if (align
< TYPE_ALIGN (exp_type
))
1498 exp_type
= build_aligned_type (exp_type
, align
);
1500 return fold_build2_loc (loc
, MEM_REF
, exp_type
, base
, off
);
1503 /* Construct a memory reference to a part of an aggregate BASE at the given
1504 OFFSET and of the same type as MODEL. In case this is a reference to a
1505 bit-field, the function will replicate the last component_ref of model's
1506 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1507 build_ref_for_offset. */
1510 build_ref_for_model (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1511 struct access
*model
, gimple_stmt_iterator
*gsi
,
1514 if (TREE_CODE (model
->expr
) == COMPONENT_REF
1515 && DECL_BIT_FIELD (TREE_OPERAND (model
->expr
, 1)))
1517 /* This access represents a bit-field. */
1518 tree t
, exp_type
, fld
= TREE_OPERAND (model
->expr
, 1);
1520 offset
-= int_bit_position (fld
);
1521 exp_type
= TREE_TYPE (TREE_OPERAND (model
->expr
, 0));
1522 t
= build_ref_for_offset (loc
, base
, offset
, exp_type
, gsi
, insert_after
);
1523 return fold_build3_loc (loc
, COMPONENT_REF
, TREE_TYPE (fld
), t
, fld
,
1527 return build_ref_for_offset (loc
, base
, offset
, model
->type
,
1531 /* Construct a memory reference consisting of component_refs and array_refs to
1532 a part of an aggregate *RES (which is of type TYPE). The requested part
1533 should have type EXP_TYPE at be the given OFFSET. This function might not
1534 succeed, it returns true when it does and only then *RES points to something
1535 meaningful. This function should be used only to build expressions that we
1536 might need to present to user (e.g. in warnings). In all other situations,
1537 build_ref_for_model or build_ref_for_offset should be used instead. */
1540 build_user_friendly_ref_for_offset (tree
*res
, tree type
, HOST_WIDE_INT offset
,
1546 tree tr_size
, index
, minidx
;
1547 HOST_WIDE_INT el_size
;
1549 if (offset
== 0 && exp_type
1550 && types_compatible_p (exp_type
, type
))
1553 switch (TREE_CODE (type
))
1556 case QUAL_UNION_TYPE
:
1558 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
1560 HOST_WIDE_INT pos
, size
;
1561 tree tr_pos
, expr
, *expr_ptr
;
1563 if (TREE_CODE (fld
) != FIELD_DECL
)
1566 tr_pos
= bit_position (fld
);
1567 if (!tr_pos
|| !host_integerp (tr_pos
, 1))
1569 pos
= TREE_INT_CST_LOW (tr_pos
);
1570 gcc_assert (TREE_CODE (type
) == RECORD_TYPE
|| pos
== 0);
1571 tr_size
= DECL_SIZE (fld
);
1572 if (!tr_size
|| !host_integerp (tr_size
, 1))
1574 size
= TREE_INT_CST_LOW (tr_size
);
1580 else if (pos
> offset
|| (pos
+ size
) <= offset
)
1583 expr
= build3 (COMPONENT_REF
, TREE_TYPE (fld
), *res
, fld
,
1586 if (build_user_friendly_ref_for_offset (expr_ptr
, TREE_TYPE (fld
),
1587 offset
- pos
, exp_type
))
1596 tr_size
= TYPE_SIZE (TREE_TYPE (type
));
1597 if (!tr_size
|| !host_integerp (tr_size
, 1))
1599 el_size
= tree_low_cst (tr_size
, 1);
1601 minidx
= TYPE_MIN_VALUE (TYPE_DOMAIN (type
));
1602 if (TREE_CODE (minidx
) != INTEGER_CST
|| el_size
== 0)
1604 index
= build_int_cst (TYPE_DOMAIN (type
), offset
/ el_size
);
1605 if (!integer_zerop (minidx
))
1606 index
= int_const_binop (PLUS_EXPR
, index
, minidx
);
1607 *res
= build4 (ARRAY_REF
, TREE_TYPE (type
), *res
, index
,
1608 NULL_TREE
, NULL_TREE
);
1609 offset
= offset
% el_size
;
1610 type
= TREE_TYPE (type
);
1625 /* Return true iff TYPE is stdarg va_list type. */
1628 is_va_list_type (tree type
)
1630 return TYPE_MAIN_VARIANT (type
) == TYPE_MAIN_VARIANT (va_list_type_node
);
1633 /* Print message to dump file why a variable was rejected. */
1636 reject (tree var
, const char *msg
)
1638 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1640 fprintf (dump_file
, "Rejected (%d): %s: ", DECL_UID (var
), msg
);
1641 print_generic_expr (dump_file
, var
, 0);
1642 fprintf (dump_file
, "\n");
1646 /* Return true if VAR is a candidate for SRA. */
1649 maybe_add_sra_candidate (tree var
)
1651 tree type
= TREE_TYPE (var
);
1655 if (!AGGREGATE_TYPE_P (type
))
1657 reject (var
, "not aggregate");
1660 if (needs_to_live_in_memory (var
))
1662 reject (var
, "needs to live in memory");
1665 if (TREE_THIS_VOLATILE (var
))
1667 reject (var
, "is volatile");
1670 if (!COMPLETE_TYPE_P (type
))
1672 reject (var
, "has incomplete type");
1675 if (!host_integerp (TYPE_SIZE (type
), 1))
1677 reject (var
, "type size not fixed");
1680 if (tree_low_cst (TYPE_SIZE (type
), 1) == 0)
1682 reject (var
, "type size is zero");
1685 if (type_internals_preclude_sra_p (type
, &msg
))
1690 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1691 we also want to schedule it rather late. Thus we ignore it in
1693 (sra_mode
== SRA_MODE_EARLY_INTRA
1694 && is_va_list_type (type
)))
1696 reject (var
, "is va_list");
1700 bitmap_set_bit (candidate_bitmap
, DECL_UID (var
));
1701 slot
= htab_find_slot_with_hash (candidates
, var
, DECL_UID (var
), INSERT
);
1702 *slot
= (void *) var
;
1704 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1706 fprintf (dump_file
, "Candidate (%d): ", DECL_UID (var
));
1707 print_generic_expr (dump_file
, var
, 0);
1708 fprintf (dump_file
, "\n");
1714 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1715 those with type which is suitable for scalarization. */
1718 find_var_candidates (void)
1724 for (parm
= DECL_ARGUMENTS (current_function_decl
);
1726 parm
= DECL_CHAIN (parm
))
1727 ret
|= maybe_add_sra_candidate (parm
);
1729 FOR_EACH_LOCAL_DECL (cfun
, i
, var
)
1731 if (TREE_CODE (var
) != VAR_DECL
)
1734 ret
|= maybe_add_sra_candidate (var
);
1740 /* Sort all accesses for the given variable, check for partial overlaps and
1741 return NULL if there are any. If there are none, pick a representative for
1742 each combination of offset and size and create a linked list out of them.
1743 Return the pointer to the first representative and make sure it is the first
1744 one in the vector of accesses. */
1746 static struct access
*
1747 sort_and_splice_var_accesses (tree var
)
1749 int i
, j
, access_count
;
1750 struct access
*res
, **prev_acc_ptr
= &res
;
1751 VEC (access_p
, heap
) *access_vec
;
1753 HOST_WIDE_INT low
= -1, high
= 0;
1755 access_vec
= get_base_access_vector (var
);
1758 access_count
= VEC_length (access_p
, access_vec
);
1760 /* Sort by <OFFSET, SIZE>. */
1761 VEC_qsort (access_p
, access_vec
, compare_access_positions
);
1764 while (i
< access_count
)
1766 struct access
*access
= VEC_index (access_p
, access_vec
, i
);
1767 bool grp_write
= access
->write
;
1768 bool grp_read
= !access
->write
;
1769 bool grp_scalar_write
= access
->write
1770 && is_gimple_reg_type (access
->type
);
1771 bool grp_scalar_read
= !access
->write
1772 && is_gimple_reg_type (access
->type
);
1773 bool grp_assignment_read
= access
->grp_assignment_read
;
1774 bool grp_assignment_write
= access
->grp_assignment_write
;
1775 bool multiple_scalar_reads
= false;
1776 bool total_scalarization
= access
->grp_total_scalarization
;
1777 bool grp_partial_lhs
= access
->grp_partial_lhs
;
1778 bool first_scalar
= is_gimple_reg_type (access
->type
);
1779 bool unscalarizable_region
= access
->grp_unscalarizable_region
;
1781 if (first
|| access
->offset
>= high
)
1784 low
= access
->offset
;
1785 high
= access
->offset
+ access
->size
;
1787 else if (access
->offset
> low
&& access
->offset
+ access
->size
> high
)
1790 gcc_assert (access
->offset
>= low
1791 && access
->offset
+ access
->size
<= high
);
1794 while (j
< access_count
)
1796 struct access
*ac2
= VEC_index (access_p
, access_vec
, j
);
1797 if (ac2
->offset
!= access
->offset
|| ac2
->size
!= access
->size
)
1802 grp_scalar_write
= (grp_scalar_write
1803 || is_gimple_reg_type (ac2
->type
));
1808 if (is_gimple_reg_type (ac2
->type
))
1810 if (grp_scalar_read
)
1811 multiple_scalar_reads
= true;
1813 grp_scalar_read
= true;
1816 grp_assignment_read
|= ac2
->grp_assignment_read
;
1817 grp_assignment_write
|= ac2
->grp_assignment_write
;
1818 grp_partial_lhs
|= ac2
->grp_partial_lhs
;
1819 unscalarizable_region
|= ac2
->grp_unscalarizable_region
;
1820 total_scalarization
|= ac2
->grp_total_scalarization
;
1821 relink_to_new_repr (access
, ac2
);
1823 /* If there are both aggregate-type and scalar-type accesses with
1824 this combination of size and offset, the comparison function
1825 should have put the scalars first. */
1826 gcc_assert (first_scalar
|| !is_gimple_reg_type (ac2
->type
));
1827 ac2
->group_representative
= access
;
1833 access
->group_representative
= access
;
1834 access
->grp_write
= grp_write
;
1835 access
->grp_read
= grp_read
;
1836 access
->grp_scalar_read
= grp_scalar_read
;
1837 access
->grp_scalar_write
= grp_scalar_write
;
1838 access
->grp_assignment_read
= grp_assignment_read
;
1839 access
->grp_assignment_write
= grp_assignment_write
;
1840 access
->grp_hint
= multiple_scalar_reads
|| total_scalarization
;
1841 access
->grp_total_scalarization
= total_scalarization
;
1842 access
->grp_partial_lhs
= grp_partial_lhs
;
1843 access
->grp_unscalarizable_region
= unscalarizable_region
;
1844 if (access
->first_link
)
1845 add_access_to_work_queue (access
);
1847 *prev_acc_ptr
= access
;
1848 prev_acc_ptr
= &access
->next_grp
;
1851 gcc_assert (res
== VEC_index (access_p
, access_vec
, 0));
1855 /* Create a variable for the given ACCESS which determines the type, name and a
1856 few other properties. Return the variable declaration and store it also to
1857 ACCESS->replacement. */
1860 create_access_replacement (struct access
*access
)
1864 repl
= create_tmp_var (access
->type
, "SR");
1865 if (TREE_CODE (access
->type
) == COMPLEX_TYPE
1866 || TREE_CODE (access
->type
) == VECTOR_TYPE
)
1868 if (!access
->grp_partial_lhs
)
1869 DECL_GIMPLE_REG_P (repl
) = 1;
1871 else if (access
->grp_partial_lhs
1872 && is_gimple_reg_type (access
->type
))
1873 TREE_ADDRESSABLE (repl
) = 1;
1875 DECL_SOURCE_LOCATION (repl
) = DECL_SOURCE_LOCATION (access
->base
);
1876 DECL_ARTIFICIAL (repl
) = 1;
1877 DECL_IGNORED_P (repl
) = DECL_IGNORED_P (access
->base
);
1879 if (DECL_NAME (access
->base
)
1880 && !DECL_IGNORED_P (access
->base
)
1881 && !DECL_ARTIFICIAL (access
->base
))
1883 char *pretty_name
= make_fancy_name (access
->expr
);
1884 tree debug_expr
= unshare_expr (access
->expr
), d
;
1887 DECL_NAME (repl
) = get_identifier (pretty_name
);
1888 obstack_free (&name_obstack
, pretty_name
);
1890 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1891 as DECL_DEBUG_EXPR isn't considered when looking for still
1892 used SSA_NAMEs and thus they could be freed. All debug info
1893 generation cares is whether something is constant or variable
1894 and that get_ref_base_and_extent works properly on the
1895 expression. It cannot handle accesses at a non-constant offset
1896 though, so just give up in those cases. */
1897 for (d
= debug_expr
; !fail
&& handled_component_p (d
);
1898 d
= TREE_OPERAND (d
, 0))
1899 switch (TREE_CODE (d
))
1902 case ARRAY_RANGE_REF
:
1903 if (TREE_OPERAND (d
, 1)
1904 && TREE_CODE (TREE_OPERAND (d
, 1)) != INTEGER_CST
)
1906 if (TREE_OPERAND (d
, 3)
1907 && TREE_CODE (TREE_OPERAND (d
, 3)) != INTEGER_CST
)
1911 if (TREE_OPERAND (d
, 2)
1912 && TREE_CODE (TREE_OPERAND (d
, 2)) != INTEGER_CST
)
1920 SET_DECL_DEBUG_EXPR (repl
, debug_expr
);
1921 DECL_DEBUG_EXPR_IS_FROM (repl
) = 1;
1923 if (access
->grp_no_warning
)
1924 TREE_NO_WARNING (repl
) = 1;
1926 TREE_NO_WARNING (repl
) = TREE_NO_WARNING (access
->base
);
1929 TREE_NO_WARNING (repl
) = 1;
1933 fprintf (dump_file
, "Created a replacement for ");
1934 print_generic_expr (dump_file
, access
->base
, 0);
1935 fprintf (dump_file
, " offset: %u, size: %u: ",
1936 (unsigned) access
->offset
, (unsigned) access
->size
);
1937 print_generic_expr (dump_file
, repl
, 0);
1938 fprintf (dump_file
, "\n");
1940 sra_stats
.replacements
++;
1945 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1948 get_access_replacement (struct access
*access
)
1950 if (!access
->replacement_decl
)
1951 access
->replacement_decl
= create_access_replacement (access
);
1952 return access
->replacement_decl
;
1956 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1957 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1958 to it is not "within" the root. Return false iff some accesses partially
1962 build_access_subtree (struct access
**access
)
1964 struct access
*root
= *access
, *last_child
= NULL
;
1965 HOST_WIDE_INT limit
= root
->offset
+ root
->size
;
1967 *access
= (*access
)->next_grp
;
1968 while (*access
&& (*access
)->offset
+ (*access
)->size
<= limit
)
1971 root
->first_child
= *access
;
1973 last_child
->next_sibling
= *access
;
1974 last_child
= *access
;
1976 if (!build_access_subtree (access
))
1980 if (*access
&& (*access
)->offset
< limit
)
1986 /* Build a tree of access representatives, ACCESS is the pointer to the first
1987 one, others are linked in a list by the next_grp field. Return false iff
1988 some accesses partially overlap. */
1991 build_access_trees (struct access
*access
)
1995 struct access
*root
= access
;
1997 if (!build_access_subtree (&access
))
1999 root
->next_grp
= access
;
2004 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2008 expr_with_var_bounded_array_refs_p (tree expr
)
2010 while (handled_component_p (expr
))
2012 if (TREE_CODE (expr
) == ARRAY_REF
2013 && !host_integerp (array_ref_low_bound (expr
), 0))
2015 expr
= TREE_OPERAND (expr
, 0);
2020 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2021 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2022 sorts of access flags appropriately along the way, notably always set
2023 grp_read and grp_assign_read according to MARK_READ and grp_write when
2026 Creating a replacement for a scalar access is considered beneficial if its
2027 grp_hint is set (this means we are either attempting total scalarization or
2028 there is more than one direct read access) or according to the following
2031 Access written to through a scalar type (once or more times)
2033 | Written to in an assignment statement
2035 | | Access read as scalar _once_
2037 | | | Read in an assignment statement
2039 | | | | Scalarize Comment
2040 -----------------------------------------------------------------------------
2041 0 0 0 0 No access for the scalar
2042 0 0 0 1 No access for the scalar
2043 0 0 1 0 No Single read - won't help
2044 0 0 1 1 No The same case
2045 0 1 0 0 No access for the scalar
2046 0 1 0 1 No access for the scalar
2047 0 1 1 0 Yes s = *g; return s.i;
2048 0 1 1 1 Yes The same case as above
2049 1 0 0 0 No Won't help
2050 1 0 0 1 Yes s.i = 1; *g = s;
2051 1 0 1 0 Yes s.i = 5; g = s.i;
2052 1 0 1 1 Yes The same case as above
2053 1 1 0 0 No Won't help.
2054 1 1 0 1 Yes s.i = 1; *g = s;
2055 1 1 1 0 Yes s = *g; return s.i;
2056 1 1 1 1 Yes Any of the above yeses */
2059 analyze_access_subtree (struct access
*root
, struct access
*parent
,
2060 bool allow_replacements
)
2062 struct access
*child
;
2063 HOST_WIDE_INT limit
= root
->offset
+ root
->size
;
2064 HOST_WIDE_INT covered_to
= root
->offset
;
2065 bool scalar
= is_gimple_reg_type (root
->type
);
2066 bool hole
= false, sth_created
= false;
2070 if (parent
->grp_read
)
2072 if (parent
->grp_assignment_read
)
2073 root
->grp_assignment_read
= 1;
2074 if (parent
->grp_write
)
2075 root
->grp_write
= 1;
2076 if (parent
->grp_assignment_write
)
2077 root
->grp_assignment_write
= 1;
2078 if (parent
->grp_total_scalarization
)
2079 root
->grp_total_scalarization
= 1;
2082 if (root
->grp_unscalarizable_region
)
2083 allow_replacements
= false;
2085 if (allow_replacements
&& expr_with_var_bounded_array_refs_p (root
->expr
))
2086 allow_replacements
= false;
2088 for (child
= root
->first_child
; child
; child
= child
->next_sibling
)
2090 hole
|= covered_to
< child
->offset
;
2091 sth_created
|= analyze_access_subtree (child
, root
,
2092 allow_replacements
&& !scalar
);
2094 root
->grp_unscalarized_data
|= child
->grp_unscalarized_data
;
2095 root
->grp_total_scalarization
&= child
->grp_total_scalarization
;
2096 if (child
->grp_covered
)
2097 covered_to
+= child
->size
;
2102 if (allow_replacements
&& scalar
&& !root
->first_child
2104 || ((root
->grp_scalar_read
|| root
->grp_assignment_read
)
2105 && (root
->grp_scalar_write
|| root
->grp_assignment_write
))))
2107 bool new_integer_type
;
2108 /* Always create access replacements that cover the whole access.
2109 For integral types this means the precision has to match.
2110 Avoid assumptions based on the integral type kind, too. */
2111 if (INTEGRAL_TYPE_P (root
->type
)
2112 && (TREE_CODE (root
->type
) != INTEGER_TYPE
2113 || TYPE_PRECISION (root
->type
) != root
->size
)
2114 /* But leave bitfield accesses alone. */
2115 && (TREE_CODE (root
->expr
) != COMPONENT_REF
2116 || !DECL_BIT_FIELD (TREE_OPERAND (root
->expr
, 1))))
2118 tree rt
= root
->type
;
2119 gcc_assert ((root
->offset
% BITS_PER_UNIT
) == 0
2120 && (root
->size
% BITS_PER_UNIT
) == 0);
2121 root
->type
= build_nonstandard_integer_type (root
->size
,
2122 TYPE_UNSIGNED (rt
));
2123 root
->expr
= build_ref_for_offset (UNKNOWN_LOCATION
,
2124 root
->base
, root
->offset
,
2125 root
->type
, NULL
, false);
2126 new_integer_type
= true;
2129 new_integer_type
= false;
2131 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2133 fprintf (dump_file
, "Marking ");
2134 print_generic_expr (dump_file
, root
->base
, 0);
2135 fprintf (dump_file
, " offset: %u, size: %u ",
2136 (unsigned) root
->offset
, (unsigned) root
->size
);
2137 fprintf (dump_file
, " to be replaced%s.\n",
2138 new_integer_type
? " with an integer": "");
2141 root
->grp_to_be_replaced
= 1;
2147 if (covered_to
< limit
)
2150 root
->grp_total_scalarization
= 0;
2154 && (!hole
|| root
->grp_total_scalarization
))
2156 root
->grp_covered
= 1;
2159 if (root
->grp_write
|| TREE_CODE (root
->base
) == PARM_DECL
)
2160 root
->grp_unscalarized_data
= 1; /* not covered and written to */
2166 /* Analyze all access trees linked by next_grp by the means of
2167 analyze_access_subtree. */
2169 analyze_access_trees (struct access
*access
)
2175 if (analyze_access_subtree (access
, NULL
, true))
2177 access
= access
->next_grp
;
2183 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2184 SIZE would conflict with an already existing one. If exactly such a child
2185 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2188 child_would_conflict_in_lacc (struct access
*lacc
, HOST_WIDE_INT norm_offset
,
2189 HOST_WIDE_INT size
, struct access
**exact_match
)
2191 struct access
*child
;
2193 for (child
= lacc
->first_child
; child
; child
= child
->next_sibling
)
2195 if (child
->offset
== norm_offset
&& child
->size
== size
)
2197 *exact_match
= child
;
2201 if (child
->offset
< norm_offset
+ size
2202 && child
->offset
+ child
->size
> norm_offset
)
2209 /* Create a new child access of PARENT, with all properties just like MODEL
2210 except for its offset and with its grp_write false and grp_read true.
2211 Return the new access or NULL if it cannot be created. Note that this access
2212 is created long after all splicing and sorting, it's not located in any
2213 access vector and is automatically a representative of its group. */
2215 static struct access
*
2216 create_artificial_child_access (struct access
*parent
, struct access
*model
,
2217 HOST_WIDE_INT new_offset
)
2219 struct access
*access
;
2220 struct access
**child
;
2221 tree expr
= parent
->base
;
2223 gcc_assert (!model
->grp_unscalarizable_region
);
2225 access
= (struct access
*) pool_alloc (access_pool
);
2226 memset (access
, 0, sizeof (struct access
));
2227 if (!build_user_friendly_ref_for_offset (&expr
, TREE_TYPE (expr
), new_offset
,
2230 access
->grp_no_warning
= true;
2231 expr
= build_ref_for_model (EXPR_LOCATION (parent
->base
), parent
->base
,
2232 new_offset
, model
, NULL
, false);
2235 access
->base
= parent
->base
;
2236 access
->expr
= expr
;
2237 access
->offset
= new_offset
;
2238 access
->size
= model
->size
;
2239 access
->type
= model
->type
;
2240 access
->grp_write
= true;
2241 access
->grp_read
= false;
2243 child
= &parent
->first_child
;
2244 while (*child
&& (*child
)->offset
< new_offset
)
2245 child
= &(*child
)->next_sibling
;
2247 access
->next_sibling
= *child
;
2254 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2255 true if any new subaccess was created. Additionally, if RACC is a scalar
2256 access but LACC is not, change the type of the latter, if possible. */
2259 propagate_subaccesses_across_link (struct access
*lacc
, struct access
*racc
)
2261 struct access
*rchild
;
2262 HOST_WIDE_INT norm_delta
= lacc
->offset
- racc
->offset
;
2265 if (is_gimple_reg_type (lacc
->type
)
2266 || lacc
->grp_unscalarizable_region
2267 || racc
->grp_unscalarizable_region
)
2270 if (is_gimple_reg_type (racc
->type
))
2272 if (!lacc
->first_child
&& !racc
->first_child
)
2274 tree t
= lacc
->base
;
2276 lacc
->type
= racc
->type
;
2277 if (build_user_friendly_ref_for_offset (&t
, TREE_TYPE (t
),
2278 lacc
->offset
, racc
->type
))
2282 lacc
->expr
= build_ref_for_model (EXPR_LOCATION (lacc
->base
),
2283 lacc
->base
, lacc
->offset
,
2285 lacc
->grp_no_warning
= true;
2291 for (rchild
= racc
->first_child
; rchild
; rchild
= rchild
->next_sibling
)
2293 struct access
*new_acc
= NULL
;
2294 HOST_WIDE_INT norm_offset
= rchild
->offset
+ norm_delta
;
2296 if (rchild
->grp_unscalarizable_region
)
2299 if (child_would_conflict_in_lacc (lacc
, norm_offset
, rchild
->size
,
2304 rchild
->grp_hint
= 1;
2305 new_acc
->grp_hint
|= new_acc
->grp_read
;
2306 if (rchild
->first_child
)
2307 ret
|= propagate_subaccesses_across_link (new_acc
, rchild
);
2312 rchild
->grp_hint
= 1;
2313 new_acc
= create_artificial_child_access (lacc
, rchild
, norm_offset
);
2317 if (racc
->first_child
)
2318 propagate_subaccesses_across_link (new_acc
, rchild
);
2325 /* Propagate all subaccesses across assignment links. */
2328 propagate_all_subaccesses (void)
2330 while (work_queue_head
)
2332 struct access
*racc
= pop_access_from_work_queue ();
2333 struct assign_link
*link
;
2335 gcc_assert (racc
->first_link
);
2337 for (link
= racc
->first_link
; link
; link
= link
->next
)
2339 struct access
*lacc
= link
->lacc
;
2341 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (lacc
->base
)))
2343 lacc
= lacc
->group_representative
;
2344 if (propagate_subaccesses_across_link (lacc
, racc
)
2345 && lacc
->first_link
)
2346 add_access_to_work_queue (lacc
);
2351 /* Go through all accesses collected throughout the (intraprocedural) analysis
2352 stage, exclude overlapping ones, identify representatives and build trees
2353 out of them, making decisions about scalarization on the way. Return true
2354 iff there are any to-be-scalarized variables after this stage. */
2357 analyze_all_variable_accesses (void)
2360 bitmap tmp
= BITMAP_ALLOC (NULL
);
2362 unsigned i
, max_total_scalarization_size
;
2364 max_total_scalarization_size
= UNITS_PER_WORD
* BITS_PER_UNIT
2365 * MOVE_RATIO (optimize_function_for_speed_p (cfun
));
2367 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap
, 0, i
, bi
)
2368 if (bitmap_bit_p (should_scalarize_away_bitmap
, i
)
2369 && !bitmap_bit_p (cannot_scalarize_away_bitmap
, i
))
2371 tree var
= candidate (i
);
2373 if (TREE_CODE (var
) == VAR_DECL
2374 && type_consists_of_records_p (TREE_TYPE (var
)))
2376 if ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var
)), 1)
2377 <= max_total_scalarization_size
)
2379 completely_scalarize_var (var
);
2380 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2382 fprintf (dump_file
, "Will attempt to totally scalarize ");
2383 print_generic_expr (dump_file
, var
, 0);
2384 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (var
));
2387 else if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2389 fprintf (dump_file
, "Too big to totally scalarize: ");
2390 print_generic_expr (dump_file
, var
, 0);
2391 fprintf (dump_file
, " (UID: %u)\n", DECL_UID (var
));
2396 bitmap_copy (tmp
, candidate_bitmap
);
2397 EXECUTE_IF_SET_IN_BITMAP (tmp
, 0, i
, bi
)
2399 tree var
= candidate (i
);
2400 struct access
*access
;
2402 access
= sort_and_splice_var_accesses (var
);
2403 if (!access
|| !build_access_trees (access
))
2404 disqualify_candidate (var
,
2405 "No or inhibitingly overlapping accesses.");
2408 propagate_all_subaccesses ();
2410 bitmap_copy (tmp
, candidate_bitmap
);
2411 EXECUTE_IF_SET_IN_BITMAP (tmp
, 0, i
, bi
)
2413 tree var
= candidate (i
);
2414 struct access
*access
= get_first_repr_for_decl (var
);
2416 if (analyze_access_trees (access
))
2419 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2421 fprintf (dump_file
, "\nAccess trees for ");
2422 print_generic_expr (dump_file
, var
, 0);
2423 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (var
));
2424 dump_access_tree (dump_file
, access
);
2425 fprintf (dump_file
, "\n");
2429 disqualify_candidate (var
, "No scalar replacements to be created.");
2436 statistics_counter_event (cfun
, "Scalarized aggregates", res
);
2443 /* Generate statements copying scalar replacements of accesses within a subtree
2444 into or out of AGG. ACCESS, all its children, siblings and their children
2445 are to be processed. AGG is an aggregate type expression (can be a
2446 declaration but does not have to be, it can for example also be a mem_ref or
2447 a series of handled components). TOP_OFFSET is the offset of the processed
2448 subtree which has to be subtracted from offsets of individual accesses to
2449 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2450 replacements in the interval <start_offset, start_offset + chunk_size>,
2451 otherwise copy all. GSI is a statement iterator used to place the new
2452 statements. WRITE should be true when the statements should write from AGG
2453 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2454 statements will be added after the current statement in GSI, they will be
2455 added before the statement otherwise. */
2458 generate_subtree_copies (struct access
*access
, tree agg
,
2459 HOST_WIDE_INT top_offset
,
2460 HOST_WIDE_INT start_offset
, HOST_WIDE_INT chunk_size
,
2461 gimple_stmt_iterator
*gsi
, bool write
,
2462 bool insert_after
, location_t loc
)
2466 if (chunk_size
&& access
->offset
>= start_offset
+ chunk_size
)
2469 if (access
->grp_to_be_replaced
2471 || access
->offset
+ access
->size
> start_offset
))
2473 tree expr
, repl
= get_access_replacement (access
);
2476 expr
= build_ref_for_model (loc
, agg
, access
->offset
- top_offset
,
2477 access
, gsi
, insert_after
);
2481 if (access
->grp_partial_lhs
)
2482 expr
= force_gimple_operand_gsi (gsi
, expr
, true, NULL_TREE
,
2484 insert_after
? GSI_NEW_STMT
2486 stmt
= gimple_build_assign (repl
, expr
);
2490 TREE_NO_WARNING (repl
) = 1;
2491 if (access
->grp_partial_lhs
)
2492 repl
= force_gimple_operand_gsi (gsi
, repl
, true, NULL_TREE
,
2494 insert_after
? GSI_NEW_STMT
2496 stmt
= gimple_build_assign (expr
, repl
);
2498 gimple_set_location (stmt
, loc
);
2501 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2503 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2505 sra_stats
.subtree_copies
++;
2508 if (access
->first_child
)
2509 generate_subtree_copies (access
->first_child
, agg
, top_offset
,
2510 start_offset
, chunk_size
, gsi
,
2511 write
, insert_after
, loc
);
2513 access
= access
->next_sibling
;
2518 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2519 the root of the subtree to be processed. GSI is the statement iterator used
2520 for inserting statements which are added after the current statement if
2521 INSERT_AFTER is true or before it otherwise. */
2524 init_subtree_with_zero (struct access
*access
, gimple_stmt_iterator
*gsi
,
2525 bool insert_after
, location_t loc
)
2528 struct access
*child
;
2530 if (access
->grp_to_be_replaced
)
2534 stmt
= gimple_build_assign (get_access_replacement (access
),
2535 build_zero_cst (access
->type
));
2537 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2539 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2541 gimple_set_location (stmt
, loc
);
2544 for (child
= access
->first_child
; child
; child
= child
->next_sibling
)
2545 init_subtree_with_zero (child
, gsi
, insert_after
, loc
);
2548 /* Search for an access representative for the given expression EXPR and
2549 return it or NULL if it cannot be found. */
2551 static struct access
*
2552 get_access_for_expr (tree expr
)
2554 HOST_WIDE_INT offset
, size
, max_size
;
2557 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2558 a different size than the size of its argument and we need the latter
2560 if (TREE_CODE (expr
) == VIEW_CONVERT_EXPR
)
2561 expr
= TREE_OPERAND (expr
, 0);
2563 base
= get_ref_base_and_extent (expr
, &offset
, &size
, &max_size
);
2564 if (max_size
== -1 || !DECL_P (base
))
2567 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (base
)))
2570 return get_var_base_offset_size_access (base
, offset
, max_size
);
2573 /* Replace the expression EXPR with a scalar replacement if there is one and
2574 generate other statements to do type conversion or subtree copying if
2575 necessary. GSI is used to place newly created statements, WRITE is true if
2576 the expression is being written to (it is on a LHS of a statement or output
2577 in an assembly statement). */
2580 sra_modify_expr (tree
*expr
, gimple_stmt_iterator
*gsi
, bool write
)
2583 struct access
*access
;
2586 if (TREE_CODE (*expr
) == BIT_FIELD_REF
)
2589 expr
= &TREE_OPERAND (*expr
, 0);
2594 if (TREE_CODE (*expr
) == REALPART_EXPR
|| TREE_CODE (*expr
) == IMAGPART_EXPR
)
2595 expr
= &TREE_OPERAND (*expr
, 0);
2596 access
= get_access_for_expr (*expr
);
2599 type
= TREE_TYPE (*expr
);
2601 loc
= gimple_location (gsi_stmt (*gsi
));
2602 if (access
->grp_to_be_replaced
)
2604 tree repl
= get_access_replacement (access
);
2605 /* If we replace a non-register typed access simply use the original
2606 access expression to extract the scalar component afterwards.
2607 This happens if scalarizing a function return value or parameter
2608 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2609 gcc.c-torture/compile/20011217-1.c.
2611 We also want to use this when accessing a complex or vector which can
2612 be accessed as a different type too, potentially creating a need for
2613 type conversion (see PR42196) and when scalarized unions are involved
2614 in assembler statements (see PR42398). */
2615 if (!useless_type_conversion_p (type
, access
->type
))
2619 ref
= build_ref_for_model (loc
, access
->base
, access
->offset
, access
,
2626 if (access
->grp_partial_lhs
)
2627 ref
= force_gimple_operand_gsi (gsi
, ref
, true, NULL_TREE
,
2628 false, GSI_NEW_STMT
);
2629 stmt
= gimple_build_assign (repl
, ref
);
2630 gimple_set_location (stmt
, loc
);
2631 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2637 if (access
->grp_partial_lhs
)
2638 repl
= force_gimple_operand_gsi (gsi
, repl
, true, NULL_TREE
,
2639 true, GSI_SAME_STMT
);
2640 stmt
= gimple_build_assign (ref
, repl
);
2641 gimple_set_location (stmt
, loc
);
2642 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2650 if (access
->first_child
)
2652 HOST_WIDE_INT start_offset
, chunk_size
;
2654 && host_integerp (TREE_OPERAND (bfr
, 1), 1)
2655 && host_integerp (TREE_OPERAND (bfr
, 2), 1))
2657 chunk_size
= tree_low_cst (TREE_OPERAND (bfr
, 1), 1);
2658 start_offset
= access
->offset
2659 + tree_low_cst (TREE_OPERAND (bfr
, 2), 1);
2662 start_offset
= chunk_size
= 0;
2664 generate_subtree_copies (access
->first_child
, access
->base
, 0,
2665 start_offset
, chunk_size
, gsi
, write
, write
,
2671 /* Where scalar replacements of the RHS have been written to when a replacement
2672 of a LHS of an assigments cannot be direclty loaded from a replacement of
2674 enum unscalarized_data_handling
{ SRA_UDH_NONE
, /* Nothing done so far. */
2675 SRA_UDH_RIGHT
, /* Data flushed to the RHS. */
2676 SRA_UDH_LEFT
}; /* Data flushed to the LHS. */
2678 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2679 base aggregate if there are unscalarized data or directly to LHS of the
2680 statement that is pointed to by GSI otherwise. */
2682 static enum unscalarized_data_handling
2683 handle_unscalarized_data_in_subtree (struct access
*top_racc
,
2684 gimple_stmt_iterator
*gsi
)
2686 if (top_racc
->grp_unscalarized_data
)
2688 generate_subtree_copies (top_racc
->first_child
, top_racc
->base
, 0, 0, 0,
2690 gimple_location (gsi_stmt (*gsi
)));
2691 return SRA_UDH_RIGHT
;
2695 tree lhs
= gimple_assign_lhs (gsi_stmt (*gsi
));
2696 generate_subtree_copies (top_racc
->first_child
, lhs
, top_racc
->offset
,
2697 0, 0, gsi
, false, false,
2698 gimple_location (gsi_stmt (*gsi
)));
2699 return SRA_UDH_LEFT
;
2704 /* Try to generate statements to load all sub-replacements in an access subtree
2705 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2706 If that is not possible, refresh the TOP_RACC base aggregate and load the
2707 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2708 copied. NEW_GSI is stmt iterator used for statement insertions after the
2709 original assignment, OLD_GSI is used to insert statements before the
2710 assignment. *REFRESHED keeps the information whether we have needed to
2711 refresh replacements of the LHS and from which side of the assignments this
2715 load_assign_lhs_subreplacements (struct access
*lacc
, struct access
*top_racc
,
2716 HOST_WIDE_INT left_offset
,
2717 gimple_stmt_iterator
*old_gsi
,
2718 gimple_stmt_iterator
*new_gsi
,
2719 enum unscalarized_data_handling
*refreshed
)
2721 location_t loc
= gimple_location (gsi_stmt (*old_gsi
));
2722 for (lacc
= lacc
->first_child
; lacc
; lacc
= lacc
->next_sibling
)
2724 if (lacc
->grp_to_be_replaced
)
2726 struct access
*racc
;
2727 HOST_WIDE_INT offset
= lacc
->offset
- left_offset
+ top_racc
->offset
;
2731 racc
= find_access_in_subtree (top_racc
, offset
, lacc
->size
);
2732 if (racc
&& racc
->grp_to_be_replaced
)
2734 rhs
= get_access_replacement (racc
);
2735 if (!useless_type_conversion_p (lacc
->type
, racc
->type
))
2736 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
, lacc
->type
, rhs
);
2738 if (racc
->grp_partial_lhs
&& lacc
->grp_partial_lhs
)
2739 rhs
= force_gimple_operand_gsi (old_gsi
, rhs
, true, NULL_TREE
,
2740 true, GSI_SAME_STMT
);
2744 /* No suitable access on the right hand side, need to load from
2745 the aggregate. See if we have to update it first... */
2746 if (*refreshed
== SRA_UDH_NONE
)
2747 *refreshed
= handle_unscalarized_data_in_subtree (top_racc
,
2750 if (*refreshed
== SRA_UDH_LEFT
)
2751 rhs
= build_ref_for_model (loc
, lacc
->base
, lacc
->offset
, lacc
,
2754 rhs
= build_ref_for_model (loc
, top_racc
->base
, offset
, lacc
,
2756 if (lacc
->grp_partial_lhs
)
2757 rhs
= force_gimple_operand_gsi (new_gsi
, rhs
, true, NULL_TREE
,
2758 false, GSI_NEW_STMT
);
2761 stmt
= gimple_build_assign (get_access_replacement (lacc
), rhs
);
2762 gsi_insert_after (new_gsi
, stmt
, GSI_NEW_STMT
);
2763 gimple_set_location (stmt
, loc
);
2765 sra_stats
.subreplacements
++;
2767 else if (*refreshed
== SRA_UDH_NONE
2768 && lacc
->grp_read
&& !lacc
->grp_covered
)
2769 *refreshed
= handle_unscalarized_data_in_subtree (top_racc
,
2772 if (lacc
->first_child
)
2773 load_assign_lhs_subreplacements (lacc
, top_racc
, left_offset
,
2774 old_gsi
, new_gsi
, refreshed
);
2778 /* Result code for SRA assignment modification. */
2779 enum assignment_mod_result
{ SRA_AM_NONE
, /* nothing done for the stmt */
2780 SRA_AM_MODIFIED
, /* stmt changed but not
2782 SRA_AM_REMOVED
}; /* stmt eliminated */
2784 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2785 to the assignment and GSI is the statement iterator pointing at it. Returns
2786 the same values as sra_modify_assign. */
2788 static enum assignment_mod_result
2789 sra_modify_constructor_assign (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
2791 tree lhs
= gimple_assign_lhs (*stmt
);
2795 acc
= get_access_for_expr (lhs
);
2799 if (gimple_clobber_p (*stmt
))
2801 /* Remove clobbers of fully scalarized variables, otherwise
2803 if (acc
->grp_covered
)
2805 unlink_stmt_vdef (*stmt
);
2806 gsi_remove (gsi
, true);
2807 release_defs (*stmt
);
2808 return SRA_AM_REMOVED
;
2814 loc
= gimple_location (*stmt
);
2815 if (VEC_length (constructor_elt
,
2816 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt
))) > 0)
2818 /* I have never seen this code path trigger but if it can happen the
2819 following should handle it gracefully. */
2820 if (access_has_children_p (acc
))
2821 generate_subtree_copies (acc
->first_child
, acc
->base
, 0, 0, 0, gsi
,
2823 return SRA_AM_MODIFIED
;
2826 if (acc
->grp_covered
)
2828 init_subtree_with_zero (acc
, gsi
, false, loc
);
2829 unlink_stmt_vdef (*stmt
);
2830 gsi_remove (gsi
, true);
2831 release_defs (*stmt
);
2832 return SRA_AM_REMOVED
;
2836 init_subtree_with_zero (acc
, gsi
, true, loc
);
2837 return SRA_AM_MODIFIED
;
2841 /* Create and return a new suitable default definition SSA_NAME for RACC which
2842 is an access describing an uninitialized part of an aggregate that is being
2846 get_repl_default_def_ssa_name (struct access
*racc
)
2848 return get_or_create_ssa_default_def (cfun
, get_access_replacement (racc
));
2851 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2855 contains_bitfld_comp_ref_p (const_tree ref
)
2857 while (handled_component_p (ref
))
2859 if (TREE_CODE (ref
) == COMPONENT_REF
2860 && DECL_BIT_FIELD (TREE_OPERAND (ref
, 1)))
2862 ref
= TREE_OPERAND (ref
, 0);
2868 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2869 bit-field field declaration somewhere in it. */
2872 contains_vce_or_bfcref_p (const_tree ref
)
2874 while (handled_component_p (ref
))
2876 if (TREE_CODE (ref
) == VIEW_CONVERT_EXPR
2877 || (TREE_CODE (ref
) == COMPONENT_REF
2878 && DECL_BIT_FIELD (TREE_OPERAND (ref
, 1))))
2880 ref
= TREE_OPERAND (ref
, 0);
2886 /* Examine both sides of the assignment statement pointed to by STMT, replace
2887 them with a scalare replacement if there is one and generate copying of
2888 replacements if scalarized aggregates have been used in the assignment. GSI
2889 is used to hold generated statements for type conversions and subtree
2892 static enum assignment_mod_result
2893 sra_modify_assign (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
2895 struct access
*lacc
, *racc
;
2897 bool modify_this_stmt
= false;
2898 bool force_gimple_rhs
= false;
2900 gimple_stmt_iterator orig_gsi
= *gsi
;
2902 if (!gimple_assign_single_p (*stmt
))
2904 lhs
= gimple_assign_lhs (*stmt
);
2905 rhs
= gimple_assign_rhs1 (*stmt
);
2907 if (TREE_CODE (rhs
) == CONSTRUCTOR
)
2908 return sra_modify_constructor_assign (stmt
, gsi
);
2910 if (TREE_CODE (rhs
) == REALPART_EXPR
|| TREE_CODE (lhs
) == REALPART_EXPR
2911 || TREE_CODE (rhs
) == IMAGPART_EXPR
|| TREE_CODE (lhs
) == IMAGPART_EXPR
2912 || TREE_CODE (rhs
) == BIT_FIELD_REF
|| TREE_CODE (lhs
) == BIT_FIELD_REF
)
2914 modify_this_stmt
= sra_modify_expr (gimple_assign_rhs1_ptr (*stmt
),
2916 modify_this_stmt
|= sra_modify_expr (gimple_assign_lhs_ptr (*stmt
),
2918 return modify_this_stmt
? SRA_AM_MODIFIED
: SRA_AM_NONE
;
2921 lacc
= get_access_for_expr (lhs
);
2922 racc
= get_access_for_expr (rhs
);
2926 loc
= gimple_location (*stmt
);
2927 if (lacc
&& lacc
->grp_to_be_replaced
)
2929 lhs
= get_access_replacement (lacc
);
2930 gimple_assign_set_lhs (*stmt
, lhs
);
2931 modify_this_stmt
= true;
2932 if (lacc
->grp_partial_lhs
)
2933 force_gimple_rhs
= true;
2937 if (racc
&& racc
->grp_to_be_replaced
)
2939 rhs
= get_access_replacement (racc
);
2940 modify_this_stmt
= true;
2941 if (racc
->grp_partial_lhs
)
2942 force_gimple_rhs
= true;
2946 && !racc
->grp_unscalarized_data
2947 && TREE_CODE (lhs
) == SSA_NAME
2948 && !access_has_replacements_p (racc
))
2950 rhs
= get_repl_default_def_ssa_name (racc
);
2951 modify_this_stmt
= true;
2955 if (modify_this_stmt
)
2957 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
2959 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2960 ??? This should move to fold_stmt which we simply should
2961 call after building a VIEW_CONVERT_EXPR here. */
2962 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs
))
2963 && !contains_bitfld_comp_ref_p (lhs
)
2964 && !access_has_children_p (lacc
))
2966 lhs
= build_ref_for_model (loc
, lhs
, 0, racc
, gsi
, false);
2967 gimple_assign_set_lhs (*stmt
, lhs
);
2969 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs
))
2970 && !contains_vce_or_bfcref_p (rhs
)
2971 && !access_has_children_p (racc
))
2972 rhs
= build_ref_for_model (loc
, rhs
, 0, lacc
, gsi
, false);
2974 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
2976 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
, TREE_TYPE (lhs
),
2978 if (is_gimple_reg_type (TREE_TYPE (lhs
))
2979 && TREE_CODE (lhs
) != SSA_NAME
)
2980 force_gimple_rhs
= true;
2985 /* From this point on, the function deals with assignments in between
2986 aggregates when at least one has scalar reductions of some of its
2987 components. There are three possible scenarios: Both the LHS and RHS have
2988 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2990 In the first case, we would like to load the LHS components from RHS
2991 components whenever possible. If that is not possible, we would like to
2992 read it directly from the RHS (after updating it by storing in it its own
2993 components). If there are some necessary unscalarized data in the LHS,
2994 those will be loaded by the original assignment too. If neither of these
2995 cases happen, the original statement can be removed. Most of this is done
2996 by load_assign_lhs_subreplacements.
2998 In the second case, we would like to store all RHS scalarized components
2999 directly into LHS and if they cover the aggregate completely, remove the
3000 statement too. In the third case, we want the LHS components to be loaded
3001 directly from the RHS (DSE will remove the original statement if it
3004 This is a bit complex but manageable when types match and when unions do
3005 not cause confusion in a way that we cannot really load a component of LHS
3006 from the RHS or vice versa (the access representing this level can have
3007 subaccesses that are accessible only through a different union field at a
3008 higher level - different from the one used in the examined expression).
3011 Therefore, I specially handle a fourth case, happening when there is a
3012 specific type cast or it is impossible to locate a scalarized subaccess on
3013 the other side of the expression. If that happens, I simply "refresh" the
3014 RHS by storing in it is scalarized components leave the original statement
3015 there to do the copying and then load the scalar replacements of the LHS.
3016 This is what the first branch does. */
3018 if (modify_this_stmt
3019 || gimple_has_volatile_ops (*stmt
)
3020 || contains_vce_or_bfcref_p (rhs
)
3021 || contains_vce_or_bfcref_p (lhs
))
3023 if (access_has_children_p (racc
))
3024 generate_subtree_copies (racc
->first_child
, racc
->base
, 0, 0, 0,
3025 gsi
, false, false, loc
);
3026 if (access_has_children_p (lacc
))
3027 generate_subtree_copies (lacc
->first_child
, lacc
->base
, 0, 0, 0,
3028 gsi
, true, true, loc
);
3029 sra_stats
.separate_lhs_rhs_handling
++;
3031 /* This gimplification must be done after generate_subtree_copies,
3032 lest we insert the subtree copies in the middle of the gimplified
3034 if (force_gimple_rhs
)
3035 rhs
= force_gimple_operand_gsi (&orig_gsi
, rhs
, true, NULL_TREE
,
3036 true, GSI_SAME_STMT
);
3037 if (gimple_assign_rhs1 (*stmt
) != rhs
)
3039 modify_this_stmt
= true;
3040 gimple_assign_set_rhs_from_tree (&orig_gsi
, rhs
);
3041 gcc_assert (*stmt
== gsi_stmt (orig_gsi
));
3044 return modify_this_stmt
? SRA_AM_MODIFIED
: SRA_AM_NONE
;
3048 if (access_has_children_p (lacc
)
3049 && access_has_children_p (racc
)
3050 /* When an access represents an unscalarizable region, it usually
3051 represents accesses with variable offset and thus must not be used
3052 to generate new memory accesses. */
3053 && !lacc
->grp_unscalarizable_region
3054 && !racc
->grp_unscalarizable_region
)
3056 gimple_stmt_iterator orig_gsi
= *gsi
;
3057 enum unscalarized_data_handling refreshed
;
3059 if (lacc
->grp_read
&& !lacc
->grp_covered
)
3060 refreshed
= handle_unscalarized_data_in_subtree (racc
, gsi
);
3062 refreshed
= SRA_UDH_NONE
;
3064 load_assign_lhs_subreplacements (lacc
, racc
, lacc
->offset
,
3065 &orig_gsi
, gsi
, &refreshed
);
3066 if (refreshed
!= SRA_UDH_RIGHT
)
3069 unlink_stmt_vdef (*stmt
);
3070 gsi_remove (&orig_gsi
, true);
3071 release_defs (*stmt
);
3072 sra_stats
.deleted
++;
3073 return SRA_AM_REMOVED
;
3078 if (access_has_children_p (racc
)
3079 && !racc
->grp_unscalarized_data
)
3083 fprintf (dump_file
, "Removing load: ");
3084 print_gimple_stmt (dump_file
, *stmt
, 0, 0);
3086 generate_subtree_copies (racc
->first_child
, lhs
,
3087 racc
->offset
, 0, 0, gsi
,
3089 gcc_assert (*stmt
== gsi_stmt (*gsi
));
3090 unlink_stmt_vdef (*stmt
);
3091 gsi_remove (gsi
, true);
3092 release_defs (*stmt
);
3093 sra_stats
.deleted
++;
3094 return SRA_AM_REMOVED
;
3096 /* Restore the aggregate RHS from its components so the
3097 prevailing aggregate copy does the right thing. */
3098 if (access_has_children_p (racc
))
3099 generate_subtree_copies (racc
->first_child
, racc
->base
, 0, 0, 0,
3100 gsi
, false, false, loc
);
3101 /* Re-load the components of the aggregate copy destination.
3102 But use the RHS aggregate to load from to expose more
3103 optimization opportunities. */
3104 if (access_has_children_p (lacc
))
3105 generate_subtree_copies (lacc
->first_child
, rhs
, lacc
->offset
,
3106 0, 0, gsi
, true, true, loc
);
3113 /* Traverse the function body and all modifications as decided in
3114 analyze_all_variable_accesses. Return true iff the CFG has been
3118 sra_modify_function_body (void)
3120 bool cfg_changed
= false;
3125 gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
3126 while (!gsi_end_p (gsi
))
3128 gimple stmt
= gsi_stmt (gsi
);
3129 enum assignment_mod_result assign_result
;
3130 bool modified
= false, deleted
= false;
3134 switch (gimple_code (stmt
))
3137 t
= gimple_return_retval_ptr (stmt
);
3138 if (*t
!= NULL_TREE
)
3139 modified
|= sra_modify_expr (t
, &gsi
, false);
3143 assign_result
= sra_modify_assign (&stmt
, &gsi
);
3144 modified
|= assign_result
== SRA_AM_MODIFIED
;
3145 deleted
= assign_result
== SRA_AM_REMOVED
;
3149 /* Operands must be processed before the lhs. */
3150 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
3152 t
= gimple_call_arg_ptr (stmt
, i
);
3153 modified
|= sra_modify_expr (t
, &gsi
, false);
3156 if (gimple_call_lhs (stmt
))
3158 t
= gimple_call_lhs_ptr (stmt
);
3159 modified
|= sra_modify_expr (t
, &gsi
, true);
3164 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
3166 t
= &TREE_VALUE (gimple_asm_input_op (stmt
, i
));
3167 modified
|= sra_modify_expr (t
, &gsi
, false);
3169 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
3171 t
= &TREE_VALUE (gimple_asm_output_op (stmt
, i
));
3172 modified
|= sra_modify_expr (t
, &gsi
, true);
3183 if (maybe_clean_eh_stmt (stmt
)
3184 && gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
3195 /* Generate statements initializing scalar replacements of parts of function
3199 initialize_parameter_reductions (void)
3201 gimple_stmt_iterator gsi
;
3202 gimple_seq seq
= NULL
;
3205 gsi
= gsi_start (seq
);
3206 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3208 parm
= DECL_CHAIN (parm
))
3210 VEC (access_p
, heap
) *access_vec
;
3211 struct access
*access
;
3213 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
3215 access_vec
= get_base_access_vector (parm
);
3219 for (access
= VEC_index (access_p
, access_vec
, 0);
3221 access
= access
->next_grp
)
3222 generate_subtree_copies (access
, parm
, 0, 0, 0, &gsi
, true, true,
3223 EXPR_LOCATION (parm
));
3226 seq
= gsi_seq (gsi
);
3228 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR
), seq
);
3231 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3232 it reveals there are components of some aggregates to be scalarized, it runs
3233 the required transformations. */
3235 perform_intra_sra (void)
3240 if (!find_var_candidates ())
3243 if (!scan_function ())
3246 if (!analyze_all_variable_accesses ())
3249 if (sra_modify_function_body ())
3250 ret
= TODO_update_ssa
| TODO_cleanup_cfg
;
3252 ret
= TODO_update_ssa
;
3253 initialize_parameter_reductions ();
3255 statistics_counter_event (cfun
, "Scalar replacements created",
3256 sra_stats
.replacements
);
3257 statistics_counter_event (cfun
, "Modified expressions", sra_stats
.exprs
);
3258 statistics_counter_event (cfun
, "Subtree copy stmts",
3259 sra_stats
.subtree_copies
);
3260 statistics_counter_event (cfun
, "Subreplacement stmts",
3261 sra_stats
.subreplacements
);
3262 statistics_counter_event (cfun
, "Deleted stmts", sra_stats
.deleted
);
3263 statistics_counter_event (cfun
, "Separate LHS and RHS handling",
3264 sra_stats
.separate_lhs_rhs_handling
);
3267 sra_deinitialize ();
3271 /* Perform early intraprocedural SRA. */
3273 early_intra_sra (void)
3275 sra_mode
= SRA_MODE_EARLY_INTRA
;
3276 return perform_intra_sra ();
3279 /* Perform "late" intraprocedural SRA. */
3281 late_intra_sra (void)
3283 sra_mode
= SRA_MODE_INTRA
;
3284 return perform_intra_sra ();
3289 gate_intra_sra (void)
3291 return flag_tree_sra
!= 0 && dbg_cnt (tree_sra
);
3295 struct gimple_opt_pass pass_sra_early
=
3300 gate_intra_sra
, /* gate */
3301 early_intra_sra
, /* execute */
3304 0, /* static_pass_number */
3305 TV_TREE_SRA
, /* tv_id */
3306 PROP_cfg
| PROP_ssa
, /* properties_required */
3307 0, /* properties_provided */
3308 0, /* properties_destroyed */
3309 0, /* todo_flags_start */
3312 | TODO_verify_ssa
/* todo_flags_finish */
3316 struct gimple_opt_pass pass_sra
=
3321 gate_intra_sra
, /* gate */
3322 late_intra_sra
, /* execute */
3325 0, /* static_pass_number */
3326 TV_TREE_SRA
, /* tv_id */
3327 PROP_cfg
| PROP_ssa
, /* properties_required */
3328 0, /* properties_provided */
3329 0, /* properties_destroyed */
3330 TODO_update_address_taken
, /* todo_flags_start */
3333 | TODO_verify_ssa
/* todo_flags_finish */
3338 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3342 is_unused_scalar_param (tree parm
)
3345 return (is_gimple_reg (parm
)
3346 && (!(name
= ssa_default_def (cfun
, parm
))
3347 || has_zero_uses (name
)));
3350 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3351 examine whether there are any direct or otherwise infeasible ones. If so,
3352 return true, otherwise return false. PARM must be a gimple register with a
3353 non-NULL default definition. */
3356 ptr_parm_has_direct_uses (tree parm
)
3358 imm_use_iterator ui
;
3360 tree name
= ssa_default_def (cfun
, parm
);
3363 FOR_EACH_IMM_USE_STMT (stmt
, ui
, name
)
3366 use_operand_p use_p
;
3368 if (is_gimple_debug (stmt
))
3371 /* Valid uses include dereferences on the lhs and the rhs. */
3372 if (gimple_has_lhs (stmt
))
3374 tree lhs
= gimple_get_lhs (stmt
);
3375 while (handled_component_p (lhs
))
3376 lhs
= TREE_OPERAND (lhs
, 0);
3377 if (TREE_CODE (lhs
) == MEM_REF
3378 && TREE_OPERAND (lhs
, 0) == name
3379 && integer_zerop (TREE_OPERAND (lhs
, 1))
3380 && types_compatible_p (TREE_TYPE (lhs
),
3381 TREE_TYPE (TREE_TYPE (name
)))
3382 && !TREE_THIS_VOLATILE (lhs
))
3385 if (gimple_assign_single_p (stmt
))
3387 tree rhs
= gimple_assign_rhs1 (stmt
);
3388 while (handled_component_p (rhs
))
3389 rhs
= TREE_OPERAND (rhs
, 0);
3390 if (TREE_CODE (rhs
) == MEM_REF
3391 && TREE_OPERAND (rhs
, 0) == name
3392 && integer_zerop (TREE_OPERAND (rhs
, 1))
3393 && types_compatible_p (TREE_TYPE (rhs
),
3394 TREE_TYPE (TREE_TYPE (name
)))
3395 && !TREE_THIS_VOLATILE (rhs
))
3398 else if (is_gimple_call (stmt
))
3401 for (i
= 0; i
< gimple_call_num_args (stmt
); ++i
)
3403 tree arg
= gimple_call_arg (stmt
, i
);
3404 while (handled_component_p (arg
))
3405 arg
= TREE_OPERAND (arg
, 0);
3406 if (TREE_CODE (arg
) == MEM_REF
3407 && TREE_OPERAND (arg
, 0) == name
3408 && integer_zerop (TREE_OPERAND (arg
, 1))
3409 && types_compatible_p (TREE_TYPE (arg
),
3410 TREE_TYPE (TREE_TYPE (name
)))
3411 && !TREE_THIS_VOLATILE (arg
))
3416 /* If the number of valid uses does not match the number of
3417 uses in this stmt there is an unhandled use. */
3418 FOR_EACH_IMM_USE_ON_STMT (use_p
, ui
)
3425 BREAK_FROM_IMM_USE_STMT (ui
);
3431 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3432 them in candidate_bitmap. Note that these do not necessarily include
3433 parameter which are unused and thus can be removed. Return true iff any
3434 such candidate has been found. */
3437 find_param_candidates (void)
3444 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3446 parm
= DECL_CHAIN (parm
))
3448 tree type
= TREE_TYPE (parm
);
3453 if (TREE_THIS_VOLATILE (parm
)
3454 || TREE_ADDRESSABLE (parm
)
3455 || (!is_gimple_reg_type (type
) && is_va_list_type (type
)))
3458 if (is_unused_scalar_param (parm
))
3464 if (POINTER_TYPE_P (type
))
3466 type
= TREE_TYPE (type
);
3468 if (TREE_CODE (type
) == FUNCTION_TYPE
3469 || TYPE_VOLATILE (type
)
3470 || (TREE_CODE (type
) == ARRAY_TYPE
3471 && TYPE_NONALIASED_COMPONENT (type
))
3472 || !is_gimple_reg (parm
)
3473 || is_va_list_type (type
)
3474 || ptr_parm_has_direct_uses (parm
))
3477 else if (!AGGREGATE_TYPE_P (type
))
3480 if (!COMPLETE_TYPE_P (type
)
3481 || !host_integerp (TYPE_SIZE (type
), 1)
3482 || tree_low_cst (TYPE_SIZE (type
), 1) == 0
3483 || (AGGREGATE_TYPE_P (type
)
3484 && type_internals_preclude_sra_p (type
, &msg
)))
3487 bitmap_set_bit (candidate_bitmap
, DECL_UID (parm
));
3488 slot
= htab_find_slot_with_hash (candidates
, parm
,
3489 DECL_UID (parm
), INSERT
);
3490 *slot
= (void *) parm
;
3493 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3495 fprintf (dump_file
, "Candidate (%d): ", DECL_UID (parm
));
3496 print_generic_expr (dump_file
, parm
, 0);
3497 fprintf (dump_file
, "\n");
3501 func_param_count
= count
;
3505 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3509 mark_maybe_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
3512 struct access
*repr
= (struct access
*) data
;
3514 repr
->grp_maybe_modified
= 1;
3518 /* Analyze what representatives (in linked lists accessible from
3519 REPRESENTATIVES) can be modified by side effects of statements in the
3520 current function. */
3523 analyze_modified_params (VEC (access_p
, heap
) *representatives
)
3527 for (i
= 0; i
< func_param_count
; i
++)
3529 struct access
*repr
;
3531 for (repr
= VEC_index (access_p
, representatives
, i
);
3533 repr
= repr
->next_grp
)
3535 struct access
*access
;
3539 if (no_accesses_p (repr
))
3541 if (!POINTER_TYPE_P (TREE_TYPE (repr
->base
))
3542 || repr
->grp_maybe_modified
)
3545 ao_ref_init (&ar
, repr
->expr
);
3546 visited
= BITMAP_ALLOC (NULL
);
3547 for (access
= repr
; access
; access
= access
->next_sibling
)
3549 /* All accesses are read ones, otherwise grp_maybe_modified would
3550 be trivially set. */
3551 walk_aliased_vdefs (&ar
, gimple_vuse (access
->stmt
),
3552 mark_maybe_modified
, repr
, &visited
);
3553 if (repr
->grp_maybe_modified
)
3556 BITMAP_FREE (visited
);
3561 /* Propagate distances in bb_dereferences in the opposite direction than the
3562 control flow edges, in each step storing the maximum of the current value
3563 and the minimum of all successors. These steps are repeated until the table
3564 stabilizes. Note that BBs which might terminate the functions (according to
3565 final_bbs bitmap) never updated in this way. */
3568 propagate_dereference_distances (void)
3570 VEC (basic_block
, heap
) *queue
;
3573 queue
= VEC_alloc (basic_block
, heap
, last_basic_block_for_function (cfun
));
3574 VEC_quick_push (basic_block
, queue
, ENTRY_BLOCK_PTR
);
3577 VEC_quick_push (basic_block
, queue
, bb
);
3581 while (!VEC_empty (basic_block
, queue
))
3585 bool change
= false;
3588 bb
= VEC_pop (basic_block
, queue
);
3591 if (bitmap_bit_p (final_bbs
, bb
->index
))
3594 for (i
= 0; i
< func_param_count
; i
++)
3596 int idx
= bb
->index
* func_param_count
+ i
;
3598 HOST_WIDE_INT inh
= 0;
3600 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3602 int succ_idx
= e
->dest
->index
* func_param_count
+ i
;
3604 if (e
->src
== EXIT_BLOCK_PTR
)
3610 inh
= bb_dereferences
[succ_idx
];
3612 else if (bb_dereferences
[succ_idx
] < inh
)
3613 inh
= bb_dereferences
[succ_idx
];
3616 if (!first
&& bb_dereferences
[idx
] < inh
)
3618 bb_dereferences
[idx
] = inh
;
3623 if (change
&& !bitmap_bit_p (final_bbs
, bb
->index
))
3624 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3629 e
->src
->aux
= e
->src
;
3630 VEC_quick_push (basic_block
, queue
, e
->src
);
3634 VEC_free (basic_block
, heap
, queue
);
3637 /* Dump a dereferences TABLE with heading STR to file F. */
3640 dump_dereferences_table (FILE *f
, const char *str
, HOST_WIDE_INT
*table
)
3644 fprintf (dump_file
, str
);
3645 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
, next_bb
)
3647 fprintf (f
, "%4i %i ", bb
->index
, bitmap_bit_p (final_bbs
, bb
->index
));
3648 if (bb
!= EXIT_BLOCK_PTR
)
3651 for (i
= 0; i
< func_param_count
; i
++)
3653 int idx
= bb
->index
* func_param_count
+ i
;
3654 fprintf (f
, " %4" HOST_WIDE_INT_PRINT
"d", table
[idx
]);
3659 fprintf (dump_file
, "\n");
3662 /* Determine what (parts of) parameters passed by reference that are not
3663 assigned to are not certainly dereferenced in this function and thus the
3664 dereferencing cannot be safely moved to the caller without potentially
3665 introducing a segfault. Mark such REPRESENTATIVES as
3666 grp_not_necessarilly_dereferenced.
3668 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3669 part is calculated rather than simple booleans are calculated for each
3670 pointer parameter to handle cases when only a fraction of the whole
3671 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3674 The maximum dereference distances for each pointer parameter and BB are
3675 already stored in bb_dereference. This routine simply propagates these
3676 values upwards by propagate_dereference_distances and then compares the
3677 distances of individual parameters in the ENTRY BB to the equivalent
3678 distances of each representative of a (fraction of a) parameter. */
3681 analyze_caller_dereference_legality (VEC (access_p
, heap
) *representatives
)
3685 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3686 dump_dereferences_table (dump_file
,
3687 "Dereference table before propagation:\n",
3690 propagate_dereference_distances ();
3692 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3693 dump_dereferences_table (dump_file
,
3694 "Dereference table after propagation:\n",
3697 for (i
= 0; i
< func_param_count
; i
++)
3699 struct access
*repr
= VEC_index (access_p
, representatives
, i
);
3700 int idx
= ENTRY_BLOCK_PTR
->index
* func_param_count
+ i
;
3702 if (!repr
|| no_accesses_p (repr
))
3707 if ((repr
->offset
+ repr
->size
) > bb_dereferences
[idx
])
3708 repr
->grp_not_necessarilly_dereferenced
= 1;
3709 repr
= repr
->next_grp
;
3715 /* Return the representative access for the parameter declaration PARM if it is
3716 a scalar passed by reference which is not written to and the pointer value
3717 is not used directly. Thus, if it is legal to dereference it in the caller
3718 and we can rule out modifications through aliases, such parameter should be
3719 turned into one passed by value. Return NULL otherwise. */
3721 static struct access
*
3722 unmodified_by_ref_scalar_representative (tree parm
)
3724 int i
, access_count
;
3725 struct access
*repr
;
3726 VEC (access_p
, heap
) *access_vec
;
3728 access_vec
= get_base_access_vector (parm
);
3729 gcc_assert (access_vec
);
3730 repr
= VEC_index (access_p
, access_vec
, 0);
3733 repr
->group_representative
= repr
;
3735 access_count
= VEC_length (access_p
, access_vec
);
3736 for (i
= 1; i
< access_count
; i
++)
3738 struct access
*access
= VEC_index (access_p
, access_vec
, i
);
3741 access
->group_representative
= repr
;
3742 access
->next_sibling
= repr
->next_sibling
;
3743 repr
->next_sibling
= access
;
3747 repr
->grp_scalar_ptr
= 1;
3751 /* Return true iff this access precludes IPA-SRA of the parameter it is
3755 access_precludes_ipa_sra_p (struct access
*access
)
3757 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3758 is incompatible assign in a call statement (and possibly even in asm
3759 statements). This can be relaxed by using a new temporary but only for
3760 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3761 intraprocedural SRA we deal with this by keeping the old aggregate around,
3762 something we cannot do in IPA-SRA.) */
3764 && (is_gimple_call (access
->stmt
)
3765 || gimple_code (access
->stmt
) == GIMPLE_ASM
))
3772 /* Sort collected accesses for parameter PARM, identify representatives for
3773 each accessed region and link them together. Return NULL if there are
3774 different but overlapping accesses, return the special ptr value meaning
3775 there are no accesses for this parameter if that is the case and return the
3776 first representative otherwise. Set *RO_GRP if there is a group of accesses
3777 with only read (i.e. no write) accesses. */
3779 static struct access
*
3780 splice_param_accesses (tree parm
, bool *ro_grp
)
3782 int i
, j
, access_count
, group_count
;
3783 int agg_size
, total_size
= 0;
3784 struct access
*access
, *res
, **prev_acc_ptr
= &res
;
3785 VEC (access_p
, heap
) *access_vec
;
3787 access_vec
= get_base_access_vector (parm
);
3789 return &no_accesses_representant
;
3790 access_count
= VEC_length (access_p
, access_vec
);
3792 VEC_qsort (access_p
, access_vec
, compare_access_positions
);
3797 while (i
< access_count
)
3801 access
= VEC_index (access_p
, access_vec
, i
);
3802 modification
= access
->write
;
3803 if (access_precludes_ipa_sra_p (access
))
3805 a1_alias_type
= reference_alias_ptr_type (access
->expr
);
3807 /* Access is about to become group representative unless we find some
3808 nasty overlap which would preclude us from breaking this parameter
3812 while (j
< access_count
)
3814 struct access
*ac2
= VEC_index (access_p
, access_vec
, j
);
3815 if (ac2
->offset
!= access
->offset
)
3817 /* All or nothing law for parameters. */
3818 if (access
->offset
+ access
->size
> ac2
->offset
)
3823 else if (ac2
->size
!= access
->size
)
3826 if (access_precludes_ipa_sra_p (ac2
)
3827 || (ac2
->type
!= access
->type
3828 && (TREE_ADDRESSABLE (ac2
->type
)
3829 || TREE_ADDRESSABLE (access
->type
)))
3830 || (reference_alias_ptr_type (ac2
->expr
) != a1_alias_type
))
3833 modification
|= ac2
->write
;
3834 ac2
->group_representative
= access
;
3835 ac2
->next_sibling
= access
->next_sibling
;
3836 access
->next_sibling
= ac2
;
3841 access
->grp_maybe_modified
= modification
;
3844 *prev_acc_ptr
= access
;
3845 prev_acc_ptr
= &access
->next_grp
;
3846 total_size
+= access
->size
;
3850 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
3851 agg_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm
))), 1);
3853 agg_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (parm
)), 1);
3854 if (total_size
>= agg_size
)
3857 gcc_assert (group_count
> 0);
3861 /* Decide whether parameters with representative accesses given by REPR should
3862 be reduced into components. */
3865 decide_one_param_reduction (struct access
*repr
)
3867 int total_size
, cur_parm_size
, agg_size
, new_param_count
, parm_size_limit
;
3872 cur_parm_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (parm
)), 1);
3873 gcc_assert (cur_parm_size
> 0);
3875 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
3878 agg_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm
))), 1);
3883 agg_size
= cur_parm_size
;
3889 fprintf (dump_file
, "Evaluating PARAM group sizes for ");
3890 print_generic_expr (dump_file
, parm
, 0);
3891 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (parm
));
3892 for (acc
= repr
; acc
; acc
= acc
->next_grp
)
3893 dump_access (dump_file
, acc
, true);
3897 new_param_count
= 0;
3899 for (; repr
; repr
= repr
->next_grp
)
3901 gcc_assert (parm
== repr
->base
);
3903 /* Taking the address of a non-addressable field is verboten. */
3904 if (by_ref
&& repr
->non_addressable
)
3907 /* Do not decompose a non-BLKmode param in a way that would
3908 create BLKmode params. Especially for by-reference passing
3909 (thus, pointer-type param) this is hardly worthwhile. */
3910 if (DECL_MODE (parm
) != BLKmode
3911 && TYPE_MODE (repr
->type
) == BLKmode
)
3914 if (!by_ref
|| (!repr
->grp_maybe_modified
3915 && !repr
->grp_not_necessarilly_dereferenced
))
3916 total_size
+= repr
->size
;
3918 total_size
+= cur_parm_size
;
3923 gcc_assert (new_param_count
> 0);
3925 if (optimize_function_for_size_p (cfun
))
3926 parm_size_limit
= cur_parm_size
;
3928 parm_size_limit
= (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR
)
3931 if (total_size
< agg_size
3932 && total_size
<= parm_size_limit
)
3935 fprintf (dump_file
, " ....will be split into %i components\n",
3937 return new_param_count
;
3943 /* The order of the following enums is important, we need to do extra work for
3944 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3945 enum ipa_splicing_result
{ NO_GOOD_ACCESS
, UNUSED_PARAMS
, BY_VAL_ACCESSES
,
3946 MODIF_BY_REF_ACCESSES
, UNMODIF_BY_REF_ACCESSES
};
3948 /* Identify representatives of all accesses to all candidate parameters for
3949 IPA-SRA. Return result based on what representatives have been found. */
3951 static enum ipa_splicing_result
3952 splice_all_param_accesses (VEC (access_p
, heap
) **representatives
)
3954 enum ipa_splicing_result result
= NO_GOOD_ACCESS
;
3956 struct access
*repr
;
3958 *representatives
= VEC_alloc (access_p
, heap
, func_param_count
);
3960 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3962 parm
= DECL_CHAIN (parm
))
3964 if (is_unused_scalar_param (parm
))
3966 VEC_quick_push (access_p
, *representatives
,
3967 &no_accesses_representant
);
3968 if (result
== NO_GOOD_ACCESS
)
3969 result
= UNUSED_PARAMS
;
3971 else if (POINTER_TYPE_P (TREE_TYPE (parm
))
3972 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm
)))
3973 && bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
3975 repr
= unmodified_by_ref_scalar_representative (parm
);
3976 VEC_quick_push (access_p
, *representatives
, repr
);
3978 result
= UNMODIF_BY_REF_ACCESSES
;
3980 else if (bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
3982 bool ro_grp
= false;
3983 repr
= splice_param_accesses (parm
, &ro_grp
);
3984 VEC_quick_push (access_p
, *representatives
, repr
);
3986 if (repr
&& !no_accesses_p (repr
))
3988 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
3991 result
= UNMODIF_BY_REF_ACCESSES
;
3992 else if (result
< MODIF_BY_REF_ACCESSES
)
3993 result
= MODIF_BY_REF_ACCESSES
;
3995 else if (result
< BY_VAL_ACCESSES
)
3996 result
= BY_VAL_ACCESSES
;
3998 else if (no_accesses_p (repr
) && (result
== NO_GOOD_ACCESS
))
3999 result
= UNUSED_PARAMS
;
4002 VEC_quick_push (access_p
, *representatives
, (access_p
) NULL
);
4005 if (result
== NO_GOOD_ACCESS
)
4007 VEC_free (access_p
, heap
, *representatives
);
4008 *representatives
= NULL
;
4009 return NO_GOOD_ACCESS
;
4015 /* Return the index of BASE in PARMS. Abort if it is not found. */
4018 get_param_index (tree base
, VEC(tree
, heap
) *parms
)
4022 len
= VEC_length (tree
, parms
);
4023 for (i
= 0; i
< len
; i
++)
4024 if (VEC_index (tree
, parms
, i
) == base
)
4029 /* Convert the decisions made at the representative level into compact
4030 parameter adjustments. REPRESENTATIVES are pointers to first
4031 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4032 final number of adjustments. */
4034 static ipa_parm_adjustment_vec
4035 turn_representatives_into_adjustments (VEC (access_p
, heap
) *representatives
,
4036 int adjustments_count
)
4038 VEC (tree
, heap
) *parms
;
4039 ipa_parm_adjustment_vec adjustments
;
4043 gcc_assert (adjustments_count
> 0);
4044 parms
= ipa_get_vector_of_formal_parms (current_function_decl
);
4045 adjustments
= VEC_alloc (ipa_parm_adjustment_t
, heap
, adjustments_count
);
4046 parm
= DECL_ARGUMENTS (current_function_decl
);
4047 for (i
= 0; i
< func_param_count
; i
++, parm
= DECL_CHAIN (parm
))
4049 struct access
*repr
= VEC_index (access_p
, representatives
, i
);
4051 if (!repr
|| no_accesses_p (repr
))
4053 struct ipa_parm_adjustment
*adj
;
4055 adj
= VEC_quick_push (ipa_parm_adjustment_t
, adjustments
, NULL
);
4056 memset (adj
, 0, sizeof (*adj
));
4057 adj
->base_index
= get_param_index (parm
, parms
);
4060 adj
->copy_param
= 1;
4062 adj
->remove_param
= 1;
4066 struct ipa_parm_adjustment
*adj
;
4067 int index
= get_param_index (parm
, parms
);
4069 for (; repr
; repr
= repr
->next_grp
)
4071 adj
= VEC_quick_push (ipa_parm_adjustment_t
, adjustments
, NULL
);
4072 memset (adj
, 0, sizeof (*adj
));
4073 gcc_assert (repr
->base
== parm
);
4074 adj
->base_index
= index
;
4075 adj
->base
= repr
->base
;
4076 adj
->type
= repr
->type
;
4077 adj
->alias_ptr_type
= reference_alias_ptr_type (repr
->expr
);
4078 adj
->offset
= repr
->offset
;
4079 adj
->by_ref
= (POINTER_TYPE_P (TREE_TYPE (repr
->base
))
4080 && (repr
->grp_maybe_modified
4081 || repr
->grp_not_necessarilly_dereferenced
));
4086 VEC_free (tree
, heap
, parms
);
4090 /* Analyze the collected accesses and produce a plan what to do with the
4091 parameters in the form of adjustments, NULL meaning nothing. */
4093 static ipa_parm_adjustment_vec
4094 analyze_all_param_acesses (void)
4096 enum ipa_splicing_result repr_state
;
4097 bool proceed
= false;
4098 int i
, adjustments_count
= 0;
4099 VEC (access_p
, heap
) *representatives
;
4100 ipa_parm_adjustment_vec adjustments
;
4102 repr_state
= splice_all_param_accesses (&representatives
);
4103 if (repr_state
== NO_GOOD_ACCESS
)
4106 /* If there are any parameters passed by reference which are not modified
4107 directly, we need to check whether they can be modified indirectly. */
4108 if (repr_state
== UNMODIF_BY_REF_ACCESSES
)
4110 analyze_caller_dereference_legality (representatives
);
4111 analyze_modified_params (representatives
);
4114 for (i
= 0; i
< func_param_count
; i
++)
4116 struct access
*repr
= VEC_index (access_p
, representatives
, i
);
4118 if (repr
&& !no_accesses_p (repr
))
4120 if (repr
->grp_scalar_ptr
)
4122 adjustments_count
++;
4123 if (repr
->grp_not_necessarilly_dereferenced
4124 || repr
->grp_maybe_modified
)
4125 VEC_replace (access_p
, representatives
, i
, NULL
);
4129 sra_stats
.scalar_by_ref_to_by_val
++;
4134 int new_components
= decide_one_param_reduction (repr
);
4136 if (new_components
== 0)
4138 VEC_replace (access_p
, representatives
, i
, NULL
);
4139 adjustments_count
++;
4143 adjustments_count
+= new_components
;
4144 sra_stats
.aggregate_params_reduced
++;
4145 sra_stats
.param_reductions_created
+= new_components
;
4152 if (no_accesses_p (repr
))
4155 sra_stats
.deleted_unused_parameters
++;
4157 adjustments_count
++;
4161 if (!proceed
&& dump_file
)
4162 fprintf (dump_file
, "NOT proceeding to change params.\n");
4165 adjustments
= turn_representatives_into_adjustments (representatives
,
4170 VEC_free (access_p
, heap
, representatives
);
4174 /* If a parameter replacement identified by ADJ does not yet exist in the form
4175 of declaration, create it and record it, otherwise return the previously
4179 get_replaced_param_substitute (struct ipa_parm_adjustment
*adj
)
4182 if (!adj
->new_ssa_base
)
4184 char *pretty_name
= make_fancy_name (adj
->base
);
4186 repl
= create_tmp_reg (TREE_TYPE (adj
->base
), "ISR");
4187 DECL_NAME (repl
) = get_identifier (pretty_name
);
4188 obstack_free (&name_obstack
, pretty_name
);
4190 adj
->new_ssa_base
= repl
;
4193 repl
= adj
->new_ssa_base
;
4197 /* Find the first adjustment for a particular parameter BASE in a vector of
4198 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4201 static struct ipa_parm_adjustment
*
4202 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments
, tree base
)
4206 len
= VEC_length (ipa_parm_adjustment_t
, adjustments
);
4207 for (i
= 0; i
< len
; i
++)
4209 struct ipa_parm_adjustment
*adj
;
4211 adj
= &VEC_index (ipa_parm_adjustment_t
, adjustments
, i
);
4212 if (!adj
->copy_param
&& adj
->base
== base
)
4219 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4220 removed because its value is not used, replace the SSA_NAME with a one
4221 relating to a created VAR_DECL together all of its uses and return true.
4222 ADJUSTMENTS is a pointer to an adjustments vector. */
4225 replace_removed_params_ssa_names (gimple stmt
,
4226 ipa_parm_adjustment_vec adjustments
)
4228 struct ipa_parm_adjustment
*adj
;
4229 tree lhs
, decl
, repl
, name
;
4231 if (gimple_code (stmt
) == GIMPLE_PHI
)
4232 lhs
= gimple_phi_result (stmt
);
4233 else if (is_gimple_assign (stmt
))
4234 lhs
= gimple_assign_lhs (stmt
);
4235 else if (is_gimple_call (stmt
))
4236 lhs
= gimple_call_lhs (stmt
);
4240 if (TREE_CODE (lhs
) != SSA_NAME
)
4243 decl
= SSA_NAME_VAR (lhs
);
4244 if (decl
== NULL_TREE
4245 || TREE_CODE (decl
) != PARM_DECL
)
4248 adj
= get_adjustment_for_base (adjustments
, decl
);
4252 repl
= get_replaced_param_substitute (adj
);
4253 name
= make_ssa_name (repl
, stmt
);
4257 fprintf (dump_file
, "replacing an SSA name of a removed param ");
4258 print_generic_expr (dump_file
, lhs
, 0);
4259 fprintf (dump_file
, " with ");
4260 print_generic_expr (dump_file
, name
, 0);
4261 fprintf (dump_file
, "\n");
4264 if (is_gimple_assign (stmt
))
4265 gimple_assign_set_lhs (stmt
, name
);
4266 else if (is_gimple_call (stmt
))
4267 gimple_call_set_lhs (stmt
, name
);
4269 gimple_phi_set_result (stmt
, name
);
4271 replace_uses_by (lhs
, name
);
4272 release_ssa_name (lhs
);
4276 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4277 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4278 specifies whether the function should care about type incompatibility the
4279 current and new expressions. If it is false, the function will leave
4280 incompatibility issues to the caller. Return true iff the expression
4284 sra_ipa_modify_expr (tree
*expr
, bool convert
,
4285 ipa_parm_adjustment_vec adjustments
)
4288 struct ipa_parm_adjustment
*adj
, *cand
= NULL
;
4289 HOST_WIDE_INT offset
, size
, max_size
;
4292 len
= VEC_length (ipa_parm_adjustment_t
, adjustments
);
4294 if (TREE_CODE (*expr
) == BIT_FIELD_REF
4295 || TREE_CODE (*expr
) == IMAGPART_EXPR
4296 || TREE_CODE (*expr
) == REALPART_EXPR
)
4298 expr
= &TREE_OPERAND (*expr
, 0);
4302 base
= get_ref_base_and_extent (*expr
, &offset
, &size
, &max_size
);
4303 if (!base
|| size
== -1 || max_size
== -1)
4306 if (TREE_CODE (base
) == MEM_REF
)
4308 offset
+= mem_ref_offset (base
).low
* BITS_PER_UNIT
;
4309 base
= TREE_OPERAND (base
, 0);
4312 base
= get_ssa_base_param (base
);
4313 if (!base
|| TREE_CODE (base
) != PARM_DECL
)
4316 for (i
= 0; i
< len
; i
++)
4318 adj
= &VEC_index (ipa_parm_adjustment_t
, adjustments
, i
);
4320 if (adj
->base
== base
&&
4321 (adj
->offset
== offset
|| adj
->remove_param
))
4327 if (!cand
|| cand
->copy_param
|| cand
->remove_param
)
4331 src
= build_simple_mem_ref (cand
->reduction
);
4333 src
= cand
->reduction
;
4335 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
4337 fprintf (dump_file
, "About to replace expr ");
4338 print_generic_expr (dump_file
, *expr
, 0);
4339 fprintf (dump_file
, " with ");
4340 print_generic_expr (dump_file
, src
, 0);
4341 fprintf (dump_file
, "\n");
4344 if (convert
&& !useless_type_conversion_p (TREE_TYPE (*expr
), cand
->type
))
4346 tree vce
= build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (*expr
), src
);
4354 /* If the statement pointed to by STMT_PTR contains any expressions that need
4355 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4356 potential type incompatibilities (GSI is used to accommodate conversion
4357 statements and must point to the statement). Return true iff the statement
4361 sra_ipa_modify_assign (gimple
*stmt_ptr
, gimple_stmt_iterator
*gsi
,
4362 ipa_parm_adjustment_vec adjustments
)
4364 gimple stmt
= *stmt_ptr
;
4365 tree
*lhs_p
, *rhs_p
;
4368 if (!gimple_assign_single_p (stmt
))
4371 rhs_p
= gimple_assign_rhs1_ptr (stmt
);
4372 lhs_p
= gimple_assign_lhs_ptr (stmt
);
4374 any
= sra_ipa_modify_expr (rhs_p
, false, adjustments
);
4375 any
|= sra_ipa_modify_expr (lhs_p
, false, adjustments
);
4378 tree new_rhs
= NULL_TREE
;
4380 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p
), TREE_TYPE (*rhs_p
)))
4382 if (TREE_CODE (*rhs_p
) == CONSTRUCTOR
)
4384 /* V_C_Es of constructors can cause trouble (PR 42714). */
4385 if (is_gimple_reg_type (TREE_TYPE (*lhs_p
)))
4386 *rhs_p
= build_zero_cst (TREE_TYPE (*lhs_p
));
4388 *rhs_p
= build_constructor (TREE_TYPE (*lhs_p
), 0);
4391 new_rhs
= fold_build1_loc (gimple_location (stmt
),
4392 VIEW_CONVERT_EXPR
, TREE_TYPE (*lhs_p
),
4395 else if (REFERENCE_CLASS_P (*rhs_p
)
4396 && is_gimple_reg_type (TREE_TYPE (*lhs_p
))
4397 && !is_gimple_reg (*lhs_p
))
4398 /* This can happen when an assignment in between two single field
4399 structures is turned into an assignment in between two pointers to
4400 scalars (PR 42237). */
4405 tree tmp
= force_gimple_operand_gsi (gsi
, new_rhs
, true, NULL_TREE
,
4406 true, GSI_SAME_STMT
);
4408 gimple_assign_set_rhs_from_tree (gsi
, tmp
);
4417 /* Traverse the function body and all modifications as described in
4418 ADJUSTMENTS. Return true iff the CFG has been changed. */
4421 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments
)
4423 bool cfg_changed
= false;
4428 gimple_stmt_iterator gsi
;
4430 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4431 replace_removed_params_ssa_names (gsi_stmt (gsi
), adjustments
);
4433 gsi
= gsi_start_bb (bb
);
4434 while (!gsi_end_p (gsi
))
4436 gimple stmt
= gsi_stmt (gsi
);
4437 bool modified
= false;
4441 switch (gimple_code (stmt
))
4444 t
= gimple_return_retval_ptr (stmt
);
4445 if (*t
!= NULL_TREE
)
4446 modified
|= sra_ipa_modify_expr (t
, true, adjustments
);
4450 modified
|= sra_ipa_modify_assign (&stmt
, &gsi
, adjustments
);
4451 modified
|= replace_removed_params_ssa_names (stmt
, adjustments
);
4455 /* Operands must be processed before the lhs. */
4456 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
4458 t
= gimple_call_arg_ptr (stmt
, i
);
4459 modified
|= sra_ipa_modify_expr (t
, true, adjustments
);
4462 if (gimple_call_lhs (stmt
))
4464 t
= gimple_call_lhs_ptr (stmt
);
4465 modified
|= sra_ipa_modify_expr (t
, false, adjustments
);
4466 modified
|= replace_removed_params_ssa_names (stmt
,
4472 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
4474 t
= &TREE_VALUE (gimple_asm_input_op (stmt
, i
));
4475 modified
|= sra_ipa_modify_expr (t
, true, adjustments
);
4477 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
4479 t
= &TREE_VALUE (gimple_asm_output_op (stmt
, i
));
4480 modified
|= sra_ipa_modify_expr (t
, false, adjustments
);
4491 if (maybe_clean_eh_stmt (stmt
)
4492 && gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
4502 /* Call gimple_debug_bind_reset_value on all debug statements describing
4503 gimple register parameters that are being removed or replaced. */
4506 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments
)
4509 gimple_stmt_iterator
*gsip
= NULL
, gsi
;
4511 if (MAY_HAVE_DEBUG_STMTS
&& single_succ_p (ENTRY_BLOCK_PTR
))
4513 gsi
= gsi_after_labels (single_succ (ENTRY_BLOCK_PTR
));
4516 len
= VEC_length (ipa_parm_adjustment_t
, adjustments
);
4517 for (i
= 0; i
< len
; i
++)
4519 struct ipa_parm_adjustment
*adj
;
4520 imm_use_iterator ui
;
4521 gimple stmt
, def_temp
;
4522 tree name
, vexpr
, copy
= NULL_TREE
;
4523 use_operand_p use_p
;
4525 adj
= &VEC_index (ipa_parm_adjustment_t
, adjustments
, i
);
4526 if (adj
->copy_param
|| !is_gimple_reg (adj
->base
))
4528 name
= ssa_default_def (cfun
, adj
->base
);
4531 FOR_EACH_IMM_USE_STMT (stmt
, ui
, name
)
4533 /* All other users must have been removed by
4534 ipa_sra_modify_function_body. */
4535 gcc_assert (is_gimple_debug (stmt
));
4536 if (vexpr
== NULL
&& gsip
!= NULL
)
4538 gcc_assert (TREE_CODE (adj
->base
) == PARM_DECL
);
4539 vexpr
= make_node (DEBUG_EXPR_DECL
);
4540 def_temp
= gimple_build_debug_source_bind (vexpr
, adj
->base
,
4542 DECL_ARTIFICIAL (vexpr
) = 1;
4543 TREE_TYPE (vexpr
) = TREE_TYPE (name
);
4544 DECL_MODE (vexpr
) = DECL_MODE (adj
->base
);
4545 gsi_insert_before (gsip
, def_temp
, GSI_SAME_STMT
);
4549 FOR_EACH_IMM_USE_ON_STMT (use_p
, ui
)
4550 SET_USE (use_p
, vexpr
);
4553 gimple_debug_bind_reset_value (stmt
);
4556 /* Create a VAR_DECL for debug info purposes. */
4557 if (!DECL_IGNORED_P (adj
->base
))
4559 copy
= build_decl (DECL_SOURCE_LOCATION (current_function_decl
),
4560 VAR_DECL
, DECL_NAME (adj
->base
),
4561 TREE_TYPE (adj
->base
));
4562 if (DECL_PT_UID_SET_P (adj
->base
))
4563 SET_DECL_PT_UID (copy
, DECL_PT_UID (adj
->base
));
4564 TREE_ADDRESSABLE (copy
) = TREE_ADDRESSABLE (adj
->base
);
4565 TREE_READONLY (copy
) = TREE_READONLY (adj
->base
);
4566 TREE_THIS_VOLATILE (copy
) = TREE_THIS_VOLATILE (adj
->base
);
4567 DECL_GIMPLE_REG_P (copy
) = DECL_GIMPLE_REG_P (adj
->base
);
4568 DECL_ARTIFICIAL (copy
) = DECL_ARTIFICIAL (adj
->base
);
4569 DECL_IGNORED_P (copy
) = DECL_IGNORED_P (adj
->base
);
4570 DECL_ABSTRACT_ORIGIN (copy
) = DECL_ORIGIN (adj
->base
);
4571 DECL_SEEN_IN_BIND_EXPR_P (copy
) = 1;
4572 SET_DECL_RTL (copy
, 0);
4573 TREE_USED (copy
) = 1;
4574 DECL_CONTEXT (copy
) = current_function_decl
;
4575 add_local_decl (cfun
, copy
);
4577 BLOCK_VARS (DECL_INITIAL (current_function_decl
));
4578 BLOCK_VARS (DECL_INITIAL (current_function_decl
)) = copy
;
4580 if (gsip
!= NULL
&& copy
&& target_for_debug_bind (adj
->base
))
4582 gcc_assert (TREE_CODE (adj
->base
) == PARM_DECL
);
4584 def_temp
= gimple_build_debug_bind (copy
, vexpr
, NULL
);
4586 def_temp
= gimple_build_debug_source_bind (copy
, adj
->base
,
4588 gsi_insert_before (gsip
, def_temp
, GSI_SAME_STMT
);
4593 /* Return false iff all callers have at least as many actual arguments as there
4594 are formal parameters in the current function. */
4597 not_all_callers_have_enough_arguments_p (struct cgraph_node
*node
,
4598 void *data ATTRIBUTE_UNUSED
)
4600 struct cgraph_edge
*cs
;
4601 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4602 if (!callsite_has_enough_arguments_p (cs
->call_stmt
))
4608 /* Convert all callers of NODE. */
4611 convert_callers_for_node (struct cgraph_node
*node
,
4614 ipa_parm_adjustment_vec adjustments
= (ipa_parm_adjustment_vec
)data
;
4615 bitmap recomputed_callers
= BITMAP_ALLOC (NULL
);
4616 struct cgraph_edge
*cs
;
4618 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4620 current_function_decl
= cs
->caller
->symbol
.decl
;
4621 push_cfun (DECL_STRUCT_FUNCTION (cs
->caller
->symbol
.decl
));
4624 fprintf (dump_file
, "Adjusting call (%i -> %i) %s -> %s\n",
4625 cs
->caller
->uid
, cs
->callee
->uid
,
4626 xstrdup (cgraph_node_name (cs
->caller
)),
4627 xstrdup (cgraph_node_name (cs
->callee
)));
4629 ipa_modify_call_arguments (cs
, cs
->call_stmt
, adjustments
);
4634 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4635 if (bitmap_set_bit (recomputed_callers
, cs
->caller
->uid
)
4636 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs
->caller
->symbol
.decl
)))
4637 compute_inline_parameters (cs
->caller
, true);
4638 BITMAP_FREE (recomputed_callers
);
4643 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4646 convert_callers (struct cgraph_node
*node
, tree old_decl
,
4647 ipa_parm_adjustment_vec adjustments
)
4649 tree old_cur_fndecl
= current_function_decl
;
4650 basic_block this_block
;
4652 cgraph_for_node_and_aliases (node
, convert_callers_for_node
,
4653 adjustments
, false);
4655 current_function_decl
= old_cur_fndecl
;
4657 if (!encountered_recursive_call
)
4660 FOR_EACH_BB (this_block
)
4662 gimple_stmt_iterator gsi
;
4664 for (gsi
= gsi_start_bb (this_block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4666 gimple stmt
= gsi_stmt (gsi
);
4668 if (gimple_code (stmt
) != GIMPLE_CALL
)
4670 call_fndecl
= gimple_call_fndecl (stmt
);
4671 if (call_fndecl
== old_decl
)
4674 fprintf (dump_file
, "Adjusting recursive call");
4675 gimple_call_set_fndecl (stmt
, node
->symbol
.decl
);
4676 ipa_modify_call_arguments (NULL
, stmt
, adjustments
);
4684 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4685 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4688 modify_function (struct cgraph_node
*node
, ipa_parm_adjustment_vec adjustments
)
4690 struct cgraph_node
*new_node
;
4692 VEC (cgraph_edge_p
, heap
) * redirect_callers
= collect_callers_of_node (node
);
4694 rebuild_cgraph_edges ();
4695 free_dominance_info (CDI_DOMINATORS
);
4697 current_function_decl
= NULL_TREE
;
4699 new_node
= cgraph_function_versioning (node
, redirect_callers
, NULL
, NULL
,
4700 false, NULL
, NULL
, "isra");
4701 VEC_free (cgraph_edge_p
, heap
, redirect_callers
);
4703 current_function_decl
= new_node
->symbol
.decl
;
4704 push_cfun (DECL_STRUCT_FUNCTION (new_node
->symbol
.decl
));
4706 ipa_modify_formal_parameters (current_function_decl
, adjustments
, "ISRA");
4707 cfg_changed
= ipa_sra_modify_function_body (adjustments
);
4708 sra_ipa_reset_debug_stmts (adjustments
);
4709 convert_callers (new_node
, node
->symbol
.decl
, adjustments
);
4710 cgraph_make_node_local (new_node
);
4714 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4715 attributes, return true otherwise. NODE is the cgraph node of the current
4719 ipa_sra_preliminary_function_checks (struct cgraph_node
*node
)
4721 if (!cgraph_node_can_be_local_p (node
))
4724 fprintf (dump_file
, "Function not local to this compilation unit.\n");
4728 if (!node
->local
.can_change_signature
)
4731 fprintf (dump_file
, "Function can not change signature.\n");
4735 if (!tree_versionable_function_p (node
->symbol
.decl
))
4738 fprintf (dump_file
, "Function is not versionable.\n");
4742 if (DECL_VIRTUAL_P (current_function_decl
))
4745 fprintf (dump_file
, "Function is a virtual method.\n");
4749 if ((DECL_COMDAT (node
->symbol
.decl
) || DECL_EXTERNAL (node
->symbol
.decl
))
4750 && inline_summary(node
)->size
>= MAX_INLINE_INSNS_AUTO
)
4753 fprintf (dump_file
, "Function too big to be made truly local.\n");
4761 "Function has no callers in this compilation unit.\n");
4768 fprintf (dump_file
, "Function uses stdarg. \n");
4772 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->symbol
.decl
)))
4778 /* Perform early interprocedural SRA. */
4781 ipa_early_sra (void)
4783 struct cgraph_node
*node
= cgraph_get_node (current_function_decl
);
4784 ipa_parm_adjustment_vec adjustments
;
4787 if (!ipa_sra_preliminary_function_checks (node
))
4791 sra_mode
= SRA_MODE_EARLY_IPA
;
4793 if (!find_param_candidates ())
4796 fprintf (dump_file
, "Function has no IPA-SRA candidates.\n");
4800 if (cgraph_for_node_and_aliases (node
, not_all_callers_have_enough_arguments_p
,
4804 fprintf (dump_file
, "There are callers with insufficient number of "
4809 bb_dereferences
= XCNEWVEC (HOST_WIDE_INT
,
4811 * last_basic_block_for_function (cfun
));
4812 final_bbs
= BITMAP_ALLOC (NULL
);
4815 if (encountered_apply_args
)
4818 fprintf (dump_file
, "Function calls __builtin_apply_args().\n");
4822 if (encountered_unchangable_recursive_call
)
4825 fprintf (dump_file
, "Function calls itself with insufficient "
4826 "number of arguments.\n");
4830 adjustments
= analyze_all_param_acesses ();
4834 ipa_dump_param_adjustments (dump_file
, adjustments
, current_function_decl
);
4836 if (modify_function (node
, adjustments
))
4837 ret
= TODO_update_ssa
| TODO_cleanup_cfg
;
4839 ret
= TODO_update_ssa
;
4840 VEC_free (ipa_parm_adjustment_t
, heap
, adjustments
);
4842 statistics_counter_event (cfun
, "Unused parameters deleted",
4843 sra_stats
.deleted_unused_parameters
);
4844 statistics_counter_event (cfun
, "Scalar parameters converted to by-value",
4845 sra_stats
.scalar_by_ref_to_by_val
);
4846 statistics_counter_event (cfun
, "Aggregate parameters broken up",
4847 sra_stats
.aggregate_params_reduced
);
4848 statistics_counter_event (cfun
, "Aggregate parameter components created",
4849 sra_stats
.param_reductions_created
);
4852 BITMAP_FREE (final_bbs
);
4853 free (bb_dereferences
);
4855 sra_deinitialize ();
4859 /* Return if early ipa sra shall be performed. */
4861 ipa_early_sra_gate (void)
4863 return flag_ipa_sra
&& dbg_cnt (eipa_sra
);
4866 struct gimple_opt_pass pass_early_ipa_sra
=
4870 "eipa_sra", /* name */
4871 ipa_early_sra_gate
, /* gate */
4872 ipa_early_sra
, /* execute */
4875 0, /* static_pass_number */
4876 TV_IPA_SRA
, /* tv_id */
4877 0, /* properties_required */
4878 0, /* properties_provided */
4879 0, /* properties_destroyed */
4880 0, /* todo_flags_start */
4881 TODO_dump_symtab
/* todo_flags_finish */