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 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"
83 #include "tree-flow.h"
85 #include "tree-pretty-print.h"
86 #include "statistics.h"
87 #include "tree-dump.h"
93 #include "tree-inline.h"
94 #include "gimple-pretty-print.h"
96 /* Enumeration of all aggregate reductions we can do. */
97 enum sra_mode
{ SRA_MODE_EARLY_IPA
, /* early call regularization */
98 SRA_MODE_EARLY_INTRA
, /* early intraprocedural SRA */
99 SRA_MODE_INTRA
}; /* late intraprocedural SRA */
101 /* Global variable describing which aggregate reduction we are performing at
103 static enum sra_mode sra_mode
;
107 /* ACCESS represents each access to an aggregate variable (as a whole or a
108 part). It can also represent a group of accesses that refer to exactly the
109 same fragment of an aggregate (i.e. those that have exactly the same offset
110 and size). Such representatives for a single aggregate, once determined,
111 are linked in a linked list and have the group fields set.
113 Moreover, when doing intraprocedural SRA, a tree is built from those
114 representatives (by the means of first_child and next_sibling pointers), in
115 which all items in a subtree are "within" the root, i.e. their offset is
116 greater or equal to offset of the root and offset+size is smaller or equal
117 to offset+size of the root. Children of an access are sorted by offset.
119 Note that accesses to parts of vector and complex number types always
120 represented by an access to the whole complex number or a vector. It is a
121 duty of the modifying functions to replace them appropriately. */
125 /* Values returned by `get_ref_base_and_extent' for each component reference
126 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
127 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
128 HOST_WIDE_INT offset
;
132 /* Expression. It is context dependent so do not use it to create new
133 expressions to access the original aggregate. See PR 42154 for a
139 /* The statement this access belongs to. */
142 /* Next group representative for this aggregate. */
143 struct access
*next_grp
;
145 /* Pointer to the group representative. Pointer to itself if the struct is
146 the representative. */
147 struct access
*group_representative
;
149 /* If this access has any children (in terms of the definition above), this
150 points to the first one. */
151 struct access
*first_child
;
153 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
154 described above. In IPA-SRA this is a pointer to the next access
155 belonging to the same group (having the same representative). */
156 struct access
*next_sibling
;
158 /* Pointers to the first and last element in the linked list of assign
160 struct assign_link
*first_link
, *last_link
;
162 /* Pointer to the next access in the work queue. */
163 struct access
*next_queued
;
165 /* Replacement variable for this access "region." Never to be accessed
166 directly, always only by the means of get_access_replacement() and only
167 when grp_to_be_replaced flag is set. */
168 tree replacement_decl
;
170 /* Is this particular access write access? */
173 /* Is this access an artificial one created to scalarize some record
175 unsigned total_scalarization
: 1;
177 /* Is this access currently in the work queue? */
178 unsigned grp_queued
: 1;
180 /* Does this group contain a write access? This flag is propagated down the
182 unsigned grp_write
: 1;
184 /* Does this group contain a read access? This flag is propagated down the
186 unsigned grp_read
: 1;
188 /* Does this group contain a read access that comes from an assignment
189 statement? This flag is propagated down the access tree. */
190 unsigned grp_assignment_read
: 1;
192 /* Other passes of the analysis use this bit to make function
193 analyze_access_subtree create scalar replacements for this group if
195 unsigned grp_hint
: 1;
197 /* Is the subtree rooted in this access fully covered by scalar
199 unsigned grp_covered
: 1;
201 /* If set to true, this access and all below it in an access tree must not be
203 unsigned grp_unscalarizable_region
: 1;
205 /* Whether data have been written to parts of the aggregate covered by this
206 access which is not to be scalarized. This flag is propagated up in the
208 unsigned grp_unscalarized_data
: 1;
210 /* Does this access and/or group contain a write access through a
212 unsigned grp_partial_lhs
: 1;
214 /* Set when a scalar replacement should be created for this variable. We do
215 the decision and creation at different places because create_tmp_var
216 cannot be called from within FOR_EACH_REFERENCED_VAR. */
217 unsigned grp_to_be_replaced
: 1;
219 /* Is it possible that the group refers to data which might be (directly or
220 otherwise) modified? */
221 unsigned grp_maybe_modified
: 1;
223 /* Set when this is a representative of a pointer to scalar (i.e. by
224 reference) parameter which we consider for turning into a plain scalar
225 (i.e. a by value parameter). */
226 unsigned grp_scalar_ptr
: 1;
228 /* Set when we discover that this pointer is not safe to dereference in the
230 unsigned grp_not_necessarilly_dereferenced
: 1;
233 typedef struct access
*access_p
;
235 DEF_VEC_P (access_p
);
236 DEF_VEC_ALLOC_P (access_p
, heap
);
238 /* Alloc pool for allocating access structures. */
239 static alloc_pool access_pool
;
241 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
242 are used to propagate subaccesses from rhs to lhs as long as they don't
243 conflict with what is already there. */
246 struct access
*lacc
, *racc
;
247 struct assign_link
*next
;
250 /* Alloc pool for allocating assign link structures. */
251 static alloc_pool link_pool
;
253 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
254 static struct pointer_map_t
*base_access_vec
;
256 /* Bitmap of candidates. */
257 static bitmap candidate_bitmap
;
259 /* Bitmap of candidates which we should try to entirely scalarize away and
260 those which cannot be (because they are and need be used as a whole). */
261 static bitmap should_scalarize_away_bitmap
, cannot_scalarize_away_bitmap
;
263 /* Obstack for creation of fancy names. */
264 static struct obstack name_obstack
;
266 /* Head of a linked list of accesses that need to have its subaccesses
267 propagated to their assignment counterparts. */
268 static struct access
*work_queue_head
;
270 /* Number of parameters of the analyzed function when doing early ipa SRA. */
271 static int func_param_count
;
273 /* scan_function sets the following to true if it encounters a call to
274 __builtin_apply_args. */
275 static bool encountered_apply_args
;
277 /* Set by scan_function when it finds a recursive call. */
278 static bool encountered_recursive_call
;
280 /* Set by scan_function when it finds a recursive call with less actual
281 arguments than formal parameters.. */
282 static bool encountered_unchangable_recursive_call
;
284 /* This is a table in which for each basic block and parameter there is a
285 distance (offset + size) in that parameter which is dereferenced and
286 accessed in that BB. */
287 static HOST_WIDE_INT
*bb_dereferences
;
288 /* Bitmap of BBs that can cause the function to "stop" progressing by
289 returning, throwing externally, looping infinitely or calling a function
290 which might abort etc.. */
291 static bitmap final_bbs
;
293 /* Representative of no accesses at all. */
294 static struct access no_accesses_representant
;
296 /* Predicate to test the special value. */
299 no_accesses_p (struct access
*access
)
301 return access
== &no_accesses_representant
;
304 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
305 representative fields are dumped, otherwise those which only describe the
306 individual access are. */
310 /* Number of processed aggregates is readily available in
311 analyze_all_variable_accesses and so is not stored here. */
313 /* Number of created scalar replacements. */
316 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
320 /* Number of statements created by generate_subtree_copies. */
323 /* Number of statements created by load_assign_lhs_subreplacements. */
326 /* Number of times sra_modify_assign has deleted a statement. */
329 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
330 RHS reparately due to type conversions or nonexistent matching
332 int separate_lhs_rhs_handling
;
334 /* Number of parameters that were removed because they were unused. */
335 int deleted_unused_parameters
;
337 /* Number of scalars passed as parameters by reference that have been
338 converted to be passed by value. */
339 int scalar_by_ref_to_by_val
;
341 /* Number of aggregate parameters that were replaced by one or more of their
343 int aggregate_params_reduced
;
345 /* Numbber of components created when splitting aggregate parameters. */
346 int param_reductions_created
;
350 dump_access (FILE *f
, struct access
*access
, bool grp
)
352 fprintf (f
, "access { ");
353 fprintf (f
, "base = (%d)'", DECL_UID (access
->base
));
354 print_generic_expr (f
, access
->base
, 0);
355 fprintf (f
, "', offset = " HOST_WIDE_INT_PRINT_DEC
, access
->offset
);
356 fprintf (f
, ", size = " HOST_WIDE_INT_PRINT_DEC
, access
->size
);
357 fprintf (f
, ", expr = ");
358 print_generic_expr (f
, access
->expr
, 0);
359 fprintf (f
, ", type = ");
360 print_generic_expr (f
, access
->type
, 0);
362 fprintf (f
, ", grp_write = %d, total_scalarization = %d, "
363 "grp_read = %d, grp_hint = %d, grp_assignment_read = %d,"
364 "grp_covered = %d, grp_unscalarizable_region = %d, "
365 "grp_unscalarized_data = %d, grp_partial_lhs = %d, "
366 "grp_to_be_replaced = %d, grp_maybe_modified = %d, "
367 "grp_not_necessarilly_dereferenced = %d\n",
368 access
->grp_write
, access
->total_scalarization
,
369 access
->grp_read
, access
->grp_hint
, access
->grp_assignment_read
,
370 access
->grp_covered
, access
->grp_unscalarizable_region
,
371 access
->grp_unscalarized_data
, access
->grp_partial_lhs
,
372 access
->grp_to_be_replaced
, access
->grp_maybe_modified
,
373 access
->grp_not_necessarilly_dereferenced
);
375 fprintf (f
, ", write = %d, total_scalarization = %d, "
376 "grp_partial_lhs = %d\n",
377 access
->write
, access
->total_scalarization
,
378 access
->grp_partial_lhs
);
381 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
384 dump_access_tree_1 (FILE *f
, struct access
*access
, int level
)
390 for (i
= 0; i
< level
; i
++)
391 fputs ("* ", dump_file
);
393 dump_access (f
, access
, true);
395 if (access
->first_child
)
396 dump_access_tree_1 (f
, access
->first_child
, level
+ 1);
398 access
= access
->next_sibling
;
403 /* Dump all access trees for a variable, given the pointer to the first root in
407 dump_access_tree (FILE *f
, struct access
*access
)
409 for (; access
; access
= access
->next_grp
)
410 dump_access_tree_1 (f
, access
, 0);
413 /* Return true iff ACC is non-NULL and has subaccesses. */
416 access_has_children_p (struct access
*acc
)
418 return acc
&& acc
->first_child
;
421 /* Return a vector of pointers to accesses for the variable given in BASE or
422 NULL if there is none. */
424 static VEC (access_p
, heap
) *
425 get_base_access_vector (tree base
)
429 slot
= pointer_map_contains (base_access_vec
, base
);
433 return *(VEC (access_p
, heap
) **) slot
;
436 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
437 in ACCESS. Return NULL if it cannot be found. */
439 static struct access
*
440 find_access_in_subtree (struct access
*access
, HOST_WIDE_INT offset
,
443 while (access
&& (access
->offset
!= offset
|| access
->size
!= size
))
445 struct access
*child
= access
->first_child
;
447 while (child
&& (child
->offset
+ child
->size
<= offset
))
448 child
= child
->next_sibling
;
455 /* Return the first group representative for DECL or NULL if none exists. */
457 static struct access
*
458 get_first_repr_for_decl (tree base
)
460 VEC (access_p
, heap
) *access_vec
;
462 access_vec
= get_base_access_vector (base
);
466 return VEC_index (access_p
, access_vec
, 0);
469 /* Find an access representative for the variable BASE and given OFFSET and
470 SIZE. Requires that access trees have already been built. Return NULL if
471 it cannot be found. */
473 static struct access
*
474 get_var_base_offset_size_access (tree base
, HOST_WIDE_INT offset
,
477 struct access
*access
;
479 access
= get_first_repr_for_decl (base
);
480 while (access
&& (access
->offset
+ access
->size
<= offset
))
481 access
= access
->next_grp
;
485 return find_access_in_subtree (access
, offset
, size
);
488 /* Add LINK to the linked list of assign links of RACC. */
490 add_link_to_rhs (struct access
*racc
, struct assign_link
*link
)
492 gcc_assert (link
->racc
== racc
);
494 if (!racc
->first_link
)
496 gcc_assert (!racc
->last_link
);
497 racc
->first_link
= link
;
500 racc
->last_link
->next
= link
;
502 racc
->last_link
= link
;
506 /* Move all link structures in their linked list in OLD_RACC to the linked list
509 relink_to_new_repr (struct access
*new_racc
, struct access
*old_racc
)
511 if (!old_racc
->first_link
)
513 gcc_assert (!old_racc
->last_link
);
517 if (new_racc
->first_link
)
519 gcc_assert (!new_racc
->last_link
->next
);
520 gcc_assert (!old_racc
->last_link
|| !old_racc
->last_link
->next
);
522 new_racc
->last_link
->next
= old_racc
->first_link
;
523 new_racc
->last_link
= old_racc
->last_link
;
527 gcc_assert (!new_racc
->last_link
);
529 new_racc
->first_link
= old_racc
->first_link
;
530 new_racc
->last_link
= old_racc
->last_link
;
532 old_racc
->first_link
= old_racc
->last_link
= NULL
;
535 /* Add ACCESS to the work queue (which is actually a stack). */
538 add_access_to_work_queue (struct access
*access
)
540 if (!access
->grp_queued
)
542 gcc_assert (!access
->next_queued
);
543 access
->next_queued
= work_queue_head
;
544 access
->grp_queued
= 1;
545 work_queue_head
= access
;
549 /* Pop an access from the work queue, and return it, assuming there is one. */
551 static struct access
*
552 pop_access_from_work_queue (void)
554 struct access
*access
= work_queue_head
;
556 work_queue_head
= access
->next_queued
;
557 access
->next_queued
= NULL
;
558 access
->grp_queued
= 0;
563 /* Allocate necessary structures. */
566 sra_initialize (void)
568 candidate_bitmap
= BITMAP_ALLOC (NULL
);
569 should_scalarize_away_bitmap
= BITMAP_ALLOC (NULL
);
570 cannot_scalarize_away_bitmap
= BITMAP_ALLOC (NULL
);
571 gcc_obstack_init (&name_obstack
);
572 access_pool
= create_alloc_pool ("SRA accesses", sizeof (struct access
), 16);
573 link_pool
= create_alloc_pool ("SRA links", sizeof (struct assign_link
), 16);
574 base_access_vec
= pointer_map_create ();
575 memset (&sra_stats
, 0, sizeof (sra_stats
));
576 encountered_apply_args
= false;
577 encountered_recursive_call
= false;
578 encountered_unchangable_recursive_call
= false;
581 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
584 delete_base_accesses (const void *key ATTRIBUTE_UNUSED
, void **value
,
585 void *data ATTRIBUTE_UNUSED
)
587 VEC (access_p
, heap
) *access_vec
;
588 access_vec
= (VEC (access_p
, heap
) *) *value
;
589 VEC_free (access_p
, heap
, access_vec
);
594 /* Deallocate all general structures. */
597 sra_deinitialize (void)
599 BITMAP_FREE (candidate_bitmap
);
600 BITMAP_FREE (should_scalarize_away_bitmap
);
601 BITMAP_FREE (cannot_scalarize_away_bitmap
);
602 free_alloc_pool (access_pool
);
603 free_alloc_pool (link_pool
);
604 obstack_free (&name_obstack
, NULL
);
606 pointer_map_traverse (base_access_vec
, delete_base_accesses
, NULL
);
607 pointer_map_destroy (base_access_vec
);
610 /* Remove DECL from candidates for SRA and write REASON to the dump file if
613 disqualify_candidate (tree decl
, const char *reason
)
615 bitmap_clear_bit (candidate_bitmap
, DECL_UID (decl
));
617 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
619 fprintf (dump_file
, "! Disqualifying ");
620 print_generic_expr (dump_file
, decl
, 0);
621 fprintf (dump_file
, " - %s\n", reason
);
625 /* Return true iff the type contains a field or an element which does not allow
629 type_internals_preclude_sra_p (tree type
)
634 switch (TREE_CODE (type
))
638 case QUAL_UNION_TYPE
:
639 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
640 if (TREE_CODE (fld
) == FIELD_DECL
)
642 tree ft
= TREE_TYPE (fld
);
644 if (TREE_THIS_VOLATILE (fld
)
645 || !DECL_FIELD_OFFSET (fld
) || !DECL_SIZE (fld
)
646 || !host_integerp (DECL_FIELD_OFFSET (fld
), 1)
647 || !host_integerp (DECL_SIZE (fld
), 1))
650 if (AGGREGATE_TYPE_P (ft
)
651 && type_internals_preclude_sra_p (ft
))
658 et
= TREE_TYPE (type
);
660 if (AGGREGATE_TYPE_P (et
))
661 return type_internals_preclude_sra_p (et
);
670 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
671 base variable if it is. Return T if it is not an SSA_NAME. */
674 get_ssa_base_param (tree t
)
676 if (TREE_CODE (t
) == SSA_NAME
)
678 if (SSA_NAME_IS_DEFAULT_DEF (t
))
679 return SSA_NAME_VAR (t
);
686 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
687 belongs to, unless the BB has already been marked as a potentially
691 mark_parm_dereference (tree base
, HOST_WIDE_INT dist
, gimple stmt
)
693 basic_block bb
= gimple_bb (stmt
);
694 int idx
, parm_index
= 0;
697 if (bitmap_bit_p (final_bbs
, bb
->index
))
700 for (parm
= DECL_ARGUMENTS (current_function_decl
);
701 parm
&& parm
!= base
;
702 parm
= DECL_CHAIN (parm
))
705 gcc_assert (parm_index
< func_param_count
);
707 idx
= bb
->index
* func_param_count
+ parm_index
;
708 if (bb_dereferences
[idx
] < dist
)
709 bb_dereferences
[idx
] = dist
;
712 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
713 the three fields. Also add it to the vector of accesses corresponding to
714 the base. Finally, return the new access. */
716 static struct access
*
717 create_access_1 (tree base
, HOST_WIDE_INT offset
, HOST_WIDE_INT size
)
719 VEC (access_p
, heap
) *vec
;
720 struct access
*access
;
723 access
= (struct access
*) pool_alloc (access_pool
);
724 memset (access
, 0, sizeof (struct access
));
726 access
->offset
= offset
;
729 slot
= pointer_map_contains (base_access_vec
, base
);
731 vec
= (VEC (access_p
, heap
) *) *slot
;
733 vec
= VEC_alloc (access_p
, heap
, 32);
735 VEC_safe_push (access_p
, heap
, vec
, access
);
737 *((struct VEC (access_p
,heap
) **)
738 pointer_map_insert (base_access_vec
, base
)) = vec
;
743 /* Create and insert access for EXPR. Return created access, or NULL if it is
746 static struct access
*
747 create_access (tree expr
, gimple stmt
, bool write
)
749 struct access
*access
;
750 HOST_WIDE_INT offset
, size
, max_size
;
752 bool ptr
, unscalarizable_region
= false;
754 base
= get_ref_base_and_extent (expr
, &offset
, &size
, &max_size
);
756 if (sra_mode
== SRA_MODE_EARLY_IPA
757 && TREE_CODE (base
) == MEM_REF
)
759 base
= get_ssa_base_param (TREE_OPERAND (base
, 0));
767 if (!DECL_P (base
) || !bitmap_bit_p (candidate_bitmap
, DECL_UID (base
)))
770 if (sra_mode
== SRA_MODE_EARLY_IPA
)
772 if (size
< 0 || size
!= max_size
)
774 disqualify_candidate (base
, "Encountered a variable sized access.");
777 if (TREE_CODE (expr
) == COMPONENT_REF
778 && DECL_BIT_FIELD (TREE_OPERAND (expr
, 1)))
780 disqualify_candidate (base
, "Encountered a bit-field access.");
783 gcc_checking_assert ((offset
% BITS_PER_UNIT
) == 0);
786 mark_parm_dereference (base
, offset
+ size
, stmt
);
790 if (size
!= max_size
)
793 unscalarizable_region
= true;
797 disqualify_candidate (base
, "Encountered an unconstrained access.");
802 access
= create_access_1 (base
, offset
, size
);
804 access
->type
= TREE_TYPE (expr
);
805 access
->write
= write
;
806 access
->grp_unscalarizable_region
= unscalarizable_region
;
813 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
814 register types or (recursively) records with only these two kinds of fields.
815 It also returns false if any of these records has a zero-size field as its
816 last field or has a bit-field. */
819 type_consists_of_records_p (tree type
)
822 bool last_fld_has_zero_size
= false;
824 if (TREE_CODE (type
) != RECORD_TYPE
)
827 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
828 if (TREE_CODE (fld
) == FIELD_DECL
)
830 tree ft
= TREE_TYPE (fld
);
832 if (DECL_BIT_FIELD (fld
))
835 if (!is_gimple_reg_type (ft
)
836 && !type_consists_of_records_p (ft
))
839 last_fld_has_zero_size
= tree_low_cst (DECL_SIZE (fld
), 1) == 0;
842 if (last_fld_has_zero_size
)
848 /* Create total_scalarization accesses for all scalar type fields in DECL that
849 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
850 must be the top-most VAR_DECL representing the variable, OFFSET must be the
851 offset of DECL within BASE. REF must be the memory reference expression for
855 completely_scalarize_record (tree base
, tree decl
, HOST_WIDE_INT offset
,
858 tree fld
, decl_type
= TREE_TYPE (decl
);
860 for (fld
= TYPE_FIELDS (decl_type
); fld
; fld
= DECL_CHAIN (fld
))
861 if (TREE_CODE (fld
) == FIELD_DECL
)
863 HOST_WIDE_INT pos
= offset
+ int_bit_position (fld
);
864 tree ft
= TREE_TYPE (fld
);
865 tree nref
= build3 (COMPONENT_REF
, TREE_TYPE (fld
), ref
, fld
,
868 if (is_gimple_reg_type (ft
))
870 struct access
*access
;
873 size
= tree_low_cst (DECL_SIZE (fld
), 1);
874 access
= create_access_1 (base
, pos
, size
);
877 access
->total_scalarization
= 1;
878 /* Accesses for intraprocedural SRA can have their stmt NULL. */
881 completely_scalarize_record (base
, fld
, pos
, nref
);
886 /* Search the given tree for a declaration by skipping handled components and
887 exclude it from the candidates. */
890 disqualify_base_of_expr (tree t
, const char *reason
)
892 t
= get_base_address (t
);
893 if (sra_mode
== SRA_MODE_EARLY_IPA
894 && TREE_CODE (t
) == MEM_REF
)
895 t
= get_ssa_base_param (TREE_OPERAND (t
, 0));
898 disqualify_candidate (t
, reason
);
901 /* Scan expression EXPR and create access structures for all accesses to
902 candidates for scalarization. Return the created access or NULL if none is
905 static struct access
*
906 build_access_from_expr_1 (tree expr
, gimple stmt
, bool write
)
908 struct access
*ret
= NULL
;
911 if (TREE_CODE (expr
) == BIT_FIELD_REF
912 || TREE_CODE (expr
) == IMAGPART_EXPR
913 || TREE_CODE (expr
) == REALPART_EXPR
)
915 expr
= TREE_OPERAND (expr
, 0);
921 /* We need to dive through V_C_Es in order to get the size of its parameter
922 and not the result type. Ada produces such statements. We are also
923 capable of handling the topmost V_C_E but not any of those buried in other
924 handled components. */
925 if (TREE_CODE (expr
) == VIEW_CONVERT_EXPR
)
926 expr
= TREE_OPERAND (expr
, 0);
928 if (contains_view_convert_expr_p (expr
))
930 disqualify_base_of_expr (expr
, "V_C_E under a different handled "
935 switch (TREE_CODE (expr
))
938 if (TREE_CODE (TREE_OPERAND (expr
, 0)) != ADDR_EXPR
939 && sra_mode
!= SRA_MODE_EARLY_IPA
)
947 case ARRAY_RANGE_REF
:
948 ret
= create_access (expr
, stmt
, write
);
955 if (write
&& partial_ref
&& ret
)
956 ret
->grp_partial_lhs
= 1;
961 /* Scan expression EXPR and create access structures for all accesses to
962 candidates for scalarization. Return true if any access has been inserted.
963 STMT must be the statement from which the expression is taken, WRITE must be
964 true if the expression is a store and false otherwise. */
967 build_access_from_expr (tree expr
, gimple stmt
, bool write
)
969 struct access
*access
;
971 access
= build_access_from_expr_1 (expr
, stmt
, write
);
974 /* This means the aggregate is accesses as a whole in a way other than an
975 assign statement and thus cannot be removed even if we had a scalar
976 replacement for everything. */
977 if (cannot_scalarize_away_bitmap
)
978 bitmap_set_bit (cannot_scalarize_away_bitmap
, DECL_UID (access
->base
));
984 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
985 modes in which it matters, return true iff they have been disqualified. RHS
986 may be NULL, in that case ignore it. If we scalarize an aggregate in
987 intra-SRA we may need to add statements after each statement. This is not
988 possible if a statement unconditionally has to end the basic block. */
990 disqualify_ops_if_throwing_stmt (gimple stmt
, tree lhs
, tree rhs
)
992 if ((sra_mode
== SRA_MODE_EARLY_INTRA
|| sra_mode
== SRA_MODE_INTRA
)
993 && (stmt_can_throw_internal (stmt
) || stmt_ends_bb_p (stmt
)))
995 disqualify_base_of_expr (lhs
, "LHS of a throwing stmt.");
997 disqualify_base_of_expr (rhs
, "RHS of a throwing stmt.");
1003 /* Scan expressions occuring in STMT, create access structures for all accesses
1004 to candidates for scalarization and remove those candidates which occur in
1005 statements or expressions that prevent them from being split apart. Return
1006 true if any access has been inserted. */
1009 build_accesses_from_assign (gimple stmt
)
1012 struct access
*lacc
, *racc
;
1014 if (!gimple_assign_single_p (stmt
))
1017 lhs
= gimple_assign_lhs (stmt
);
1018 rhs
= gimple_assign_rhs1 (stmt
);
1020 if (disqualify_ops_if_throwing_stmt (stmt
, lhs
, rhs
))
1023 racc
= build_access_from_expr_1 (rhs
, stmt
, false);
1024 lacc
= build_access_from_expr_1 (lhs
, stmt
, true);
1028 racc
->grp_assignment_read
= 1;
1029 if (should_scalarize_away_bitmap
&& !gimple_has_volatile_ops (stmt
)
1030 && !is_gimple_reg_type (racc
->type
))
1031 bitmap_set_bit (should_scalarize_away_bitmap
, DECL_UID (racc
->base
));
1035 && (sra_mode
== SRA_MODE_EARLY_INTRA
|| sra_mode
== SRA_MODE_INTRA
)
1036 && !lacc
->grp_unscalarizable_region
1037 && !racc
->grp_unscalarizable_region
1038 && AGGREGATE_TYPE_P (TREE_TYPE (lhs
))
1039 /* FIXME: Turn the following line into an assert after PR 40058 is
1041 && lacc
->size
== racc
->size
1042 && useless_type_conversion_p (lacc
->type
, racc
->type
))
1044 struct assign_link
*link
;
1046 link
= (struct assign_link
*) pool_alloc (link_pool
);
1047 memset (link
, 0, sizeof (struct assign_link
));
1052 add_link_to_rhs (racc
, link
);
1055 return lacc
|| racc
;
1058 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1059 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1062 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED
, tree op
,
1063 void *data ATTRIBUTE_UNUSED
)
1065 op
= get_base_address (op
);
1068 disqualify_candidate (op
, "Non-scalarizable GIMPLE_ASM operand.");
1073 /* Return true iff callsite CALL has at least as many actual arguments as there
1074 are formal parameters of the function currently processed by IPA-SRA. */
1077 callsite_has_enough_arguments_p (gimple call
)
1079 return gimple_call_num_args (call
) >= (unsigned) func_param_count
;
1082 /* Scan function and look for interesting expressions and create access
1083 structures for them. Return true iff any access is created. */
1086 scan_function (void)
1093 gimple_stmt_iterator gsi
;
1094 for (gsi
= gsi_start_bb (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
1096 gimple stmt
= gsi_stmt (gsi
);
1100 if (final_bbs
&& stmt_can_throw_external (stmt
))
1101 bitmap_set_bit (final_bbs
, bb
->index
);
1102 switch (gimple_code (stmt
))
1105 t
= gimple_return_retval (stmt
);
1107 ret
|= build_access_from_expr (t
, stmt
, false);
1109 bitmap_set_bit (final_bbs
, bb
->index
);
1113 ret
|= build_accesses_from_assign (stmt
);
1117 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
1118 ret
|= build_access_from_expr (gimple_call_arg (stmt
, i
),
1121 if (sra_mode
== SRA_MODE_EARLY_IPA
)
1123 tree dest
= gimple_call_fndecl (stmt
);
1124 int flags
= gimple_call_flags (stmt
);
1128 if (DECL_BUILT_IN_CLASS (dest
) == BUILT_IN_NORMAL
1129 && DECL_FUNCTION_CODE (dest
) == BUILT_IN_APPLY_ARGS
)
1130 encountered_apply_args
= true;
1131 if (cgraph_get_node (dest
)
1132 == cgraph_get_node (current_function_decl
))
1134 encountered_recursive_call
= true;
1135 if (!callsite_has_enough_arguments_p (stmt
))
1136 encountered_unchangable_recursive_call
= true;
1141 && (flags
& (ECF_CONST
| ECF_PURE
)) == 0)
1142 bitmap_set_bit (final_bbs
, bb
->index
);
1145 t
= gimple_call_lhs (stmt
);
1146 if (t
&& !disqualify_ops_if_throwing_stmt (stmt
, t
, NULL
))
1147 ret
|= build_access_from_expr (t
, stmt
, true);
1151 walk_stmt_load_store_addr_ops (stmt
, NULL
, NULL
, NULL
,
1154 bitmap_set_bit (final_bbs
, bb
->index
);
1156 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
1158 t
= TREE_VALUE (gimple_asm_input_op (stmt
, i
));
1159 ret
|= build_access_from_expr (t
, stmt
, false);
1161 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
1163 t
= TREE_VALUE (gimple_asm_output_op (stmt
, i
));
1164 ret
|= build_access_from_expr (t
, stmt
, true);
1177 /* Helper of QSORT function. There are pointers to accesses in the array. An
1178 access is considered smaller than another if it has smaller offset or if the
1179 offsets are the same but is size is bigger. */
1182 compare_access_positions (const void *a
, const void *b
)
1184 const access_p
*fp1
= (const access_p
*) a
;
1185 const access_p
*fp2
= (const access_p
*) b
;
1186 const access_p f1
= *fp1
;
1187 const access_p f2
= *fp2
;
1189 if (f1
->offset
!= f2
->offset
)
1190 return f1
->offset
< f2
->offset
? -1 : 1;
1192 if (f1
->size
== f2
->size
)
1194 if (f1
->type
== f2
->type
)
1196 /* Put any non-aggregate type before any aggregate type. */
1197 else if (!is_gimple_reg_type (f1
->type
)
1198 && is_gimple_reg_type (f2
->type
))
1200 else if (is_gimple_reg_type (f1
->type
)
1201 && !is_gimple_reg_type (f2
->type
))
1203 /* Put any complex or vector type before any other scalar type. */
1204 else if (TREE_CODE (f1
->type
) != COMPLEX_TYPE
1205 && TREE_CODE (f1
->type
) != VECTOR_TYPE
1206 && (TREE_CODE (f2
->type
) == COMPLEX_TYPE
1207 || TREE_CODE (f2
->type
) == VECTOR_TYPE
))
1209 else if ((TREE_CODE (f1
->type
) == COMPLEX_TYPE
1210 || TREE_CODE (f1
->type
) == VECTOR_TYPE
)
1211 && TREE_CODE (f2
->type
) != COMPLEX_TYPE
1212 && TREE_CODE (f2
->type
) != VECTOR_TYPE
)
1214 /* Put the integral type with the bigger precision first. */
1215 else if (INTEGRAL_TYPE_P (f1
->type
)
1216 && INTEGRAL_TYPE_P (f2
->type
))
1217 return TYPE_PRECISION (f2
->type
) - TYPE_PRECISION (f1
->type
);
1218 /* Put any integral type with non-full precision last. */
1219 else if (INTEGRAL_TYPE_P (f1
->type
)
1220 && (TREE_INT_CST_LOW (TYPE_SIZE (f1
->type
))
1221 != TYPE_PRECISION (f1
->type
)))
1223 else if (INTEGRAL_TYPE_P (f2
->type
)
1224 && (TREE_INT_CST_LOW (TYPE_SIZE (f2
->type
))
1225 != TYPE_PRECISION (f2
->type
)))
1227 /* Stabilize the sort. */
1228 return TYPE_UID (f1
->type
) - TYPE_UID (f2
->type
);
1231 /* We want the bigger accesses first, thus the opposite operator in the next
1233 return f1
->size
> f2
->size
? -1 : 1;
1237 /* Append a name of the declaration to the name obstack. A helper function for
1241 make_fancy_decl_name (tree decl
)
1245 tree name
= DECL_NAME (decl
);
1247 obstack_grow (&name_obstack
, IDENTIFIER_POINTER (name
),
1248 IDENTIFIER_LENGTH (name
));
1251 sprintf (buffer
, "D%u", DECL_UID (decl
));
1252 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1256 /* Helper for make_fancy_name. */
1259 make_fancy_name_1 (tree expr
)
1266 make_fancy_decl_name (expr
);
1270 switch (TREE_CODE (expr
))
1273 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1274 obstack_1grow (&name_obstack
, '$');
1275 make_fancy_decl_name (TREE_OPERAND (expr
, 1));
1279 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1280 obstack_1grow (&name_obstack
, '$');
1281 /* Arrays with only one element may not have a constant as their
1283 index
= TREE_OPERAND (expr
, 1);
1284 if (TREE_CODE (index
) != INTEGER_CST
)
1286 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
, TREE_INT_CST_LOW (index
));
1287 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1291 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1295 make_fancy_name_1 (TREE_OPERAND (expr
, 0));
1296 if (!integer_zerop (TREE_OPERAND (expr
, 1)))
1298 obstack_1grow (&name_obstack
, '$');
1299 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
,
1300 TREE_INT_CST_LOW (TREE_OPERAND (expr
, 1)));
1301 obstack_grow (&name_obstack
, buffer
, strlen (buffer
));
1308 gcc_unreachable (); /* we treat these as scalars. */
1315 /* Create a human readable name for replacement variable of ACCESS. */
1318 make_fancy_name (tree expr
)
1320 make_fancy_name_1 (expr
);
1321 obstack_1grow (&name_obstack
, '\0');
1322 return XOBFINISH (&name_obstack
, char *);
1325 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1326 EXP_TYPE at the given OFFSET. If BASE is something for which
1327 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1328 to insert new statements either before or below the current one as specified
1329 by INSERT_AFTER. This function is not capable of handling bitfields. */
1332 build_ref_for_offset (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1333 tree exp_type
, gimple_stmt_iterator
*gsi
,
1336 tree prev_base
= base
;
1338 HOST_WIDE_INT base_offset
;
1340 gcc_checking_assert (offset
% BITS_PER_UNIT
== 0);
1342 base
= get_addr_base_and_unit_offset (base
, &base_offset
);
1344 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1345 offset such as array[var_index]. */
1351 gcc_checking_assert (gsi
);
1352 tmp
= create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base
)), NULL
);
1353 add_referenced_var (tmp
);
1354 tmp
= make_ssa_name (tmp
, NULL
);
1355 addr
= build_fold_addr_expr (unshare_expr (prev_base
));
1356 stmt
= gimple_build_assign (tmp
, addr
);
1357 gimple_set_location (stmt
, loc
);
1358 SSA_NAME_DEF_STMT (tmp
) = stmt
;
1360 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
1362 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
1365 off
= build_int_cst (reference_alias_ptr_type (prev_base
),
1366 offset
/ BITS_PER_UNIT
);
1369 else if (TREE_CODE (base
) == MEM_REF
)
1371 off
= build_int_cst (TREE_TYPE (TREE_OPERAND (base
, 1)),
1372 base_offset
+ offset
/ BITS_PER_UNIT
);
1373 off
= int_const_binop (PLUS_EXPR
, TREE_OPERAND (base
, 1), off
, 0);
1374 base
= unshare_expr (TREE_OPERAND (base
, 0));
1378 off
= build_int_cst (reference_alias_ptr_type (base
),
1379 base_offset
+ offset
/ BITS_PER_UNIT
);
1380 base
= build_fold_addr_expr (unshare_expr (base
));
1383 return fold_build2_loc (loc
, MEM_REF
, exp_type
, base
, off
);
1386 /* Construct a memory reference to a part of an aggregate BASE at the given
1387 OFFSET and of the same type as MODEL. In case this is a reference to a
1388 bit-field, the function will replicate the last component_ref of model's
1389 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1390 build_ref_for_offset. */
1393 build_ref_for_model (location_t loc
, tree base
, HOST_WIDE_INT offset
,
1394 struct access
*model
, gimple_stmt_iterator
*gsi
,
1397 if (TREE_CODE (model
->expr
) == COMPONENT_REF
1398 && DECL_BIT_FIELD (TREE_OPERAND (model
->expr
, 1)))
1400 /* This access represents a bit-field. */
1403 offset
-= int_bit_position (TREE_OPERAND (model
->expr
, 1));
1404 exp_type
= TREE_TYPE (TREE_OPERAND (model
->expr
, 0));
1405 t
= build_ref_for_offset (loc
, base
, offset
, exp_type
, gsi
, insert_after
);
1406 return fold_build3_loc (loc
, COMPONENT_REF
, model
->type
, t
,
1407 TREE_OPERAND (model
->expr
, 1), NULL_TREE
);
1410 return build_ref_for_offset (loc
, base
, offset
, model
->type
,
1414 /* Construct a memory reference consisting of component_refs and array_refs to
1415 a part of an aggregate *RES (which is of type TYPE). The requested part
1416 should have type EXP_TYPE at be the given OFFSET. This function might not
1417 succeed, it returns true when it does and only then *RES points to something
1418 meaningful. This function should be used only to build expressions that we
1419 might need to present to user (e.g. in warnings). In all other situations,
1420 build_ref_for_model or build_ref_for_offset should be used instead. */
1423 build_user_friendly_ref_for_offset (tree
*res
, tree type
, HOST_WIDE_INT offset
,
1429 tree tr_size
, index
, minidx
;
1430 HOST_WIDE_INT el_size
;
1432 if (offset
== 0 && exp_type
1433 && types_compatible_p (exp_type
, type
))
1436 switch (TREE_CODE (type
))
1439 case QUAL_UNION_TYPE
:
1441 for (fld
= TYPE_FIELDS (type
); fld
; fld
= DECL_CHAIN (fld
))
1443 HOST_WIDE_INT pos
, size
;
1444 tree expr
, *expr_ptr
;
1446 if (TREE_CODE (fld
) != FIELD_DECL
)
1449 pos
= int_bit_position (fld
);
1450 gcc_assert (TREE_CODE (type
) == RECORD_TYPE
|| pos
== 0);
1451 tr_size
= DECL_SIZE (fld
);
1452 if (!tr_size
|| !host_integerp (tr_size
, 1))
1454 size
= tree_low_cst (tr_size
, 1);
1460 else if (pos
> offset
|| (pos
+ size
) <= offset
)
1463 expr
= build3 (COMPONENT_REF
, TREE_TYPE (fld
), *res
, fld
,
1466 if (build_user_friendly_ref_for_offset (expr_ptr
, TREE_TYPE (fld
),
1467 offset
- pos
, exp_type
))
1476 tr_size
= TYPE_SIZE (TREE_TYPE (type
));
1477 if (!tr_size
|| !host_integerp (tr_size
, 1))
1479 el_size
= tree_low_cst (tr_size
, 1);
1481 minidx
= TYPE_MIN_VALUE (TYPE_DOMAIN (type
));
1482 if (TREE_CODE (minidx
) != INTEGER_CST
|| el_size
== 0)
1484 index
= build_int_cst (TYPE_DOMAIN (type
), offset
/ el_size
);
1485 if (!integer_zerop (minidx
))
1486 index
= int_const_binop (PLUS_EXPR
, index
, minidx
, 0);
1487 *res
= build4 (ARRAY_REF
, TREE_TYPE (type
), *res
, index
,
1488 NULL_TREE
, NULL_TREE
);
1489 offset
= offset
% el_size
;
1490 type
= TREE_TYPE (type
);
1505 /* Return true iff TYPE is stdarg va_list type. */
1508 is_va_list_type (tree type
)
1510 return TYPE_MAIN_VARIANT (type
) == TYPE_MAIN_VARIANT (va_list_type_node
);
1513 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1514 those with type which is suitable for scalarization. */
1517 find_var_candidates (void)
1520 referenced_var_iterator rvi
;
1523 FOR_EACH_REFERENCED_VAR (var
, rvi
)
1525 if (TREE_CODE (var
) != VAR_DECL
&& TREE_CODE (var
) != PARM_DECL
)
1527 type
= TREE_TYPE (var
);
1529 if (!AGGREGATE_TYPE_P (type
)
1530 || needs_to_live_in_memory (var
)
1531 || TREE_THIS_VOLATILE (var
)
1532 || !COMPLETE_TYPE_P (type
)
1533 || !host_integerp (TYPE_SIZE (type
), 1)
1534 || tree_low_cst (TYPE_SIZE (type
), 1) == 0
1535 || type_internals_preclude_sra_p (type
)
1536 /* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1537 we also want to schedule it rather late. Thus we ignore it in
1539 || (sra_mode
== SRA_MODE_EARLY_INTRA
1540 && is_va_list_type (type
)))
1543 bitmap_set_bit (candidate_bitmap
, DECL_UID (var
));
1545 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1547 fprintf (dump_file
, "Candidate (%d): ", DECL_UID (var
));
1548 print_generic_expr (dump_file
, var
, 0);
1549 fprintf (dump_file
, "\n");
1557 /* Sort all accesses for the given variable, check for partial overlaps and
1558 return NULL if there are any. If there are none, pick a representative for
1559 each combination of offset and size and create a linked list out of them.
1560 Return the pointer to the first representative and make sure it is the first
1561 one in the vector of accesses. */
1563 static struct access
*
1564 sort_and_splice_var_accesses (tree var
)
1566 int i
, j
, access_count
;
1567 struct access
*res
, **prev_acc_ptr
= &res
;
1568 VEC (access_p
, heap
) *access_vec
;
1570 HOST_WIDE_INT low
= -1, high
= 0;
1572 access_vec
= get_base_access_vector (var
);
1575 access_count
= VEC_length (access_p
, access_vec
);
1577 /* Sort by <OFFSET, SIZE>. */
1578 qsort (VEC_address (access_p
, access_vec
), access_count
, sizeof (access_p
),
1579 compare_access_positions
);
1582 while (i
< access_count
)
1584 struct access
*access
= VEC_index (access_p
, access_vec
, i
);
1585 bool grp_write
= access
->write
;
1586 bool grp_read
= !access
->write
;
1587 bool grp_assignment_read
= access
->grp_assignment_read
;
1588 bool multiple_reads
= false;
1589 bool total_scalarization
= access
->total_scalarization
;
1590 bool grp_partial_lhs
= access
->grp_partial_lhs
;
1591 bool first_scalar
= is_gimple_reg_type (access
->type
);
1592 bool unscalarizable_region
= access
->grp_unscalarizable_region
;
1594 if (first
|| access
->offset
>= high
)
1597 low
= access
->offset
;
1598 high
= access
->offset
+ access
->size
;
1600 else if (access
->offset
> low
&& access
->offset
+ access
->size
> high
)
1603 gcc_assert (access
->offset
>= low
1604 && access
->offset
+ access
->size
<= high
);
1607 while (j
< access_count
)
1609 struct access
*ac2
= VEC_index (access_p
, access_vec
, j
);
1610 if (ac2
->offset
!= access
->offset
|| ac2
->size
!= access
->size
)
1617 multiple_reads
= true;
1621 grp_assignment_read
|= ac2
->grp_assignment_read
;
1622 grp_partial_lhs
|= ac2
->grp_partial_lhs
;
1623 unscalarizable_region
|= ac2
->grp_unscalarizable_region
;
1624 total_scalarization
|= ac2
->total_scalarization
;
1625 relink_to_new_repr (access
, ac2
);
1627 /* If there are both aggregate-type and scalar-type accesses with
1628 this combination of size and offset, the comparison function
1629 should have put the scalars first. */
1630 gcc_assert (first_scalar
|| !is_gimple_reg_type (ac2
->type
));
1631 ac2
->group_representative
= access
;
1637 access
->group_representative
= access
;
1638 access
->grp_write
= grp_write
;
1639 access
->grp_read
= grp_read
;
1640 access
->grp_assignment_read
= grp_assignment_read
;
1641 access
->grp_hint
= multiple_reads
|| total_scalarization
;
1642 access
->grp_partial_lhs
= grp_partial_lhs
;
1643 access
->grp_unscalarizable_region
= unscalarizable_region
;
1644 if (access
->first_link
)
1645 add_access_to_work_queue (access
);
1647 *prev_acc_ptr
= access
;
1648 prev_acc_ptr
= &access
->next_grp
;
1651 gcc_assert (res
== VEC_index (access_p
, access_vec
, 0));
1655 /* Create a variable for the given ACCESS which determines the type, name and a
1656 few other properties. Return the variable declaration and store it also to
1657 ACCESS->replacement. */
1660 create_access_replacement (struct access
*access
, bool rename
)
1664 repl
= create_tmp_var (access
->type
, "SR");
1666 add_referenced_var (repl
);
1668 mark_sym_for_renaming (repl
);
1670 if (!access
->grp_partial_lhs
1671 && (TREE_CODE (access
->type
) == COMPLEX_TYPE
1672 || TREE_CODE (access
->type
) == VECTOR_TYPE
))
1673 DECL_GIMPLE_REG_P (repl
) = 1;
1675 DECL_SOURCE_LOCATION (repl
) = DECL_SOURCE_LOCATION (access
->base
);
1676 DECL_ARTIFICIAL (repl
) = 1;
1677 DECL_IGNORED_P (repl
) = DECL_IGNORED_P (access
->base
);
1679 if (DECL_NAME (access
->base
)
1680 && !DECL_IGNORED_P (access
->base
)
1681 && !DECL_ARTIFICIAL (access
->base
))
1683 char *pretty_name
= make_fancy_name (access
->expr
);
1684 tree debug_expr
= unshare_expr (access
->expr
), d
;
1686 DECL_NAME (repl
) = get_identifier (pretty_name
);
1687 obstack_free (&name_obstack
, pretty_name
);
1689 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1690 as DECL_DEBUG_EXPR isn't considered when looking for still
1691 used SSA_NAMEs and thus they could be freed. All debug info
1692 generation cares is whether something is constant or variable
1693 and that get_ref_base_and_extent works properly on the
1695 for (d
= debug_expr
; handled_component_p (d
); d
= TREE_OPERAND (d
, 0))
1696 switch (TREE_CODE (d
))
1699 case ARRAY_RANGE_REF
:
1700 if (TREE_OPERAND (d
, 1)
1701 && TREE_CODE (TREE_OPERAND (d
, 1)) == SSA_NAME
)
1702 TREE_OPERAND (d
, 1) = SSA_NAME_VAR (TREE_OPERAND (d
, 1));
1703 if (TREE_OPERAND (d
, 3)
1704 && TREE_CODE (TREE_OPERAND (d
, 3)) == SSA_NAME
)
1705 TREE_OPERAND (d
, 3) = SSA_NAME_VAR (TREE_OPERAND (d
, 3));
1708 if (TREE_OPERAND (d
, 2)
1709 && TREE_CODE (TREE_OPERAND (d
, 2)) == SSA_NAME
)
1710 TREE_OPERAND (d
, 2) = SSA_NAME_VAR (TREE_OPERAND (d
, 2));
1715 SET_DECL_DEBUG_EXPR (repl
, debug_expr
);
1716 DECL_DEBUG_EXPR_IS_FROM (repl
) = 1;
1717 TREE_NO_WARNING (repl
) = TREE_NO_WARNING (access
->base
);
1720 TREE_NO_WARNING (repl
) = 1;
1724 fprintf (dump_file
, "Created a replacement for ");
1725 print_generic_expr (dump_file
, access
->base
, 0);
1726 fprintf (dump_file
, " offset: %u, size: %u: ",
1727 (unsigned) access
->offset
, (unsigned) access
->size
);
1728 print_generic_expr (dump_file
, repl
, 0);
1729 fprintf (dump_file
, "\n");
1731 sra_stats
.replacements
++;
1736 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1739 get_access_replacement (struct access
*access
)
1741 gcc_assert (access
->grp_to_be_replaced
);
1743 if (!access
->replacement_decl
)
1744 access
->replacement_decl
= create_access_replacement (access
, true);
1745 return access
->replacement_decl
;
1748 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1749 not mark it for renaming. */
1752 get_unrenamed_access_replacement (struct access
*access
)
1754 gcc_assert (!access
->grp_to_be_replaced
);
1756 if (!access
->replacement_decl
)
1757 access
->replacement_decl
= create_access_replacement (access
, false);
1758 return access
->replacement_decl
;
1762 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1763 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1764 to it is not "within" the root. Return false iff some accesses partially
1768 build_access_subtree (struct access
**access
)
1770 struct access
*root
= *access
, *last_child
= NULL
;
1771 HOST_WIDE_INT limit
= root
->offset
+ root
->size
;
1773 *access
= (*access
)->next_grp
;
1774 while (*access
&& (*access
)->offset
+ (*access
)->size
<= limit
)
1777 root
->first_child
= *access
;
1779 last_child
->next_sibling
= *access
;
1780 last_child
= *access
;
1782 if (!build_access_subtree (access
))
1786 if (*access
&& (*access
)->offset
< limit
)
1792 /* Build a tree of access representatives, ACCESS is the pointer to the first
1793 one, others are linked in a list by the next_grp field. Return false iff
1794 some accesses partially overlap. */
1797 build_access_trees (struct access
*access
)
1801 struct access
*root
= access
;
1803 if (!build_access_subtree (&access
))
1805 root
->next_grp
= access
;
1810 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1814 expr_with_var_bounded_array_refs_p (tree expr
)
1816 while (handled_component_p (expr
))
1818 if (TREE_CODE (expr
) == ARRAY_REF
1819 && !host_integerp (array_ref_low_bound (expr
), 0))
1821 expr
= TREE_OPERAND (expr
, 0);
1826 enum mark_read_status
{ SRA_MR_NOT_READ
, SRA_MR_READ
, SRA_MR_ASSIGN_READ
};
1828 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1829 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1830 sorts of access flags appropriately along the way, notably always set
1831 grp_read and grp_assign_read according to MARK_READ and grp_write when
1832 MARK_WRITE is true. */
1835 analyze_access_subtree (struct access
*root
, bool allow_replacements
,
1836 enum mark_read_status mark_read
, bool mark_write
)
1838 struct access
*child
;
1839 HOST_WIDE_INT limit
= root
->offset
+ root
->size
;
1840 HOST_WIDE_INT covered_to
= root
->offset
;
1841 bool scalar
= is_gimple_reg_type (root
->type
);
1842 bool hole
= false, sth_created
= false;
1843 bool direct_read
= root
->grp_read
;
1845 if (mark_read
== SRA_MR_ASSIGN_READ
)
1848 root
->grp_assignment_read
= 1;
1850 if (mark_read
== SRA_MR_READ
)
1852 else if (root
->grp_assignment_read
)
1853 mark_read
= SRA_MR_ASSIGN_READ
;
1854 else if (root
->grp_read
)
1855 mark_read
= SRA_MR_READ
;
1858 root
->grp_write
= true;
1859 else if (root
->grp_write
)
1862 if (root
->grp_unscalarizable_region
)
1863 allow_replacements
= false;
1865 if (allow_replacements
&& expr_with_var_bounded_array_refs_p (root
->expr
))
1866 allow_replacements
= false;
1868 for (child
= root
->first_child
; child
; child
= child
->next_sibling
)
1870 if (!hole
&& child
->offset
< covered_to
)
1873 covered_to
+= child
->size
;
1875 sth_created
|= analyze_access_subtree (child
,
1876 allow_replacements
&& !scalar
,
1877 mark_read
, mark_write
);
1879 root
->grp_unscalarized_data
|= child
->grp_unscalarized_data
;
1880 hole
|= !child
->grp_covered
;
1883 if (allow_replacements
&& scalar
&& !root
->first_child
1885 || (root
->grp_write
&& (direct_read
|| root
->grp_assignment_read
))))
1887 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1889 fprintf (dump_file
, "Marking ");
1890 print_generic_expr (dump_file
, root
->base
, 0);
1891 fprintf (dump_file
, " offset: %u, size: %u: ",
1892 (unsigned) root
->offset
, (unsigned) root
->size
);
1893 fprintf (dump_file
, " to be replaced.\n");
1896 root
->grp_to_be_replaced
= 1;
1900 else if (covered_to
< limit
)
1903 if (sth_created
&& !hole
)
1905 root
->grp_covered
= 1;
1908 if (root
->grp_write
|| TREE_CODE (root
->base
) == PARM_DECL
)
1909 root
->grp_unscalarized_data
= 1; /* not covered and written to */
1915 /* Analyze all access trees linked by next_grp by the means of
1916 analyze_access_subtree. */
1918 analyze_access_trees (struct access
*access
)
1924 if (analyze_access_subtree (access
, true, SRA_MR_NOT_READ
, false))
1926 access
= access
->next_grp
;
1932 /* Return true iff a potential new child of LACC at offset OFFSET and with size
1933 SIZE would conflict with an already existing one. If exactly such a child
1934 already exists in LACC, store a pointer to it in EXACT_MATCH. */
1937 child_would_conflict_in_lacc (struct access
*lacc
, HOST_WIDE_INT norm_offset
,
1938 HOST_WIDE_INT size
, struct access
**exact_match
)
1940 struct access
*child
;
1942 for (child
= lacc
->first_child
; child
; child
= child
->next_sibling
)
1944 if (child
->offset
== norm_offset
&& child
->size
== size
)
1946 *exact_match
= child
;
1950 if (child
->offset
< norm_offset
+ size
1951 && child
->offset
+ child
->size
> norm_offset
)
1958 /* Create a new child access of PARENT, with all properties just like MODEL
1959 except for its offset and with its grp_write false and grp_read true.
1960 Return the new access or NULL if it cannot be created. Note that this access
1961 is created long after all splicing and sorting, it's not located in any
1962 access vector and is automatically a representative of its group. */
1964 static struct access
*
1965 create_artificial_child_access (struct access
*parent
, struct access
*model
,
1966 HOST_WIDE_INT new_offset
)
1968 struct access
*access
;
1969 struct access
**child
;
1970 tree expr
= parent
->base
;
1972 gcc_assert (!model
->grp_unscalarizable_region
);
1973 if (!build_user_friendly_ref_for_offset (&expr
, TREE_TYPE (expr
), new_offset
,
1977 access
= (struct access
*) pool_alloc (access_pool
);
1978 memset (access
, 0, sizeof (struct access
));
1979 access
->base
= parent
->base
;
1980 access
->expr
= expr
;
1981 access
->offset
= new_offset
;
1982 access
->size
= model
->size
;
1983 access
->type
= model
->type
;
1984 access
->grp_write
= true;
1985 access
->grp_read
= false;
1987 child
= &parent
->first_child
;
1988 while (*child
&& (*child
)->offset
< new_offset
)
1989 child
= &(*child
)->next_sibling
;
1991 access
->next_sibling
= *child
;
1998 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
1999 true if any new subaccess was created. Additionally, if RACC is a scalar
2000 access but LACC is not, change the type of the latter, if possible. */
2003 propagate_subaccesses_across_link (struct access
*lacc
, struct access
*racc
)
2005 struct access
*rchild
;
2006 HOST_WIDE_INT norm_delta
= lacc
->offset
- racc
->offset
;
2009 if (is_gimple_reg_type (lacc
->type
)
2010 || lacc
->grp_unscalarizable_region
2011 || racc
->grp_unscalarizable_region
)
2014 if (!lacc
->first_child
&& !racc
->first_child
2015 && is_gimple_reg_type (racc
->type
))
2017 tree t
= lacc
->base
;
2019 if (build_user_friendly_ref_for_offset (&t
, TREE_TYPE (t
), lacc
->offset
,
2023 lacc
->type
= racc
->type
;
2028 for (rchild
= racc
->first_child
; rchild
; rchild
= rchild
->next_sibling
)
2030 struct access
*new_acc
= NULL
;
2031 HOST_WIDE_INT norm_offset
= rchild
->offset
+ norm_delta
;
2033 if (rchild
->grp_unscalarizable_region
)
2036 if (child_would_conflict_in_lacc (lacc
, norm_offset
, rchild
->size
,
2041 rchild
->grp_hint
= 1;
2042 new_acc
->grp_hint
|= new_acc
->grp_read
;
2043 if (rchild
->first_child
)
2044 ret
|= propagate_subaccesses_across_link (new_acc
, rchild
);
2049 rchild
->grp_hint
= 1;
2050 new_acc
= create_artificial_child_access (lacc
, rchild
, norm_offset
);
2054 if (racc
->first_child
)
2055 propagate_subaccesses_across_link (new_acc
, rchild
);
2062 /* Propagate all subaccesses across assignment links. */
2065 propagate_all_subaccesses (void)
2067 while (work_queue_head
)
2069 struct access
*racc
= pop_access_from_work_queue ();
2070 struct assign_link
*link
;
2072 gcc_assert (racc
->first_link
);
2074 for (link
= racc
->first_link
; link
; link
= link
->next
)
2076 struct access
*lacc
= link
->lacc
;
2078 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (lacc
->base
)))
2080 lacc
= lacc
->group_representative
;
2081 if (propagate_subaccesses_across_link (lacc
, racc
)
2082 && lacc
->first_link
)
2083 add_access_to_work_queue (lacc
);
2088 /* Go through all accesses collected throughout the (intraprocedural) analysis
2089 stage, exclude overlapping ones, identify representatives and build trees
2090 out of them, making decisions about scalarization on the way. Return true
2091 iff there are any to-be-scalarized variables after this stage. */
2094 analyze_all_variable_accesses (void)
2097 bitmap tmp
= BITMAP_ALLOC (NULL
);
2099 unsigned i
, max_total_scalarization_size
;
2101 max_total_scalarization_size
= UNITS_PER_WORD
* BITS_PER_UNIT
2102 * MOVE_RATIO (optimize_function_for_speed_p (cfun
));
2104 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap
, 0, i
, bi
)
2105 if (bitmap_bit_p (should_scalarize_away_bitmap
, i
)
2106 && !bitmap_bit_p (cannot_scalarize_away_bitmap
, i
))
2108 tree var
= referenced_var (i
);
2110 if (TREE_CODE (var
) == VAR_DECL
2111 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var
)), 1)
2112 <= max_total_scalarization_size
)
2113 && type_consists_of_records_p (TREE_TYPE (var
)))
2115 completely_scalarize_record (var
, var
, 0, var
);
2116 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2118 fprintf (dump_file
, "Will attempt to totally scalarize ");
2119 print_generic_expr (dump_file
, var
, 0);
2120 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (var
));
2125 bitmap_copy (tmp
, candidate_bitmap
);
2126 EXECUTE_IF_SET_IN_BITMAP (tmp
, 0, i
, bi
)
2128 tree var
= referenced_var (i
);
2129 struct access
*access
;
2131 access
= sort_and_splice_var_accesses (var
);
2132 if (!access
|| !build_access_trees (access
))
2133 disqualify_candidate (var
,
2134 "No or inhibitingly overlapping accesses.");
2137 propagate_all_subaccesses ();
2139 bitmap_copy (tmp
, candidate_bitmap
);
2140 EXECUTE_IF_SET_IN_BITMAP (tmp
, 0, i
, bi
)
2142 tree var
= referenced_var (i
);
2143 struct access
*access
= get_first_repr_for_decl (var
);
2145 if (analyze_access_trees (access
))
2148 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
2150 fprintf (dump_file
, "\nAccess trees for ");
2151 print_generic_expr (dump_file
, var
, 0);
2152 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (var
));
2153 dump_access_tree (dump_file
, access
);
2154 fprintf (dump_file
, "\n");
2158 disqualify_candidate (var
, "No scalar replacements to be created.");
2165 statistics_counter_event (cfun
, "Scalarized aggregates", res
);
2172 /* Generate statements copying scalar replacements of accesses within a subtree
2173 into or out of AGG. ACCESS, all its children, siblings and their children
2174 are to be processed. AGG is an aggregate type expression (can be a
2175 declaration but does not have to be, it can for example also be a mem_ref or
2176 a series of handled components). TOP_OFFSET is the offset of the processed
2177 subtree which has to be subtracted from offsets of individual accesses to
2178 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2179 replacements in the interval <start_offset, start_offset + chunk_size>,
2180 otherwise copy all. GSI is a statement iterator used to place the new
2181 statements. WRITE should be true when the statements should write from AGG
2182 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2183 statements will be added after the current statement in GSI, they will be
2184 added before the statement otherwise. */
2187 generate_subtree_copies (struct access
*access
, tree agg
,
2188 HOST_WIDE_INT top_offset
,
2189 HOST_WIDE_INT start_offset
, HOST_WIDE_INT chunk_size
,
2190 gimple_stmt_iterator
*gsi
, bool write
,
2191 bool insert_after
, location_t loc
)
2195 if (chunk_size
&& access
->offset
>= start_offset
+ chunk_size
)
2198 if (access
->grp_to_be_replaced
2200 || access
->offset
+ access
->size
> start_offset
))
2202 tree expr
, repl
= get_access_replacement (access
);
2205 expr
= build_ref_for_model (loc
, agg
, access
->offset
- top_offset
,
2206 access
, gsi
, insert_after
);
2210 if (access
->grp_partial_lhs
)
2211 expr
= force_gimple_operand_gsi (gsi
, expr
, true, NULL_TREE
,
2213 insert_after
? GSI_NEW_STMT
2215 stmt
= gimple_build_assign (repl
, expr
);
2219 TREE_NO_WARNING (repl
) = 1;
2220 if (access
->grp_partial_lhs
)
2221 repl
= force_gimple_operand_gsi (gsi
, repl
, true, NULL_TREE
,
2223 insert_after
? GSI_NEW_STMT
2225 stmt
= gimple_build_assign (expr
, repl
);
2227 gimple_set_location (stmt
, loc
);
2230 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2232 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2234 sra_stats
.subtree_copies
++;
2237 if (access
->first_child
)
2238 generate_subtree_copies (access
->first_child
, agg
, top_offset
,
2239 start_offset
, chunk_size
, gsi
,
2240 write
, insert_after
, loc
);
2242 access
= access
->next_sibling
;
2247 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2248 the root of the subtree to be processed. GSI is the statement iterator used
2249 for inserting statements which are added after the current statement if
2250 INSERT_AFTER is true or before it otherwise. */
2253 init_subtree_with_zero (struct access
*access
, gimple_stmt_iterator
*gsi
,
2254 bool insert_after
, location_t loc
)
2257 struct access
*child
;
2259 if (access
->grp_to_be_replaced
)
2263 stmt
= gimple_build_assign (get_access_replacement (access
),
2264 fold_convert (access
->type
,
2265 integer_zero_node
));
2267 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2269 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2271 gimple_set_location (stmt
, loc
);
2274 for (child
= access
->first_child
; child
; child
= child
->next_sibling
)
2275 init_subtree_with_zero (child
, gsi
, insert_after
, loc
);
2278 /* Search for an access representative for the given expression EXPR and
2279 return it or NULL if it cannot be found. */
2281 static struct access
*
2282 get_access_for_expr (tree expr
)
2284 HOST_WIDE_INT offset
, size
, max_size
;
2287 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2288 a different size than the size of its argument and we need the latter
2290 if (TREE_CODE (expr
) == VIEW_CONVERT_EXPR
)
2291 expr
= TREE_OPERAND (expr
, 0);
2293 base
= get_ref_base_and_extent (expr
, &offset
, &size
, &max_size
);
2294 if (max_size
== -1 || !DECL_P (base
))
2297 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (base
)))
2300 return get_var_base_offset_size_access (base
, offset
, max_size
);
2303 /* Replace the expression EXPR with a scalar replacement if there is one and
2304 generate other statements to do type conversion or subtree copying if
2305 necessary. GSI is used to place newly created statements, WRITE is true if
2306 the expression is being written to (it is on a LHS of a statement or output
2307 in an assembly statement). */
2310 sra_modify_expr (tree
*expr
, gimple_stmt_iterator
*gsi
, bool write
)
2313 struct access
*access
;
2316 if (TREE_CODE (*expr
) == BIT_FIELD_REF
)
2319 expr
= &TREE_OPERAND (*expr
, 0);
2324 if (TREE_CODE (*expr
) == REALPART_EXPR
|| TREE_CODE (*expr
) == IMAGPART_EXPR
)
2325 expr
= &TREE_OPERAND (*expr
, 0);
2326 access
= get_access_for_expr (*expr
);
2329 type
= TREE_TYPE (*expr
);
2331 loc
= gimple_location (gsi_stmt (*gsi
));
2332 if (access
->grp_to_be_replaced
)
2334 tree repl
= get_access_replacement (access
);
2335 /* If we replace a non-register typed access simply use the original
2336 access expression to extract the scalar component afterwards.
2337 This happens if scalarizing a function return value or parameter
2338 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2339 gcc.c-torture/compile/20011217-1.c.
2341 We also want to use this when accessing a complex or vector which can
2342 be accessed as a different type too, potentially creating a need for
2343 type conversion (see PR42196) and when scalarized unions are involved
2344 in assembler statements (see PR42398). */
2345 if (!useless_type_conversion_p (type
, access
->type
))
2349 ref
= build_ref_for_model (loc
, access
->base
, access
->offset
, access
,
2356 if (access
->grp_partial_lhs
)
2357 ref
= force_gimple_operand_gsi (gsi
, ref
, true, NULL_TREE
,
2358 false, GSI_NEW_STMT
);
2359 stmt
= gimple_build_assign (repl
, ref
);
2360 gimple_set_location (stmt
, loc
);
2361 gsi_insert_after (gsi
, stmt
, GSI_NEW_STMT
);
2367 if (access
->grp_partial_lhs
)
2368 repl
= force_gimple_operand_gsi (gsi
, repl
, true, NULL_TREE
,
2369 true, GSI_SAME_STMT
);
2370 stmt
= gimple_build_assign (ref
, repl
);
2371 gimple_set_location (stmt
, loc
);
2372 gsi_insert_before (gsi
, stmt
, GSI_SAME_STMT
);
2380 if (access
->first_child
)
2382 HOST_WIDE_INT start_offset
, chunk_size
;
2384 && host_integerp (TREE_OPERAND (bfr
, 1), 1)
2385 && host_integerp (TREE_OPERAND (bfr
, 2), 1))
2387 chunk_size
= tree_low_cst (TREE_OPERAND (bfr
, 1), 1);
2388 start_offset
= access
->offset
2389 + tree_low_cst (TREE_OPERAND (bfr
, 2), 1);
2392 start_offset
= chunk_size
= 0;
2394 generate_subtree_copies (access
->first_child
, access
->base
, 0,
2395 start_offset
, chunk_size
, gsi
, write
, write
,
2401 /* Where scalar replacements of the RHS have been written to when a replacement
2402 of a LHS of an assigments cannot be direclty loaded from a replacement of
2404 enum unscalarized_data_handling
{ SRA_UDH_NONE
, /* Nothing done so far. */
2405 SRA_UDH_RIGHT
, /* Data flushed to the RHS. */
2406 SRA_UDH_LEFT
}; /* Data flushed to the LHS. */
2408 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2409 base aggregate if there are unscalarized data or directly to LHS of the
2410 statement that is pointed to by GSI otherwise. */
2412 static enum unscalarized_data_handling
2413 handle_unscalarized_data_in_subtree (struct access
*top_racc
,
2414 gimple_stmt_iterator
*gsi
)
2416 if (top_racc
->grp_unscalarized_data
)
2418 generate_subtree_copies (top_racc
->first_child
, top_racc
->base
, 0, 0, 0,
2420 gimple_location (gsi_stmt (*gsi
)));
2421 return SRA_UDH_RIGHT
;
2425 tree lhs
= gimple_assign_lhs (gsi_stmt (*gsi
));
2426 generate_subtree_copies (top_racc
->first_child
, lhs
, top_racc
->offset
,
2427 0, 0, gsi
, false, false,
2428 gimple_location (gsi_stmt (*gsi
)));
2429 return SRA_UDH_LEFT
;
2434 /* Try to generate statements to load all sub-replacements in an access subtree
2435 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2436 If that is not possible, refresh the TOP_RACC base aggregate and load the
2437 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2438 copied. NEW_GSI is stmt iterator used for statement insertions after the
2439 original assignment, OLD_GSI is used to insert statements before the
2440 assignment. *REFRESHED keeps the information whether we have needed to
2441 refresh replacements of the LHS and from which side of the assignments this
2445 load_assign_lhs_subreplacements (struct access
*lacc
, struct access
*top_racc
,
2446 HOST_WIDE_INT left_offset
,
2447 gimple_stmt_iterator
*old_gsi
,
2448 gimple_stmt_iterator
*new_gsi
,
2449 enum unscalarized_data_handling
*refreshed
)
2451 location_t loc
= gimple_location (gsi_stmt (*old_gsi
));
2452 for (lacc
= lacc
->first_child
; lacc
; lacc
= lacc
->next_sibling
)
2454 if (lacc
->grp_to_be_replaced
)
2456 struct access
*racc
;
2457 HOST_WIDE_INT offset
= lacc
->offset
- left_offset
+ top_racc
->offset
;
2461 racc
= find_access_in_subtree (top_racc
, offset
, lacc
->size
);
2462 if (racc
&& racc
->grp_to_be_replaced
)
2464 rhs
= get_access_replacement (racc
);
2465 if (!useless_type_conversion_p (lacc
->type
, racc
->type
))
2466 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
, lacc
->type
, rhs
);
2470 /* No suitable access on the right hand side, need to load from
2471 the aggregate. See if we have to update it first... */
2472 if (*refreshed
== SRA_UDH_NONE
)
2473 *refreshed
= handle_unscalarized_data_in_subtree (top_racc
,
2476 if (*refreshed
== SRA_UDH_LEFT
)
2477 rhs
= build_ref_for_model (loc
, lacc
->base
, lacc
->offset
, lacc
,
2480 rhs
= build_ref_for_model (loc
, top_racc
->base
, offset
, lacc
,
2484 stmt
= gimple_build_assign (get_access_replacement (lacc
), rhs
);
2485 gsi_insert_after (new_gsi
, stmt
, GSI_NEW_STMT
);
2486 gimple_set_location (stmt
, loc
);
2488 sra_stats
.subreplacements
++;
2490 else if (*refreshed
== SRA_UDH_NONE
2491 && lacc
->grp_read
&& !lacc
->grp_covered
)
2492 *refreshed
= handle_unscalarized_data_in_subtree (top_racc
,
2495 if (lacc
->first_child
)
2496 load_assign_lhs_subreplacements (lacc
, top_racc
, left_offset
,
2497 old_gsi
, new_gsi
, refreshed
);
2501 /* Result code for SRA assignment modification. */
2502 enum assignment_mod_result
{ SRA_AM_NONE
, /* nothing done for the stmt */
2503 SRA_AM_MODIFIED
, /* stmt changed but not
2505 SRA_AM_REMOVED
}; /* stmt eliminated */
2507 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2508 to the assignment and GSI is the statement iterator pointing at it. Returns
2509 the same values as sra_modify_assign. */
2511 static enum assignment_mod_result
2512 sra_modify_constructor_assign (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
2514 tree lhs
= gimple_assign_lhs (*stmt
);
2518 acc
= get_access_for_expr (lhs
);
2522 loc
= gimple_location (*stmt
);
2523 if (VEC_length (constructor_elt
,
2524 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt
))) > 0)
2526 /* I have never seen this code path trigger but if it can happen the
2527 following should handle it gracefully. */
2528 if (access_has_children_p (acc
))
2529 generate_subtree_copies (acc
->first_child
, acc
->base
, 0, 0, 0, gsi
,
2531 return SRA_AM_MODIFIED
;
2534 if (acc
->grp_covered
)
2536 init_subtree_with_zero (acc
, gsi
, false, loc
);
2537 unlink_stmt_vdef (*stmt
);
2538 gsi_remove (gsi
, true);
2539 return SRA_AM_REMOVED
;
2543 init_subtree_with_zero (acc
, gsi
, true, loc
);
2544 return SRA_AM_MODIFIED
;
2548 /* Create and return a new suitable default definition SSA_NAME for RACC which
2549 is an access describing an uninitialized part of an aggregate that is being
2553 get_repl_default_def_ssa_name (struct access
*racc
)
2557 decl
= get_unrenamed_access_replacement (racc
);
2559 repl
= gimple_default_def (cfun
, decl
);
2562 repl
= make_ssa_name (decl
, gimple_build_nop ());
2563 set_default_def (decl
, repl
);
2569 /* Examine both sides of the assignment statement pointed to by STMT, replace
2570 them with a scalare replacement if there is one and generate copying of
2571 replacements if scalarized aggregates have been used in the assignment. GSI
2572 is used to hold generated statements for type conversions and subtree
2575 static enum assignment_mod_result
2576 sra_modify_assign (gimple
*stmt
, gimple_stmt_iterator
*gsi
)
2578 struct access
*lacc
, *racc
;
2580 bool modify_this_stmt
= false;
2581 bool force_gimple_rhs
= false;
2583 gimple_stmt_iterator orig_gsi
= *gsi
;
2585 if (!gimple_assign_single_p (*stmt
))
2587 lhs
= gimple_assign_lhs (*stmt
);
2588 rhs
= gimple_assign_rhs1 (*stmt
);
2590 if (TREE_CODE (rhs
) == CONSTRUCTOR
)
2591 return sra_modify_constructor_assign (stmt
, gsi
);
2593 if (TREE_CODE (rhs
) == REALPART_EXPR
|| TREE_CODE (lhs
) == REALPART_EXPR
2594 || TREE_CODE (rhs
) == IMAGPART_EXPR
|| TREE_CODE (lhs
) == IMAGPART_EXPR
2595 || TREE_CODE (rhs
) == BIT_FIELD_REF
|| TREE_CODE (lhs
) == BIT_FIELD_REF
)
2597 modify_this_stmt
= sra_modify_expr (gimple_assign_rhs1_ptr (*stmt
),
2599 modify_this_stmt
|= sra_modify_expr (gimple_assign_lhs_ptr (*stmt
),
2601 return modify_this_stmt
? SRA_AM_MODIFIED
: SRA_AM_NONE
;
2604 lacc
= get_access_for_expr (lhs
);
2605 racc
= get_access_for_expr (rhs
);
2609 loc
= gimple_location (*stmt
);
2610 if (lacc
&& lacc
->grp_to_be_replaced
)
2612 lhs
= get_access_replacement (lacc
);
2613 gimple_assign_set_lhs (*stmt
, lhs
);
2614 modify_this_stmt
= true;
2615 if (lacc
->grp_partial_lhs
)
2616 force_gimple_rhs
= true;
2620 if (racc
&& racc
->grp_to_be_replaced
)
2622 rhs
= get_access_replacement (racc
);
2623 modify_this_stmt
= true;
2624 if (racc
->grp_partial_lhs
)
2625 force_gimple_rhs
= true;
2629 if (modify_this_stmt
)
2631 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
2633 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2634 ??? This should move to fold_stmt which we simply should
2635 call after building a VIEW_CONVERT_EXPR here. */
2636 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs
))
2637 && !access_has_children_p (lacc
))
2639 lhs
= build_ref_for_offset (loc
, lhs
, 0, TREE_TYPE (rhs
),
2641 gimple_assign_set_lhs (*stmt
, lhs
);
2643 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs
))
2644 && !contains_view_convert_expr_p (rhs
)
2645 && !access_has_children_p (racc
))
2646 rhs
= build_ref_for_offset (loc
, rhs
, 0, TREE_TYPE (lhs
),
2649 if (!useless_type_conversion_p (TREE_TYPE (lhs
), TREE_TYPE (rhs
)))
2651 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
, TREE_TYPE (lhs
),
2653 if (is_gimple_reg_type (TREE_TYPE (lhs
))
2654 && TREE_CODE (lhs
) != SSA_NAME
)
2655 force_gimple_rhs
= true;
2660 /* From this point on, the function deals with assignments in between
2661 aggregates when at least one has scalar reductions of some of its
2662 components. There are three possible scenarios: Both the LHS and RHS have
2663 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2665 In the first case, we would like to load the LHS components from RHS
2666 components whenever possible. If that is not possible, we would like to
2667 read it directly from the RHS (after updating it by storing in it its own
2668 components). If there are some necessary unscalarized data in the LHS,
2669 those will be loaded by the original assignment too. If neither of these
2670 cases happen, the original statement can be removed. Most of this is done
2671 by load_assign_lhs_subreplacements.
2673 In the second case, we would like to store all RHS scalarized components
2674 directly into LHS and if they cover the aggregate completely, remove the
2675 statement too. In the third case, we want the LHS components to be loaded
2676 directly from the RHS (DSE will remove the original statement if it
2679 This is a bit complex but manageable when types match and when unions do
2680 not cause confusion in a way that we cannot really load a component of LHS
2681 from the RHS or vice versa (the access representing this level can have
2682 subaccesses that are accessible only through a different union field at a
2683 higher level - different from the one used in the examined expression).
2686 Therefore, I specially handle a fourth case, happening when there is a
2687 specific type cast or it is impossible to locate a scalarized subaccess on
2688 the other side of the expression. If that happens, I simply "refresh" the
2689 RHS by storing in it is scalarized components leave the original statement
2690 there to do the copying and then load the scalar replacements of the LHS.
2691 This is what the first branch does. */
2693 if (gimple_has_volatile_ops (*stmt
)
2694 || contains_view_convert_expr_p (rhs
)
2695 || contains_view_convert_expr_p (lhs
))
2697 if (access_has_children_p (racc
))
2698 generate_subtree_copies (racc
->first_child
, racc
->base
, 0, 0, 0,
2699 gsi
, false, false, loc
);
2700 if (access_has_children_p (lacc
))
2701 generate_subtree_copies (lacc
->first_child
, lacc
->base
, 0, 0, 0,
2702 gsi
, true, true, loc
);
2703 sra_stats
.separate_lhs_rhs_handling
++;
2707 if (access_has_children_p (lacc
) && access_has_children_p (racc
))
2709 gimple_stmt_iterator orig_gsi
= *gsi
;
2710 enum unscalarized_data_handling refreshed
;
2712 if (lacc
->grp_read
&& !lacc
->grp_covered
)
2713 refreshed
= handle_unscalarized_data_in_subtree (racc
, gsi
);
2715 refreshed
= SRA_UDH_NONE
;
2717 load_assign_lhs_subreplacements (lacc
, racc
, lacc
->offset
,
2718 &orig_gsi
, gsi
, &refreshed
);
2719 if (refreshed
!= SRA_UDH_RIGHT
)
2722 unlink_stmt_vdef (*stmt
);
2723 gsi_remove (&orig_gsi
, true);
2724 sra_stats
.deleted
++;
2725 return SRA_AM_REMOVED
;
2732 if (!racc
->grp_to_be_replaced
&& !racc
->grp_unscalarized_data
)
2736 fprintf (dump_file
, "Removing load: ");
2737 print_gimple_stmt (dump_file
, *stmt
, 0, 0);
2740 if (TREE_CODE (lhs
) == SSA_NAME
)
2742 rhs
= get_repl_default_def_ssa_name (racc
);
2743 if (!useless_type_conversion_p (TREE_TYPE (lhs
),
2745 rhs
= fold_build1_loc (loc
, VIEW_CONVERT_EXPR
,
2746 TREE_TYPE (lhs
), rhs
);
2750 if (racc
->first_child
)
2751 generate_subtree_copies (racc
->first_child
, lhs
,
2752 racc
->offset
, 0, 0, gsi
,
2755 gcc_assert (*stmt
== gsi_stmt (*gsi
));
2756 unlink_stmt_vdef (*stmt
);
2757 gsi_remove (gsi
, true);
2758 sra_stats
.deleted
++;
2759 return SRA_AM_REMOVED
;
2762 else if (racc
->first_child
)
2763 generate_subtree_copies (racc
->first_child
, lhs
, racc
->offset
,
2764 0, 0, gsi
, false, true, loc
);
2766 if (access_has_children_p (lacc
))
2767 generate_subtree_copies (lacc
->first_child
, rhs
, lacc
->offset
,
2768 0, 0, gsi
, true, true, loc
);
2772 /* This gimplification must be done after generate_subtree_copies, lest we
2773 insert the subtree copies in the middle of the gimplified sequence. */
2774 if (force_gimple_rhs
)
2775 rhs
= force_gimple_operand_gsi (&orig_gsi
, rhs
, true, NULL_TREE
,
2776 true, GSI_SAME_STMT
);
2777 if (gimple_assign_rhs1 (*stmt
) != rhs
)
2779 modify_this_stmt
= true;
2780 gimple_assign_set_rhs_from_tree (&orig_gsi
, rhs
);
2781 gcc_assert (*stmt
== gsi_stmt (orig_gsi
));
2784 return modify_this_stmt
? SRA_AM_MODIFIED
: SRA_AM_NONE
;
2787 /* Traverse the function body and all modifications as decided in
2788 analyze_all_variable_accesses. Return true iff the CFG has been
2792 sra_modify_function_body (void)
2794 bool cfg_changed
= false;
2799 gimple_stmt_iterator gsi
= gsi_start_bb (bb
);
2800 while (!gsi_end_p (gsi
))
2802 gimple stmt
= gsi_stmt (gsi
);
2803 enum assignment_mod_result assign_result
;
2804 bool modified
= false, deleted
= false;
2808 switch (gimple_code (stmt
))
2811 t
= gimple_return_retval_ptr (stmt
);
2812 if (*t
!= NULL_TREE
)
2813 modified
|= sra_modify_expr (t
, &gsi
, false);
2817 assign_result
= sra_modify_assign (&stmt
, &gsi
);
2818 modified
|= assign_result
== SRA_AM_MODIFIED
;
2819 deleted
= assign_result
== SRA_AM_REMOVED
;
2823 /* Operands must be processed before the lhs. */
2824 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
2826 t
= gimple_call_arg_ptr (stmt
, i
);
2827 modified
|= sra_modify_expr (t
, &gsi
, false);
2830 if (gimple_call_lhs (stmt
))
2832 t
= gimple_call_lhs_ptr (stmt
);
2833 modified
|= sra_modify_expr (t
, &gsi
, true);
2838 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
2840 t
= &TREE_VALUE (gimple_asm_input_op (stmt
, i
));
2841 modified
|= sra_modify_expr (t
, &gsi
, false);
2843 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
2845 t
= &TREE_VALUE (gimple_asm_output_op (stmt
, i
));
2846 modified
|= sra_modify_expr (t
, &gsi
, true);
2857 if (maybe_clean_eh_stmt (stmt
)
2858 && gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
2869 /* Generate statements initializing scalar replacements of parts of function
2873 initialize_parameter_reductions (void)
2875 gimple_stmt_iterator gsi
;
2876 gimple_seq seq
= NULL
;
2879 for (parm
= DECL_ARGUMENTS (current_function_decl
);
2881 parm
= DECL_CHAIN (parm
))
2883 VEC (access_p
, heap
) *access_vec
;
2884 struct access
*access
;
2886 if (!bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
2888 access_vec
= get_base_access_vector (parm
);
2894 seq
= gimple_seq_alloc ();
2895 gsi
= gsi_start (seq
);
2898 for (access
= VEC_index (access_p
, access_vec
, 0);
2900 access
= access
->next_grp
)
2901 generate_subtree_copies (access
, parm
, 0, 0, 0, &gsi
, true, true,
2902 EXPR_LOCATION (parm
));
2906 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR
), seq
);
2909 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
2910 it reveals there are components of some aggregates to be scalarized, it runs
2911 the required transformations. */
2913 perform_intra_sra (void)
2918 if (!find_var_candidates ())
2921 if (!scan_function ())
2924 if (!analyze_all_variable_accesses ())
2927 if (sra_modify_function_body ())
2928 ret
= TODO_update_ssa
| TODO_cleanup_cfg
;
2930 ret
= TODO_update_ssa
;
2931 initialize_parameter_reductions ();
2933 statistics_counter_event (cfun
, "Scalar replacements created",
2934 sra_stats
.replacements
);
2935 statistics_counter_event (cfun
, "Modified expressions", sra_stats
.exprs
);
2936 statistics_counter_event (cfun
, "Subtree copy stmts",
2937 sra_stats
.subtree_copies
);
2938 statistics_counter_event (cfun
, "Subreplacement stmts",
2939 sra_stats
.subreplacements
);
2940 statistics_counter_event (cfun
, "Deleted stmts", sra_stats
.deleted
);
2941 statistics_counter_event (cfun
, "Separate LHS and RHS handling",
2942 sra_stats
.separate_lhs_rhs_handling
);
2945 sra_deinitialize ();
2949 /* Perform early intraprocedural SRA. */
2951 early_intra_sra (void)
2953 sra_mode
= SRA_MODE_EARLY_INTRA
;
2954 return perform_intra_sra ();
2957 /* Perform "late" intraprocedural SRA. */
2959 late_intra_sra (void)
2961 sra_mode
= SRA_MODE_INTRA
;
2962 return perform_intra_sra ();
2967 gate_intra_sra (void)
2969 return flag_tree_sra
!= 0 && dbg_cnt (tree_sra
);
2973 struct gimple_opt_pass pass_sra_early
=
2978 gate_intra_sra
, /* gate */
2979 early_intra_sra
, /* execute */
2982 0, /* static_pass_number */
2983 TV_TREE_SRA
, /* tv_id */
2984 PROP_cfg
| PROP_ssa
, /* properties_required */
2985 0, /* properties_provided */
2986 0, /* properties_destroyed */
2987 0, /* todo_flags_start */
2991 | TODO_verify_ssa
/* todo_flags_finish */
2995 struct gimple_opt_pass pass_sra
=
3000 gate_intra_sra
, /* gate */
3001 late_intra_sra
, /* execute */
3004 0, /* static_pass_number */
3005 TV_TREE_SRA
, /* tv_id */
3006 PROP_cfg
| PROP_ssa
, /* properties_required */
3007 0, /* properties_provided */
3008 0, /* properties_destroyed */
3009 TODO_update_address_taken
, /* todo_flags_start */
3013 | TODO_verify_ssa
/* todo_flags_finish */
3018 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3022 is_unused_scalar_param (tree parm
)
3025 return (is_gimple_reg (parm
)
3026 && (!(name
= gimple_default_def (cfun
, parm
))
3027 || has_zero_uses (name
)));
3030 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3031 examine whether there are any direct or otherwise infeasible ones. If so,
3032 return true, otherwise return false. PARM must be a gimple register with a
3033 non-NULL default definition. */
3036 ptr_parm_has_direct_uses (tree parm
)
3038 imm_use_iterator ui
;
3040 tree name
= gimple_default_def (cfun
, parm
);
3043 FOR_EACH_IMM_USE_STMT (stmt
, ui
, name
)
3046 use_operand_p use_p
;
3048 if (is_gimple_debug (stmt
))
3051 /* Valid uses include dereferences on the lhs and the rhs. */
3052 if (gimple_has_lhs (stmt
))
3054 tree lhs
= gimple_get_lhs (stmt
);
3055 while (handled_component_p (lhs
))
3056 lhs
= TREE_OPERAND (lhs
, 0);
3057 if (TREE_CODE (lhs
) == MEM_REF
3058 && TREE_OPERAND (lhs
, 0) == name
3059 && integer_zerop (TREE_OPERAND (lhs
, 1))
3060 && types_compatible_p (TREE_TYPE (lhs
),
3061 TREE_TYPE (TREE_TYPE (name
))))
3064 if (gimple_assign_single_p (stmt
))
3066 tree rhs
= gimple_assign_rhs1 (stmt
);
3067 while (handled_component_p (rhs
))
3068 rhs
= TREE_OPERAND (rhs
, 0);
3069 if (TREE_CODE (rhs
) == MEM_REF
3070 && TREE_OPERAND (rhs
, 0) == name
3071 && integer_zerop (TREE_OPERAND (rhs
, 1))
3072 && types_compatible_p (TREE_TYPE (rhs
),
3073 TREE_TYPE (TREE_TYPE (name
))))
3076 else if (is_gimple_call (stmt
))
3079 for (i
= 0; i
< gimple_call_num_args (stmt
); ++i
)
3081 tree arg
= gimple_call_arg (stmt
, i
);
3082 while (handled_component_p (arg
))
3083 arg
= TREE_OPERAND (arg
, 0);
3084 if (TREE_CODE (arg
) == MEM_REF
3085 && TREE_OPERAND (arg
, 0) == name
3086 && integer_zerop (TREE_OPERAND (arg
, 1))
3087 && types_compatible_p (TREE_TYPE (arg
),
3088 TREE_TYPE (TREE_TYPE (name
))))
3093 /* If the number of valid uses does not match the number of
3094 uses in this stmt there is an unhandled use. */
3095 FOR_EACH_IMM_USE_ON_STMT (use_p
, ui
)
3102 BREAK_FROM_IMM_USE_STMT (ui
);
3108 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3109 them in candidate_bitmap. Note that these do not necessarily include
3110 parameter which are unused and thus can be removed. Return true iff any
3111 such candidate has been found. */
3114 find_param_candidates (void)
3120 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3122 parm
= DECL_CHAIN (parm
))
3124 tree type
= TREE_TYPE (parm
);
3128 if (TREE_THIS_VOLATILE (parm
)
3129 || TREE_ADDRESSABLE (parm
)
3130 || (!is_gimple_reg_type (type
) && is_va_list_type (type
)))
3133 if (is_unused_scalar_param (parm
))
3139 if (POINTER_TYPE_P (type
))
3141 type
= TREE_TYPE (type
);
3143 if (TREE_CODE (type
) == FUNCTION_TYPE
3144 || TYPE_VOLATILE (type
)
3145 || (TREE_CODE (type
) == ARRAY_TYPE
3146 && TYPE_NONALIASED_COMPONENT (type
))
3147 || !is_gimple_reg (parm
)
3148 || is_va_list_type (type
)
3149 || ptr_parm_has_direct_uses (parm
))
3152 else if (!AGGREGATE_TYPE_P (type
))
3155 if (!COMPLETE_TYPE_P (type
)
3156 || !host_integerp (TYPE_SIZE (type
), 1)
3157 || tree_low_cst (TYPE_SIZE (type
), 1) == 0
3158 || (AGGREGATE_TYPE_P (type
)
3159 && type_internals_preclude_sra_p (type
)))
3162 bitmap_set_bit (candidate_bitmap
, DECL_UID (parm
));
3164 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3166 fprintf (dump_file
, "Candidate (%d): ", DECL_UID (parm
));
3167 print_generic_expr (dump_file
, parm
, 0);
3168 fprintf (dump_file
, "\n");
3172 func_param_count
= count
;
3176 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3180 mark_maybe_modified (ao_ref
*ao ATTRIBUTE_UNUSED
, tree vdef ATTRIBUTE_UNUSED
,
3183 struct access
*repr
= (struct access
*) data
;
3185 repr
->grp_maybe_modified
= 1;
3189 /* Analyze what representatives (in linked lists accessible from
3190 REPRESENTATIVES) can be modified by side effects of statements in the
3191 current function. */
3194 analyze_modified_params (VEC (access_p
, heap
) *representatives
)
3198 for (i
= 0; i
< func_param_count
; i
++)
3200 struct access
*repr
;
3202 for (repr
= VEC_index (access_p
, representatives
, i
);
3204 repr
= repr
->next_grp
)
3206 struct access
*access
;
3210 if (no_accesses_p (repr
))
3212 if (!POINTER_TYPE_P (TREE_TYPE (repr
->base
))
3213 || repr
->grp_maybe_modified
)
3216 ao_ref_init (&ar
, repr
->expr
);
3217 visited
= BITMAP_ALLOC (NULL
);
3218 for (access
= repr
; access
; access
= access
->next_sibling
)
3220 /* All accesses are read ones, otherwise grp_maybe_modified would
3221 be trivially set. */
3222 walk_aliased_vdefs (&ar
, gimple_vuse (access
->stmt
),
3223 mark_maybe_modified
, repr
, &visited
);
3224 if (repr
->grp_maybe_modified
)
3227 BITMAP_FREE (visited
);
3232 /* Propagate distances in bb_dereferences in the opposite direction than the
3233 control flow edges, in each step storing the maximum of the current value
3234 and the minimum of all successors. These steps are repeated until the table
3235 stabilizes. Note that BBs which might terminate the functions (according to
3236 final_bbs bitmap) never updated in this way. */
3239 propagate_dereference_distances (void)
3241 VEC (basic_block
, heap
) *queue
;
3244 queue
= VEC_alloc (basic_block
, heap
, last_basic_block_for_function (cfun
));
3245 VEC_quick_push (basic_block
, queue
, ENTRY_BLOCK_PTR
);
3248 VEC_quick_push (basic_block
, queue
, bb
);
3252 while (!VEC_empty (basic_block
, queue
))
3256 bool change
= false;
3259 bb
= VEC_pop (basic_block
, queue
);
3262 if (bitmap_bit_p (final_bbs
, bb
->index
))
3265 for (i
= 0; i
< func_param_count
; i
++)
3267 int idx
= bb
->index
* func_param_count
+ i
;
3269 HOST_WIDE_INT inh
= 0;
3271 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
3273 int succ_idx
= e
->dest
->index
* func_param_count
+ i
;
3275 if (e
->src
== EXIT_BLOCK_PTR
)
3281 inh
= bb_dereferences
[succ_idx
];
3283 else if (bb_dereferences
[succ_idx
] < inh
)
3284 inh
= bb_dereferences
[succ_idx
];
3287 if (!first
&& bb_dereferences
[idx
] < inh
)
3289 bb_dereferences
[idx
] = inh
;
3294 if (change
&& !bitmap_bit_p (final_bbs
, bb
->index
))
3295 FOR_EACH_EDGE (e
, ei
, bb
->preds
)
3300 e
->src
->aux
= e
->src
;
3301 VEC_quick_push (basic_block
, queue
, e
->src
);
3305 VEC_free (basic_block
, heap
, queue
);
3308 /* Dump a dereferences TABLE with heading STR to file F. */
3311 dump_dereferences_table (FILE *f
, const char *str
, HOST_WIDE_INT
*table
)
3315 fprintf (dump_file
, str
);
3316 FOR_BB_BETWEEN (bb
, ENTRY_BLOCK_PTR
, EXIT_BLOCK_PTR
, next_bb
)
3318 fprintf (f
, "%4i %i ", bb
->index
, bitmap_bit_p (final_bbs
, bb
->index
));
3319 if (bb
!= EXIT_BLOCK_PTR
)
3322 for (i
= 0; i
< func_param_count
; i
++)
3324 int idx
= bb
->index
* func_param_count
+ i
;
3325 fprintf (f
, " %4" HOST_WIDE_INT_PRINT
"d", table
[idx
]);
3330 fprintf (dump_file
, "\n");
3333 /* Determine what (parts of) parameters passed by reference that are not
3334 assigned to are not certainly dereferenced in this function and thus the
3335 dereferencing cannot be safely moved to the caller without potentially
3336 introducing a segfault. Mark such REPRESENTATIVES as
3337 grp_not_necessarilly_dereferenced.
3339 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3340 part is calculated rather than simple booleans are calculated for each
3341 pointer parameter to handle cases when only a fraction of the whole
3342 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3345 The maximum dereference distances for each pointer parameter and BB are
3346 already stored in bb_dereference. This routine simply propagates these
3347 values upwards by propagate_dereference_distances and then compares the
3348 distances of individual parameters in the ENTRY BB to the equivalent
3349 distances of each representative of a (fraction of a) parameter. */
3352 analyze_caller_dereference_legality (VEC (access_p
, heap
) *representatives
)
3356 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3357 dump_dereferences_table (dump_file
,
3358 "Dereference table before propagation:\n",
3361 propagate_dereference_distances ();
3363 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3364 dump_dereferences_table (dump_file
,
3365 "Dereference table after propagation:\n",
3368 for (i
= 0; i
< func_param_count
; i
++)
3370 struct access
*repr
= VEC_index (access_p
, representatives
, i
);
3371 int idx
= ENTRY_BLOCK_PTR
->index
* func_param_count
+ i
;
3373 if (!repr
|| no_accesses_p (repr
))
3378 if ((repr
->offset
+ repr
->size
) > bb_dereferences
[idx
])
3379 repr
->grp_not_necessarilly_dereferenced
= 1;
3380 repr
= repr
->next_grp
;
3386 /* Return the representative access for the parameter declaration PARM if it is
3387 a scalar passed by reference which is not written to and the pointer value
3388 is not used directly. Thus, if it is legal to dereference it in the caller
3389 and we can rule out modifications through aliases, such parameter should be
3390 turned into one passed by value. Return NULL otherwise. */
3392 static struct access
*
3393 unmodified_by_ref_scalar_representative (tree parm
)
3395 int i
, access_count
;
3396 struct access
*repr
;
3397 VEC (access_p
, heap
) *access_vec
;
3399 access_vec
= get_base_access_vector (parm
);
3400 gcc_assert (access_vec
);
3401 repr
= VEC_index (access_p
, access_vec
, 0);
3404 repr
->group_representative
= repr
;
3406 access_count
= VEC_length (access_p
, access_vec
);
3407 for (i
= 1; i
< access_count
; i
++)
3409 struct access
*access
= VEC_index (access_p
, access_vec
, i
);
3412 access
->group_representative
= repr
;
3413 access
->next_sibling
= repr
->next_sibling
;
3414 repr
->next_sibling
= access
;
3418 repr
->grp_scalar_ptr
= 1;
3422 /* Return true iff this access precludes IPA-SRA of the parameter it is
3426 access_precludes_ipa_sra_p (struct access
*access
)
3428 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3429 is incompatible assign in a call statement (and possibly even in asm
3430 statements). This can be relaxed by using a new temporary but only for
3431 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3432 intraprocedural SRA we deal with this by keeping the old aggregate around,
3433 something we cannot do in IPA-SRA.) */
3435 && (is_gimple_call (access
->stmt
)
3436 || gimple_code (access
->stmt
) == GIMPLE_ASM
))
3443 /* Sort collected accesses for parameter PARM, identify representatives for
3444 each accessed region and link them together. Return NULL if there are
3445 different but overlapping accesses, return the special ptr value meaning
3446 there are no accesses for this parameter if that is the case and return the
3447 first representative otherwise. Set *RO_GRP if there is a group of accesses
3448 with only read (i.e. no write) accesses. */
3450 static struct access
*
3451 splice_param_accesses (tree parm
, bool *ro_grp
)
3453 int i
, j
, access_count
, group_count
;
3454 int agg_size
, total_size
= 0;
3455 struct access
*access
, *res
, **prev_acc_ptr
= &res
;
3456 VEC (access_p
, heap
) *access_vec
;
3458 access_vec
= get_base_access_vector (parm
);
3460 return &no_accesses_representant
;
3461 access_count
= VEC_length (access_p
, access_vec
);
3463 qsort (VEC_address (access_p
, access_vec
), access_count
, sizeof (access_p
),
3464 compare_access_positions
);
3469 while (i
< access_count
)
3472 access
= VEC_index (access_p
, access_vec
, i
);
3473 modification
= access
->write
;
3474 if (access_precludes_ipa_sra_p (access
))
3477 /* Access is about to become group representative unless we find some
3478 nasty overlap which would preclude us from breaking this parameter
3482 while (j
< access_count
)
3484 struct access
*ac2
= VEC_index (access_p
, access_vec
, j
);
3485 if (ac2
->offset
!= access
->offset
)
3487 /* All or nothing law for parameters. */
3488 if (access
->offset
+ access
->size
> ac2
->offset
)
3493 else if (ac2
->size
!= access
->size
)
3496 if (access_precludes_ipa_sra_p (ac2
))
3499 modification
|= ac2
->write
;
3500 ac2
->group_representative
= access
;
3501 ac2
->next_sibling
= access
->next_sibling
;
3502 access
->next_sibling
= ac2
;
3507 access
->grp_maybe_modified
= modification
;
3510 *prev_acc_ptr
= access
;
3511 prev_acc_ptr
= &access
->next_grp
;
3512 total_size
+= access
->size
;
3516 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
3517 agg_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm
))), 1);
3519 agg_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (parm
)), 1);
3520 if (total_size
>= agg_size
)
3523 gcc_assert (group_count
> 0);
3527 /* Decide whether parameters with representative accesses given by REPR should
3528 be reduced into components. */
3531 decide_one_param_reduction (struct access
*repr
)
3533 int total_size
, cur_parm_size
, agg_size
, new_param_count
, parm_size_limit
;
3538 cur_parm_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (parm
)), 1);
3539 gcc_assert (cur_parm_size
> 0);
3541 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
3544 agg_size
= tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm
))), 1);
3549 agg_size
= cur_parm_size
;
3555 fprintf (dump_file
, "Evaluating PARAM group sizes for ");
3556 print_generic_expr (dump_file
, parm
, 0);
3557 fprintf (dump_file
, " (UID: %u): \n", DECL_UID (parm
));
3558 for (acc
= repr
; acc
; acc
= acc
->next_grp
)
3559 dump_access (dump_file
, acc
, true);
3563 new_param_count
= 0;
3565 for (; repr
; repr
= repr
->next_grp
)
3567 gcc_assert (parm
== repr
->base
);
3570 if (!by_ref
|| (!repr
->grp_maybe_modified
3571 && !repr
->grp_not_necessarilly_dereferenced
))
3572 total_size
+= repr
->size
;
3574 total_size
+= cur_parm_size
;
3577 gcc_assert (new_param_count
> 0);
3579 if (optimize_function_for_size_p (cfun
))
3580 parm_size_limit
= cur_parm_size
;
3582 parm_size_limit
= (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR
)
3585 if (total_size
< agg_size
3586 && total_size
<= parm_size_limit
)
3589 fprintf (dump_file
, " ....will be split into %i components\n",
3591 return new_param_count
;
3597 /* The order of the following enums is important, we need to do extra work for
3598 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3599 enum ipa_splicing_result
{ NO_GOOD_ACCESS
, UNUSED_PARAMS
, BY_VAL_ACCESSES
,
3600 MODIF_BY_REF_ACCESSES
, UNMODIF_BY_REF_ACCESSES
};
3602 /* Identify representatives of all accesses to all candidate parameters for
3603 IPA-SRA. Return result based on what representatives have been found. */
3605 static enum ipa_splicing_result
3606 splice_all_param_accesses (VEC (access_p
, heap
) **representatives
)
3608 enum ipa_splicing_result result
= NO_GOOD_ACCESS
;
3610 struct access
*repr
;
3612 *representatives
= VEC_alloc (access_p
, heap
, func_param_count
);
3614 for (parm
= DECL_ARGUMENTS (current_function_decl
);
3616 parm
= DECL_CHAIN (parm
))
3618 if (is_unused_scalar_param (parm
))
3620 VEC_quick_push (access_p
, *representatives
,
3621 &no_accesses_representant
);
3622 if (result
== NO_GOOD_ACCESS
)
3623 result
= UNUSED_PARAMS
;
3625 else if (POINTER_TYPE_P (TREE_TYPE (parm
))
3626 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm
)))
3627 && bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
3629 repr
= unmodified_by_ref_scalar_representative (parm
);
3630 VEC_quick_push (access_p
, *representatives
, repr
);
3632 result
= UNMODIF_BY_REF_ACCESSES
;
3634 else if (bitmap_bit_p (candidate_bitmap
, DECL_UID (parm
)))
3636 bool ro_grp
= false;
3637 repr
= splice_param_accesses (parm
, &ro_grp
);
3638 VEC_quick_push (access_p
, *representatives
, repr
);
3640 if (repr
&& !no_accesses_p (repr
))
3642 if (POINTER_TYPE_P (TREE_TYPE (parm
)))
3645 result
= UNMODIF_BY_REF_ACCESSES
;
3646 else if (result
< MODIF_BY_REF_ACCESSES
)
3647 result
= MODIF_BY_REF_ACCESSES
;
3649 else if (result
< BY_VAL_ACCESSES
)
3650 result
= BY_VAL_ACCESSES
;
3652 else if (no_accesses_p (repr
) && (result
== NO_GOOD_ACCESS
))
3653 result
= UNUSED_PARAMS
;
3656 VEC_quick_push (access_p
, *representatives
, NULL
);
3659 if (result
== NO_GOOD_ACCESS
)
3661 VEC_free (access_p
, heap
, *representatives
);
3662 *representatives
= NULL
;
3663 return NO_GOOD_ACCESS
;
3669 /* Return the index of BASE in PARMS. Abort if it is not found. */
3672 get_param_index (tree base
, VEC(tree
, heap
) *parms
)
3676 len
= VEC_length (tree
, parms
);
3677 for (i
= 0; i
< len
; i
++)
3678 if (VEC_index (tree
, parms
, i
) == base
)
3683 /* Convert the decisions made at the representative level into compact
3684 parameter adjustments. REPRESENTATIVES are pointers to first
3685 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3686 final number of adjustments. */
3688 static ipa_parm_adjustment_vec
3689 turn_representatives_into_adjustments (VEC (access_p
, heap
) *representatives
,
3690 int adjustments_count
)
3692 VEC (tree
, heap
) *parms
;
3693 ipa_parm_adjustment_vec adjustments
;
3697 gcc_assert (adjustments_count
> 0);
3698 parms
= ipa_get_vector_of_formal_parms (current_function_decl
);
3699 adjustments
= VEC_alloc (ipa_parm_adjustment_t
, heap
, adjustments_count
);
3700 parm
= DECL_ARGUMENTS (current_function_decl
);
3701 for (i
= 0; i
< func_param_count
; i
++, parm
= DECL_CHAIN (parm
))
3703 struct access
*repr
= VEC_index (access_p
, representatives
, i
);
3705 if (!repr
|| no_accesses_p (repr
))
3707 struct ipa_parm_adjustment
*adj
;
3709 adj
= VEC_quick_push (ipa_parm_adjustment_t
, adjustments
, NULL
);
3710 memset (adj
, 0, sizeof (*adj
));
3711 adj
->base_index
= get_param_index (parm
, parms
);
3714 adj
->copy_param
= 1;
3716 adj
->remove_param
= 1;
3720 struct ipa_parm_adjustment
*adj
;
3721 int index
= get_param_index (parm
, parms
);
3723 for (; repr
; repr
= repr
->next_grp
)
3725 adj
= VEC_quick_push (ipa_parm_adjustment_t
, adjustments
, NULL
);
3726 memset (adj
, 0, sizeof (*adj
));
3727 gcc_assert (repr
->base
== parm
);
3728 adj
->base_index
= index
;
3729 adj
->base
= repr
->base
;
3730 adj
->type
= repr
->type
;
3731 adj
->offset
= repr
->offset
;
3732 adj
->by_ref
= (POINTER_TYPE_P (TREE_TYPE (repr
->base
))
3733 && (repr
->grp_maybe_modified
3734 || repr
->grp_not_necessarilly_dereferenced
));
3739 VEC_free (tree
, heap
, parms
);
3743 /* Analyze the collected accesses and produce a plan what to do with the
3744 parameters in the form of adjustments, NULL meaning nothing. */
3746 static ipa_parm_adjustment_vec
3747 analyze_all_param_acesses (void)
3749 enum ipa_splicing_result repr_state
;
3750 bool proceed
= false;
3751 int i
, adjustments_count
= 0;
3752 VEC (access_p
, heap
) *representatives
;
3753 ipa_parm_adjustment_vec adjustments
;
3755 repr_state
= splice_all_param_accesses (&representatives
);
3756 if (repr_state
== NO_GOOD_ACCESS
)
3759 /* If there are any parameters passed by reference which are not modified
3760 directly, we need to check whether they can be modified indirectly. */
3761 if (repr_state
== UNMODIF_BY_REF_ACCESSES
)
3763 analyze_caller_dereference_legality (representatives
);
3764 analyze_modified_params (representatives
);
3767 for (i
= 0; i
< func_param_count
; i
++)
3769 struct access
*repr
= VEC_index (access_p
, representatives
, i
);
3771 if (repr
&& !no_accesses_p (repr
))
3773 if (repr
->grp_scalar_ptr
)
3775 adjustments_count
++;
3776 if (repr
->grp_not_necessarilly_dereferenced
3777 || repr
->grp_maybe_modified
)
3778 VEC_replace (access_p
, representatives
, i
, NULL
);
3782 sra_stats
.scalar_by_ref_to_by_val
++;
3787 int new_components
= decide_one_param_reduction (repr
);
3789 if (new_components
== 0)
3791 VEC_replace (access_p
, representatives
, i
, NULL
);
3792 adjustments_count
++;
3796 adjustments_count
+= new_components
;
3797 sra_stats
.aggregate_params_reduced
++;
3798 sra_stats
.param_reductions_created
+= new_components
;
3805 if (no_accesses_p (repr
))
3808 sra_stats
.deleted_unused_parameters
++;
3810 adjustments_count
++;
3814 if (!proceed
&& dump_file
)
3815 fprintf (dump_file
, "NOT proceeding to change params.\n");
3818 adjustments
= turn_representatives_into_adjustments (representatives
,
3823 VEC_free (access_p
, heap
, representatives
);
3827 /* If a parameter replacement identified by ADJ does not yet exist in the form
3828 of declaration, create it and record it, otherwise return the previously
3832 get_replaced_param_substitute (struct ipa_parm_adjustment
*adj
)
3835 if (!adj
->new_ssa_base
)
3837 char *pretty_name
= make_fancy_name (adj
->base
);
3839 repl
= create_tmp_reg (TREE_TYPE (adj
->base
), "ISR");
3840 DECL_NAME (repl
) = get_identifier (pretty_name
);
3841 obstack_free (&name_obstack
, pretty_name
);
3844 add_referenced_var (repl
);
3845 adj
->new_ssa_base
= repl
;
3848 repl
= adj
->new_ssa_base
;
3852 /* Find the first adjustment for a particular parameter BASE in a vector of
3853 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3856 static struct ipa_parm_adjustment
*
3857 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments
, tree base
)
3861 len
= VEC_length (ipa_parm_adjustment_t
, adjustments
);
3862 for (i
= 0; i
< len
; i
++)
3864 struct ipa_parm_adjustment
*adj
;
3866 adj
= VEC_index (ipa_parm_adjustment_t
, adjustments
, i
);
3867 if (!adj
->copy_param
&& adj
->base
== base
)
3874 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3875 removed because its value is not used, replace the SSA_NAME with a one
3876 relating to a created VAR_DECL together all of its uses and return true.
3877 ADJUSTMENTS is a pointer to an adjustments vector. */
3880 replace_removed_params_ssa_names (gimple stmt
,
3881 ipa_parm_adjustment_vec adjustments
)
3883 struct ipa_parm_adjustment
*adj
;
3884 tree lhs
, decl
, repl
, name
;
3886 if (gimple_code (stmt
) == GIMPLE_PHI
)
3887 lhs
= gimple_phi_result (stmt
);
3888 else if (is_gimple_assign (stmt
))
3889 lhs
= gimple_assign_lhs (stmt
);
3890 else if (is_gimple_call (stmt
))
3891 lhs
= gimple_call_lhs (stmt
);
3895 if (TREE_CODE (lhs
) != SSA_NAME
)
3897 decl
= SSA_NAME_VAR (lhs
);
3898 if (TREE_CODE (decl
) != PARM_DECL
)
3901 adj
= get_adjustment_for_base (adjustments
, decl
);
3905 repl
= get_replaced_param_substitute (adj
);
3906 name
= make_ssa_name (repl
, stmt
);
3910 fprintf (dump_file
, "replacing an SSA name of a removed param ");
3911 print_generic_expr (dump_file
, lhs
, 0);
3912 fprintf (dump_file
, " with ");
3913 print_generic_expr (dump_file
, name
, 0);
3914 fprintf (dump_file
, "\n");
3917 if (is_gimple_assign (stmt
))
3918 gimple_assign_set_lhs (stmt
, name
);
3919 else if (is_gimple_call (stmt
))
3920 gimple_call_set_lhs (stmt
, name
);
3922 gimple_phi_set_result (stmt
, name
);
3924 replace_uses_by (lhs
, name
);
3925 release_ssa_name (lhs
);
3929 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
3930 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
3931 specifies whether the function should care about type incompatibility the
3932 current and new expressions. If it is false, the function will leave
3933 incompatibility issues to the caller. Return true iff the expression
3937 sra_ipa_modify_expr (tree
*expr
, bool convert
,
3938 ipa_parm_adjustment_vec adjustments
)
3941 struct ipa_parm_adjustment
*adj
, *cand
= NULL
;
3942 HOST_WIDE_INT offset
, size
, max_size
;
3945 len
= VEC_length (ipa_parm_adjustment_t
, adjustments
);
3947 if (TREE_CODE (*expr
) == BIT_FIELD_REF
3948 || TREE_CODE (*expr
) == IMAGPART_EXPR
3949 || TREE_CODE (*expr
) == REALPART_EXPR
)
3951 expr
= &TREE_OPERAND (*expr
, 0);
3955 base
= get_ref_base_and_extent (*expr
, &offset
, &size
, &max_size
);
3956 if (!base
|| size
== -1 || max_size
== -1)
3959 if (TREE_CODE (base
) == MEM_REF
)
3961 offset
+= mem_ref_offset (base
).low
* BITS_PER_UNIT
;
3962 base
= TREE_OPERAND (base
, 0);
3965 base
= get_ssa_base_param (base
);
3966 if (!base
|| TREE_CODE (base
) != PARM_DECL
)
3969 for (i
= 0; i
< len
; i
++)
3971 adj
= VEC_index (ipa_parm_adjustment_t
, adjustments
, i
);
3973 if (adj
->base
== base
&&
3974 (adj
->offset
== offset
|| adj
->remove_param
))
3980 if (!cand
|| cand
->copy_param
|| cand
->remove_param
)
3984 src
= build_simple_mem_ref (cand
->reduction
);
3986 src
= cand
->reduction
;
3988 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
3990 fprintf (dump_file
, "About to replace expr ");
3991 print_generic_expr (dump_file
, *expr
, 0);
3992 fprintf (dump_file
, " with ");
3993 print_generic_expr (dump_file
, src
, 0);
3994 fprintf (dump_file
, "\n");
3997 if (convert
&& !useless_type_conversion_p (TREE_TYPE (*expr
), cand
->type
))
3999 tree vce
= build1 (VIEW_CONVERT_EXPR
, TREE_TYPE (*expr
), src
);
4007 /* If the statement pointed to by STMT_PTR contains any expressions that need
4008 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4009 potential type incompatibilities (GSI is used to accommodate conversion
4010 statements and must point to the statement). Return true iff the statement
4014 sra_ipa_modify_assign (gimple
*stmt_ptr
, gimple_stmt_iterator
*gsi
,
4015 ipa_parm_adjustment_vec adjustments
)
4017 gimple stmt
= *stmt_ptr
;
4018 tree
*lhs_p
, *rhs_p
;
4021 if (!gimple_assign_single_p (stmt
))
4024 rhs_p
= gimple_assign_rhs1_ptr (stmt
);
4025 lhs_p
= gimple_assign_lhs_ptr (stmt
);
4027 any
= sra_ipa_modify_expr (rhs_p
, false, adjustments
);
4028 any
|= sra_ipa_modify_expr (lhs_p
, false, adjustments
);
4031 tree new_rhs
= NULL_TREE
;
4033 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p
), TREE_TYPE (*rhs_p
)))
4035 if (TREE_CODE (*rhs_p
) == CONSTRUCTOR
)
4037 /* V_C_Es of constructors can cause trouble (PR 42714). */
4038 if (is_gimple_reg_type (TREE_TYPE (*lhs_p
)))
4039 *rhs_p
= fold_convert (TREE_TYPE (*lhs_p
), integer_zero_node
);
4041 *rhs_p
= build_constructor (TREE_TYPE (*lhs_p
), 0);
4044 new_rhs
= fold_build1_loc (gimple_location (stmt
),
4045 VIEW_CONVERT_EXPR
, TREE_TYPE (*lhs_p
),
4048 else if (REFERENCE_CLASS_P (*rhs_p
)
4049 && is_gimple_reg_type (TREE_TYPE (*lhs_p
))
4050 && !is_gimple_reg (*lhs_p
))
4051 /* This can happen when an assignment in between two single field
4052 structures is turned into an assignment in between two pointers to
4053 scalars (PR 42237). */
4058 tree tmp
= force_gimple_operand_gsi (gsi
, new_rhs
, true, NULL_TREE
,
4059 true, GSI_SAME_STMT
);
4061 gimple_assign_set_rhs_from_tree (gsi
, tmp
);
4070 /* Traverse the function body and all modifications as described in
4071 ADJUSTMENTS. Return true iff the CFG has been changed. */
4074 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments
)
4076 bool cfg_changed
= false;
4081 gimple_stmt_iterator gsi
;
4083 for (gsi
= gsi_start_phis (bb
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4084 replace_removed_params_ssa_names (gsi_stmt (gsi
), adjustments
);
4086 gsi
= gsi_start_bb (bb
);
4087 while (!gsi_end_p (gsi
))
4089 gimple stmt
= gsi_stmt (gsi
);
4090 bool modified
= false;
4094 switch (gimple_code (stmt
))
4097 t
= gimple_return_retval_ptr (stmt
);
4098 if (*t
!= NULL_TREE
)
4099 modified
|= sra_ipa_modify_expr (t
, true, adjustments
);
4103 modified
|= sra_ipa_modify_assign (&stmt
, &gsi
, adjustments
);
4104 modified
|= replace_removed_params_ssa_names (stmt
, adjustments
);
4108 /* Operands must be processed before the lhs. */
4109 for (i
= 0; i
< gimple_call_num_args (stmt
); i
++)
4111 t
= gimple_call_arg_ptr (stmt
, i
);
4112 modified
|= sra_ipa_modify_expr (t
, true, adjustments
);
4115 if (gimple_call_lhs (stmt
))
4117 t
= gimple_call_lhs_ptr (stmt
);
4118 modified
|= sra_ipa_modify_expr (t
, false, adjustments
);
4119 modified
|= replace_removed_params_ssa_names (stmt
,
4125 for (i
= 0; i
< gimple_asm_ninputs (stmt
); i
++)
4127 t
= &TREE_VALUE (gimple_asm_input_op (stmt
, i
));
4128 modified
|= sra_ipa_modify_expr (t
, true, adjustments
);
4130 for (i
= 0; i
< gimple_asm_noutputs (stmt
); i
++)
4132 t
= &TREE_VALUE (gimple_asm_output_op (stmt
, i
));
4133 modified
|= sra_ipa_modify_expr (t
, false, adjustments
);
4144 if (maybe_clean_eh_stmt (stmt
)
4145 && gimple_purge_dead_eh_edges (gimple_bb (stmt
)))
4155 /* Call gimple_debug_bind_reset_value on all debug statements describing
4156 gimple register parameters that are being removed or replaced. */
4159 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments
)
4163 len
= VEC_length (ipa_parm_adjustment_t
, adjustments
);
4164 for (i
= 0; i
< len
; i
++)
4166 struct ipa_parm_adjustment
*adj
;
4167 imm_use_iterator ui
;
4171 adj
= VEC_index (ipa_parm_adjustment_t
, adjustments
, i
);
4172 if (adj
->copy_param
|| !is_gimple_reg (adj
->base
))
4174 name
= gimple_default_def (cfun
, adj
->base
);
4177 FOR_EACH_IMM_USE_STMT (stmt
, ui
, name
)
4179 /* All other users must have been removed by
4180 ipa_sra_modify_function_body. */
4181 gcc_assert (is_gimple_debug (stmt
));
4182 gimple_debug_bind_reset_value (stmt
);
4188 /* Return true iff all callers have at least as many actual arguments as there
4189 are formal parameters in the current function. */
4192 all_callers_have_enough_arguments_p (struct cgraph_node
*node
)
4194 struct cgraph_edge
*cs
;
4195 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4196 if (!callsite_has_enough_arguments_p (cs
->call_stmt
))
4203 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4206 convert_callers (struct cgraph_node
*node
, tree old_decl
,
4207 ipa_parm_adjustment_vec adjustments
)
4209 tree old_cur_fndecl
= current_function_decl
;
4210 struct cgraph_edge
*cs
;
4211 basic_block this_block
;
4212 bitmap recomputed_callers
= BITMAP_ALLOC (NULL
);
4214 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4216 current_function_decl
= cs
->caller
->decl
;
4217 push_cfun (DECL_STRUCT_FUNCTION (cs
->caller
->decl
));
4220 fprintf (dump_file
, "Adjusting call (%i -> %i) %s -> %s\n",
4221 cs
->caller
->uid
, cs
->callee
->uid
,
4222 cgraph_node_name (cs
->caller
),
4223 cgraph_node_name (cs
->callee
));
4225 ipa_modify_call_arguments (cs
, cs
->call_stmt
, adjustments
);
4230 for (cs
= node
->callers
; cs
; cs
= cs
->next_caller
)
4231 if (bitmap_set_bit (recomputed_callers
, cs
->caller
->uid
))
4232 compute_inline_parameters (cs
->caller
);
4233 BITMAP_FREE (recomputed_callers
);
4235 current_function_decl
= old_cur_fndecl
;
4237 if (!encountered_recursive_call
)
4240 FOR_EACH_BB (this_block
)
4242 gimple_stmt_iterator gsi
;
4244 for (gsi
= gsi_start_bb (this_block
); !gsi_end_p (gsi
); gsi_next (&gsi
))
4246 gimple stmt
= gsi_stmt (gsi
);
4248 if (gimple_code (stmt
) != GIMPLE_CALL
)
4250 call_fndecl
= gimple_call_fndecl (stmt
);
4251 if (call_fndecl
== old_decl
)
4254 fprintf (dump_file
, "Adjusting recursive call");
4255 gimple_call_set_fndecl (stmt
, node
->decl
);
4256 ipa_modify_call_arguments (NULL
, stmt
, adjustments
);
4264 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4265 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4268 modify_function (struct cgraph_node
*node
, ipa_parm_adjustment_vec adjustments
)
4270 struct cgraph_node
*new_node
;
4271 struct cgraph_edge
*cs
;
4273 VEC (cgraph_edge_p
, heap
) * redirect_callers
;
4277 for (cs
= node
->callers
; cs
!= NULL
; cs
= cs
->next_caller
)
4279 redirect_callers
= VEC_alloc (cgraph_edge_p
, heap
, node_callers
);
4280 for (cs
= node
->callers
; cs
!= NULL
; cs
= cs
->next_caller
)
4281 VEC_quick_push (cgraph_edge_p
, redirect_callers
, cs
);
4283 rebuild_cgraph_edges ();
4285 current_function_decl
= NULL_TREE
;
4287 new_node
= cgraph_function_versioning (node
, redirect_callers
, NULL
, NULL
,
4288 NULL
, NULL
, "isra");
4289 current_function_decl
= new_node
->decl
;
4290 push_cfun (DECL_STRUCT_FUNCTION (new_node
->decl
));
4292 ipa_modify_formal_parameters (current_function_decl
, adjustments
, "ISRA");
4293 cfg_changed
= ipa_sra_modify_function_body (adjustments
);
4294 sra_ipa_reset_debug_stmts (adjustments
);
4295 convert_callers (new_node
, node
->decl
, adjustments
);
4296 cgraph_make_node_local (new_node
);
4300 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4301 attributes, return true otherwise. NODE is the cgraph node of the current
4305 ipa_sra_preliminary_function_checks (struct cgraph_node
*node
)
4307 if (!cgraph_node_can_be_local_p (node
))
4310 fprintf (dump_file
, "Function not local to this compilation unit.\n");
4314 if (!tree_versionable_function_p (node
->decl
))
4317 fprintf (dump_file
, "Function is not versionable.\n");
4321 if (DECL_VIRTUAL_P (current_function_decl
))
4324 fprintf (dump_file
, "Function is a virtual method.\n");
4328 if ((DECL_COMDAT (node
->decl
) || DECL_EXTERNAL (node
->decl
))
4329 && node
->global
.size
>= MAX_INLINE_INSNS_AUTO
)
4332 fprintf (dump_file
, "Function too big to be made truly local.\n");
4340 "Function has no callers in this compilation unit.\n");
4347 fprintf (dump_file
, "Function uses stdarg. \n");
4351 if (TYPE_ATTRIBUTES (TREE_TYPE (node
->decl
)))
4357 /* Perform early interprocedural SRA. */
4360 ipa_early_sra (void)
4362 struct cgraph_node
*node
= cgraph_node (current_function_decl
);
4363 ipa_parm_adjustment_vec adjustments
;
4366 if (!ipa_sra_preliminary_function_checks (node
))
4370 sra_mode
= SRA_MODE_EARLY_IPA
;
4372 if (!find_param_candidates ())
4375 fprintf (dump_file
, "Function has no IPA-SRA candidates.\n");
4379 if (!all_callers_have_enough_arguments_p (node
))
4382 fprintf (dump_file
, "There are callers with insufficient number of "
4387 bb_dereferences
= XCNEWVEC (HOST_WIDE_INT
,
4389 * last_basic_block_for_function (cfun
));
4390 final_bbs
= BITMAP_ALLOC (NULL
);
4393 if (encountered_apply_args
)
4396 fprintf (dump_file
, "Function calls __builtin_apply_args().\n");
4400 if (encountered_unchangable_recursive_call
)
4403 fprintf (dump_file
, "Function calls itself with insufficient "
4404 "number of arguments.\n");
4408 adjustments
= analyze_all_param_acesses ();
4412 ipa_dump_param_adjustments (dump_file
, adjustments
, current_function_decl
);
4414 if (modify_function (node
, adjustments
))
4415 ret
= TODO_update_ssa
| TODO_cleanup_cfg
;
4417 ret
= TODO_update_ssa
;
4418 VEC_free (ipa_parm_adjustment_t
, heap
, adjustments
);
4420 statistics_counter_event (cfun
, "Unused parameters deleted",
4421 sra_stats
.deleted_unused_parameters
);
4422 statistics_counter_event (cfun
, "Scalar parameters converted to by-value",
4423 sra_stats
.scalar_by_ref_to_by_val
);
4424 statistics_counter_event (cfun
, "Aggregate parameters broken up",
4425 sra_stats
.aggregate_params_reduced
);
4426 statistics_counter_event (cfun
, "Aggregate parameter components created",
4427 sra_stats
.param_reductions_created
);
4430 BITMAP_FREE (final_bbs
);
4431 free (bb_dereferences
);
4433 sra_deinitialize ();
4437 /* Return if early ipa sra shall be performed. */
4439 ipa_early_sra_gate (void)
4441 return flag_ipa_sra
&& dbg_cnt (eipa_sra
);
4444 struct gimple_opt_pass pass_early_ipa_sra
=
4448 "eipa_sra", /* name */
4449 ipa_early_sra_gate
, /* gate */
4450 ipa_early_sra
, /* execute */
4453 0, /* static_pass_number */
4454 TV_IPA_SRA
, /* tv_id */
4455 0, /* properties_required */
4456 0, /* properties_provided */
4457 0, /* properties_destroyed */
4458 0, /* todo_flags_start */
4459 TODO_dump_func
| TODO_dump_cgraph
/* todo_flags_finish */