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 "diagnostic.h"
86 #include "statistics.h"
87 #include "tree-dump.h"
93 /* Enumeration of all aggregate reductions we can do. */
94 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
95 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
96 SRA_MODE_INTRA }; /* late intraprocedural SRA */
98 /* Global variable describing which aggregate reduction we are performing at
100 static enum sra_mode sra_mode;
104 /* ACCESS represents each access to an aggregate variable (as a whole or a
105 part). It can also represent a group of accesses that refer to exactly the
106 same fragment of an aggregate (i.e. those that have exactly the same offset
107 and size). Such representatives for a single aggregate, once determined,
108 are linked in a linked list and have the group fields set.
110 Moreover, when doing intraprocedural SRA, a tree is built from those
111 representatives (by the means of first_child and next_sibling pointers), in
112 which all items in a subtree are "within" the root, i.e. their offset is
113 greater or equal to offset of the root and offset+size is smaller or equal
114 to offset+size of the root. Children of an access are sorted by offset.
116 Note that accesses to parts of vector and complex number types always
117 represented by an access to the whole complex number or a vector. It is a
118 duty of the modifying functions to replace them appropriately. */
122 /* Values returned by `get_ref_base_and_extent' for each component reference
123 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
124 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
125 HOST_WIDE_INT offset;
129 /* Expression. It is context dependent so do not use it to create new
130 expressions to access the original aggregate. See PR 42154 for a
136 /* The statement this access belongs to. */
139 /* Next group representative for this aggregate. */
140 struct access *next_grp;
142 /* Pointer to the group representative. Pointer to itself if the struct is
143 the representative. */
144 struct access *group_representative;
146 /* If this access has any children (in terms of the definition above), this
147 points to the first one. */
148 struct access *first_child;
150 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
151 described above. In IPA-SRA this is a pointer to the next access
152 belonging to the same group (having the same representative). */
153 struct access *next_sibling;
155 /* Pointers to the first and last element in the linked list of assign
157 struct assign_link *first_link, *last_link;
159 /* Pointer to the next access in the work queue. */
160 struct access *next_queued;
162 /* Replacement variable for this access "region." Never to be accessed
163 directly, always only by the means of get_access_replacement() and only
164 when grp_to_be_replaced flag is set. */
165 tree replacement_decl;
167 /* Is this particular access write access? */
170 /* Is this access an artificial one created to scalarize some record
172 unsigned total_scalarization : 1;
174 /* Is this access currently in the work queue? */
175 unsigned grp_queued : 1;
177 /* Does this group contain a write access? This flag is propagated down the
179 unsigned grp_write : 1;
181 /* Does this group contain a read access? This flag is propagated down the
183 unsigned grp_read : 1;
185 /* Does this group contain a read access that comes from an assignment
186 statement? This flag is propagated down the access tree. */
187 unsigned grp_assignment_read : 1;
189 /* Other passes of the analysis use this bit to make function
190 analyze_access_subtree create scalar replacements for this group if
192 unsigned grp_hint : 1;
194 /* Is the subtree rooted in this access fully covered by scalar
196 unsigned grp_covered : 1;
198 /* If set to true, this access and all below it in an access tree must not be
200 unsigned grp_unscalarizable_region : 1;
202 /* Whether data have been written to parts of the aggregate covered by this
203 access which is not to be scalarized. This flag is propagated up in the
205 unsigned grp_unscalarized_data : 1;
207 /* Does this access and/or group contain a write access through a
209 unsigned grp_partial_lhs : 1;
211 /* Set when a scalar replacement should be created for this variable. We do
212 the decision and creation at different places because create_tmp_var
213 cannot be called from within FOR_EACH_REFERENCED_VAR. */
214 unsigned grp_to_be_replaced : 1;
216 /* Is it possible that the group refers to data which might be (directly or
217 otherwise) modified? */
218 unsigned grp_maybe_modified : 1;
220 /* Set when this is a representative of a pointer to scalar (i.e. by
221 reference) parameter which we consider for turning into a plain scalar
222 (i.e. a by value parameter). */
223 unsigned grp_scalar_ptr : 1;
225 /* Set when we discover that this pointer is not safe to dereference in the
227 unsigned grp_not_necessarilly_dereferenced : 1;
230 typedef struct access *access_p;
232 DEF_VEC_P (access_p);
233 DEF_VEC_ALLOC_P (access_p, heap);
235 /* Alloc pool for allocating access structures. */
236 static alloc_pool access_pool;
238 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
239 are used to propagate subaccesses from rhs to lhs as long as they don't
240 conflict with what is already there. */
243 struct access *lacc, *racc;
244 struct assign_link *next;
247 /* Alloc pool for allocating assign link structures. */
248 static alloc_pool link_pool;
250 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
251 static struct pointer_map_t *base_access_vec;
253 /* Bitmap of candidates. */
254 static bitmap candidate_bitmap;
256 /* Bitmap of candidates which we should try to entirely scalarize away and
257 those which cannot be (because they are and need be used as a whole). */
258 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
260 /* Obstack for creation of fancy names. */
261 static struct obstack name_obstack;
263 /* Head of a linked list of accesses that need to have its subaccesses
264 propagated to their assignment counterparts. */
265 static struct access *work_queue_head;
267 /* Number of parameters of the analyzed function when doing early ipa SRA. */
268 static int func_param_count;
270 /* scan_function sets the following to true if it encounters a call to
271 __builtin_apply_args. */
272 static bool encountered_apply_args;
274 /* Set by scan_function when it finds a recursive call. */
275 static bool encountered_recursive_call;
277 /* Set by scan_function when it finds a recursive call with less actual
278 arguments than formal parameters.. */
279 static bool encountered_unchangable_recursive_call;
281 /* This is a table in which for each basic block and parameter there is a
282 distance (offset + size) in that parameter which is dereferenced and
283 accessed in that BB. */
284 static HOST_WIDE_INT *bb_dereferences;
285 /* Bitmap of BBs that can cause the function to "stop" progressing by
286 returning, throwing externally, looping infinitely or calling a function
287 which might abort etc.. */
288 static bitmap final_bbs;
290 /* Representative of no accesses at all. */
291 static struct access no_accesses_representant;
293 /* Predicate to test the special value. */
296 no_accesses_p (struct access *access)
298 return access == &no_accesses_representant;
301 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
302 representative fields are dumped, otherwise those which only describe the
303 individual access are. */
307 /* Number of processed aggregates is readily available in
308 analyze_all_variable_accesses and so is not stored here. */
310 /* Number of created scalar replacements. */
313 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
317 /* Number of statements created by generate_subtree_copies. */
320 /* Number of statements created by load_assign_lhs_subreplacements. */
323 /* Number of times sra_modify_assign has deleted a statement. */
326 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
327 RHS reparately due to type conversions or nonexistent matching
329 int separate_lhs_rhs_handling;
331 /* Number of parameters that were removed because they were unused. */
332 int deleted_unused_parameters;
334 /* Number of scalars passed as parameters by reference that have been
335 converted to be passed by value. */
336 int scalar_by_ref_to_by_val;
338 /* Number of aggregate parameters that were replaced by one or more of their
340 int aggregate_params_reduced;
342 /* Numbber of components created when splitting aggregate parameters. */
343 int param_reductions_created;
347 dump_access (FILE *f, struct access *access, bool grp)
349 fprintf (f, "access { ");
350 fprintf (f, "base = (%d)'", DECL_UID (access->base));
351 print_generic_expr (f, access->base, 0);
352 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
353 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
354 fprintf (f, ", expr = ");
355 print_generic_expr (f, access->expr, 0);
356 fprintf (f, ", type = ");
357 print_generic_expr (f, access->type, 0);
359 fprintf (f, ", grp_write = %d, total_scalarization = %d, "
360 "grp_read = %d, grp_hint = %d, "
361 "grp_covered = %d, grp_unscalarizable_region = %d, "
362 "grp_unscalarized_data = %d, grp_partial_lhs = %d, "
363 "grp_to_be_replaced = %d, grp_maybe_modified = %d, "
364 "grp_not_necessarilly_dereferenced = %d\n",
365 access->grp_write, access->total_scalarization,
366 access->grp_read, access->grp_hint,
367 access->grp_covered, access->grp_unscalarizable_region,
368 access->grp_unscalarized_data, access->grp_partial_lhs,
369 access->grp_to_be_replaced, access->grp_maybe_modified,
370 access->grp_not_necessarilly_dereferenced);
372 fprintf (f, ", write = %d, total_scalarization = %d, "
373 "grp_partial_lhs = %d\n",
374 access->write, access->total_scalarization,
375 access->grp_partial_lhs);
378 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
381 dump_access_tree_1 (FILE *f, struct access *access, int level)
387 for (i = 0; i < level; i++)
388 fputs ("* ", dump_file);
390 dump_access (f, access, true);
392 if (access->first_child)
393 dump_access_tree_1 (f, access->first_child, level + 1);
395 access = access->next_sibling;
400 /* Dump all access trees for a variable, given the pointer to the first root in
404 dump_access_tree (FILE *f, struct access *access)
406 for (; access; access = access->next_grp)
407 dump_access_tree_1 (f, access, 0);
410 /* Return true iff ACC is non-NULL and has subaccesses. */
413 access_has_children_p (struct access *acc)
415 return acc && acc->first_child;
418 /* Return a vector of pointers to accesses for the variable given in BASE or
419 NULL if there is none. */
421 static VEC (access_p, heap) *
422 get_base_access_vector (tree base)
426 slot = pointer_map_contains (base_access_vec, base);
430 return *(VEC (access_p, heap) **) slot;
433 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
434 in ACCESS. Return NULL if it cannot be found. */
436 static struct access *
437 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
440 while (access && (access->offset != offset || access->size != size))
442 struct access *child = access->first_child;
444 while (child && (child->offset + child->size <= offset))
445 child = child->next_sibling;
452 /* Return the first group representative for DECL or NULL if none exists. */
454 static struct access *
455 get_first_repr_for_decl (tree base)
457 VEC (access_p, heap) *access_vec;
459 access_vec = get_base_access_vector (base);
463 return VEC_index (access_p, access_vec, 0);
466 /* Find an access representative for the variable BASE and given OFFSET and
467 SIZE. Requires that access trees have already been built. Return NULL if
468 it cannot be found. */
470 static struct access *
471 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
474 struct access *access;
476 access = get_first_repr_for_decl (base);
477 while (access && (access->offset + access->size <= offset))
478 access = access->next_grp;
482 return find_access_in_subtree (access, offset, size);
485 /* Add LINK to the linked list of assign links of RACC. */
487 add_link_to_rhs (struct access *racc, struct assign_link *link)
489 gcc_assert (link->racc == racc);
491 if (!racc->first_link)
493 gcc_assert (!racc->last_link);
494 racc->first_link = link;
497 racc->last_link->next = link;
499 racc->last_link = link;
503 /* Move all link structures in their linked list in OLD_RACC to the linked list
506 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
508 if (!old_racc->first_link)
510 gcc_assert (!old_racc->last_link);
514 if (new_racc->first_link)
516 gcc_assert (!new_racc->last_link->next);
517 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
519 new_racc->last_link->next = old_racc->first_link;
520 new_racc->last_link = old_racc->last_link;
524 gcc_assert (!new_racc->last_link);
526 new_racc->first_link = old_racc->first_link;
527 new_racc->last_link = old_racc->last_link;
529 old_racc->first_link = old_racc->last_link = NULL;
532 /* Add ACCESS to the work queue (which is actually a stack). */
535 add_access_to_work_queue (struct access *access)
537 if (!access->grp_queued)
539 gcc_assert (!access->next_queued);
540 access->next_queued = work_queue_head;
541 access->grp_queued = 1;
542 work_queue_head = access;
546 /* Pop an access from the work queue, and return it, assuming there is one. */
548 static struct access *
549 pop_access_from_work_queue (void)
551 struct access *access = work_queue_head;
553 work_queue_head = access->next_queued;
554 access->next_queued = NULL;
555 access->grp_queued = 0;
560 /* Allocate necessary structures. */
563 sra_initialize (void)
565 candidate_bitmap = BITMAP_ALLOC (NULL);
566 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
567 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
568 gcc_obstack_init (&name_obstack);
569 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
570 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
571 base_access_vec = pointer_map_create ();
572 memset (&sra_stats, 0, sizeof (sra_stats));
573 encountered_apply_args = false;
574 encountered_recursive_call = false;
575 encountered_unchangable_recursive_call = false;
578 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
581 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
582 void *data ATTRIBUTE_UNUSED)
584 VEC (access_p, heap) *access_vec;
585 access_vec = (VEC (access_p, heap) *) *value;
586 VEC_free (access_p, heap, access_vec);
591 /* Deallocate all general structures. */
594 sra_deinitialize (void)
596 BITMAP_FREE (candidate_bitmap);
597 BITMAP_FREE (should_scalarize_away_bitmap);
598 BITMAP_FREE (cannot_scalarize_away_bitmap);
599 free_alloc_pool (access_pool);
600 free_alloc_pool (link_pool);
601 obstack_free (&name_obstack, NULL);
603 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
604 pointer_map_destroy (base_access_vec);
607 /* Remove DECL from candidates for SRA and write REASON to the dump file if
610 disqualify_candidate (tree decl, const char *reason)
612 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
614 if (dump_file && (dump_flags & TDF_DETAILS))
616 fprintf (dump_file, "! Disqualifying ");
617 print_generic_expr (dump_file, decl, 0);
618 fprintf (dump_file, " - %s\n", reason);
622 /* Return true iff the type contains a field or an element which does not allow
626 type_internals_preclude_sra_p (tree type)
631 switch (TREE_CODE (type))
635 case QUAL_UNION_TYPE:
636 for (fld = TYPE_FIELDS (type); fld; fld = TREE_CHAIN (fld))
637 if (TREE_CODE (fld) == FIELD_DECL)
639 tree ft = TREE_TYPE (fld);
641 if (TREE_THIS_VOLATILE (fld)
642 || !DECL_FIELD_OFFSET (fld) || !DECL_SIZE (fld)
643 || !host_integerp (DECL_FIELD_OFFSET (fld), 1)
644 || !host_integerp (DECL_SIZE (fld), 1))
647 if (AGGREGATE_TYPE_P (ft)
648 && type_internals_preclude_sra_p (ft))
655 et = TREE_TYPE (type);
657 if (AGGREGATE_TYPE_P (et))
658 return type_internals_preclude_sra_p (et);
667 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
668 base variable if it is. Return T if it is not an SSA_NAME. */
671 get_ssa_base_param (tree t)
673 if (TREE_CODE (t) == SSA_NAME)
675 if (SSA_NAME_IS_DEFAULT_DEF (t))
676 return SSA_NAME_VAR (t);
683 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
684 belongs to, unless the BB has already been marked as a potentially
688 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
690 basic_block bb = gimple_bb (stmt);
691 int idx, parm_index = 0;
694 if (bitmap_bit_p (final_bbs, bb->index))
697 for (parm = DECL_ARGUMENTS (current_function_decl);
698 parm && parm != base;
699 parm = TREE_CHAIN (parm))
702 gcc_assert (parm_index < func_param_count);
704 idx = bb->index * func_param_count + parm_index;
705 if (bb_dereferences[idx] < dist)
706 bb_dereferences[idx] = dist;
709 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
710 the three fields. Also add it to the vector of accesses corresponding to
711 the base. Finally, return the new access. */
713 static struct access *
714 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
716 VEC (access_p, heap) *vec;
717 struct access *access;
720 access = (struct access *) pool_alloc (access_pool);
721 memset (access, 0, sizeof (struct access));
723 access->offset = offset;
726 slot = pointer_map_contains (base_access_vec, base);
728 vec = (VEC (access_p, heap) *) *slot;
730 vec = VEC_alloc (access_p, heap, 32);
732 VEC_safe_push (access_p, heap, vec, access);
734 *((struct VEC (access_p,heap) **)
735 pointer_map_insert (base_access_vec, base)) = vec;
740 /* Create and insert access for EXPR. Return created access, or NULL if it is
743 static struct access *
744 create_access (tree expr, gimple stmt, bool write)
746 struct access *access;
747 HOST_WIDE_INT offset, size, max_size;
749 bool ptr, unscalarizable_region = false;
751 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
753 if (sra_mode == SRA_MODE_EARLY_IPA && INDIRECT_REF_P (base))
755 base = get_ssa_base_param (TREE_OPERAND (base, 0));
763 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
766 if (sra_mode == SRA_MODE_EARLY_IPA)
768 if (size < 0 || size != max_size)
770 disqualify_candidate (base, "Encountered a variable sized access.");
773 if ((offset % BITS_PER_UNIT) != 0 || (size % BITS_PER_UNIT) != 0)
775 disqualify_candidate (base,
776 "Encountered an acces not aligned to a byte.");
781 mark_parm_dereference (base, offset + size, stmt);
785 if (size != max_size)
788 unscalarizable_region = true;
792 disqualify_candidate (base, "Encountered an unconstrained access.");
797 access = create_access_1 (base, offset, size);
799 access->type = TREE_TYPE (expr);
800 access->write = write;
801 access->grp_unscalarizable_region = unscalarizable_region;
808 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
809 register types or (recursively) records with only these two kinds of fields.
810 It also returns false if any of these records has a zero-size field as its
814 type_consists_of_records_p (tree type)
817 bool last_fld_has_zero_size = false;
819 if (TREE_CODE (type) != RECORD_TYPE)
822 for (fld = TYPE_FIELDS (type); fld; fld = TREE_CHAIN (fld))
823 if (TREE_CODE (fld) == FIELD_DECL)
825 tree ft = TREE_TYPE (fld);
827 if (!is_gimple_reg_type (ft)
828 && !type_consists_of_records_p (ft))
831 last_fld_has_zero_size = tree_low_cst (DECL_SIZE (fld), 1) == 0;
834 if (last_fld_has_zero_size)
840 /* Create total_scalarization accesses for all scalar type fields in DECL that
841 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
842 must be the top-most VAR_DECL representing the variable, OFFSET must be the
843 offset of DECL within BASE. */
846 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset)
848 tree fld, decl_type = TREE_TYPE (decl);
850 for (fld = TYPE_FIELDS (decl_type); fld; fld = TREE_CHAIN (fld))
851 if (TREE_CODE (fld) == FIELD_DECL)
853 HOST_WIDE_INT pos = offset + int_bit_position (fld);
854 tree ft = TREE_TYPE (fld);
856 if (is_gimple_reg_type (ft))
858 struct access *access;
863 size = tree_low_cst (DECL_SIZE (fld), 1);
865 ok = build_ref_for_offset (&expr, TREE_TYPE (base), pos,
869 access = create_access_1 (base, pos, size);
872 access->total_scalarization = 1;
873 /* Accesses for intraprocedural SRA can have their stmt NULL. */
876 completely_scalarize_record (base, fld, pos);
881 /* Search the given tree for a declaration by skipping handled components and
882 exclude it from the candidates. */
885 disqualify_base_of_expr (tree t, const char *reason)
887 while (handled_component_p (t))
888 t = TREE_OPERAND (t, 0);
890 if (sra_mode == SRA_MODE_EARLY_IPA)
892 if (INDIRECT_REF_P (t))
893 t = TREE_OPERAND (t, 0);
894 t = get_ssa_base_param (t);
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 (sra_mode != SRA_MODE_EARLY_IPA)
946 case ARRAY_RANGE_REF:
947 ret = create_access (expr, stmt, write);
954 if (write && partial_ref && ret)
955 ret->grp_partial_lhs = 1;
960 /* Scan expression EXPR and create access structures for all accesses to
961 candidates for scalarization. Return true if any access has been inserted.
962 STMT must be the statement from which the expression is taken, WRITE must be
963 true if the expression is a store and false otherwise. */
966 build_access_from_expr (tree expr, gimple stmt, bool write)
968 struct access *access;
970 access = build_access_from_expr_1 (expr, stmt, write);
973 /* This means the aggregate is accesses as a whole in a way other than an
974 assign statement and thus cannot be removed even if we had a scalar
975 replacement for everything. */
976 if (cannot_scalarize_away_bitmap)
977 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
983 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
984 modes in which it matters, return true iff they have been disqualified. RHS
985 may be NULL, in that case ignore it. If we scalarize an aggregate in
986 intra-SRA we may need to add statements after each statement. This is not
987 possible if a statement unconditionally has to end the basic block. */
989 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
991 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
992 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
994 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
996 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1002 /* Scan expressions occuring in STMT, create access structures for all accesses
1003 to candidates for scalarization and remove those candidates which occur in
1004 statements or expressions that prevent them from being split apart. Return
1005 true if any access has been inserted. */
1008 build_accesses_from_assign (gimple stmt)
1011 struct access *lacc, *racc;
1013 if (!gimple_assign_single_p (stmt))
1016 lhs = gimple_assign_lhs (stmt);
1017 rhs = gimple_assign_rhs1 (stmt);
1019 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1022 racc = build_access_from_expr_1 (rhs, stmt, false);
1023 lacc = build_access_from_expr_1 (lhs, stmt, true);
1027 racc->grp_assignment_read = 1;
1028 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1029 && !is_gimple_reg_type (racc->type))
1030 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1034 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1035 && !lacc->grp_unscalarizable_region
1036 && !racc->grp_unscalarizable_region
1037 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1038 /* FIXME: Turn the following line into an assert after PR 40058 is
1040 && lacc->size == racc->size
1041 && useless_type_conversion_p (lacc->type, racc->type))
1043 struct assign_link *link;
1045 link = (struct assign_link *) pool_alloc (link_pool);
1046 memset (link, 0, sizeof (struct assign_link));
1051 add_link_to_rhs (racc, link);
1054 return lacc || racc;
1057 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1058 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1061 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1062 void *data ATTRIBUTE_UNUSED)
1064 op = get_base_address (op);
1067 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1072 /* Return true iff callsite CALL has at least as many actual arguments as there
1073 are formal parameters of the function currently processed by IPA-SRA. */
1076 callsite_has_enough_arguments_p (gimple call)
1078 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1081 /* Scan function and look for interesting expressions and create access
1082 structures for them. Return true iff any access is created. */
1085 scan_function (void)
1092 gimple_stmt_iterator gsi;
1093 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1095 gimple stmt = gsi_stmt (gsi);
1099 if (final_bbs && stmt_can_throw_external (stmt))
1100 bitmap_set_bit (final_bbs, bb->index);
1101 switch (gimple_code (stmt))
1104 t = gimple_return_retval (stmt);
1106 ret |= build_access_from_expr (t, stmt, false);
1108 bitmap_set_bit (final_bbs, bb->index);
1112 ret |= build_accesses_from_assign (stmt);
1116 for (i = 0; i < gimple_call_num_args (stmt); i++)
1117 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1120 if (sra_mode == SRA_MODE_EARLY_IPA)
1122 tree dest = gimple_call_fndecl (stmt);
1123 int flags = gimple_call_flags (stmt);
1127 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1128 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1129 encountered_apply_args = true;
1130 if (cgraph_get_node (dest)
1131 == cgraph_get_node (current_function_decl))
1133 encountered_recursive_call = true;
1134 if (!callsite_has_enough_arguments_p (stmt))
1135 encountered_unchangable_recursive_call = true;
1140 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1141 bitmap_set_bit (final_bbs, bb->index);
1144 t = gimple_call_lhs (stmt);
1145 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1146 ret |= build_access_from_expr (t, stmt, true);
1150 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1153 bitmap_set_bit (final_bbs, bb->index);
1155 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1157 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1158 ret |= build_access_from_expr (t, stmt, false);
1160 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1162 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1163 ret |= build_access_from_expr (t, stmt, true);
1176 /* Helper of QSORT function. There are pointers to accesses in the array. An
1177 access is considered smaller than another if it has smaller offset or if the
1178 offsets are the same but is size is bigger. */
1181 compare_access_positions (const void *a, const void *b)
1183 const access_p *fp1 = (const access_p *) a;
1184 const access_p *fp2 = (const access_p *) b;
1185 const access_p f1 = *fp1;
1186 const access_p f2 = *fp2;
1188 if (f1->offset != f2->offset)
1189 return f1->offset < f2->offset ? -1 : 1;
1191 if (f1->size == f2->size)
1193 if (f1->type == f2->type)
1195 /* Put any non-aggregate type before any aggregate type. */
1196 else if (!is_gimple_reg_type (f1->type)
1197 && is_gimple_reg_type (f2->type))
1199 else if (is_gimple_reg_type (f1->type)
1200 && !is_gimple_reg_type (f2->type))
1202 /* Put any complex or vector type before any other scalar type. */
1203 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1204 && TREE_CODE (f1->type) != VECTOR_TYPE
1205 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1206 || TREE_CODE (f2->type) == VECTOR_TYPE))
1208 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1209 || TREE_CODE (f1->type) == VECTOR_TYPE)
1210 && TREE_CODE (f2->type) != COMPLEX_TYPE
1211 && TREE_CODE (f2->type) != VECTOR_TYPE)
1213 /* Put the integral type with the bigger precision first. */
1214 else if (INTEGRAL_TYPE_P (f1->type)
1215 && INTEGRAL_TYPE_P (f2->type))
1216 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1217 /* Put any integral type with non-full precision last. */
1218 else if (INTEGRAL_TYPE_P (f1->type)
1219 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1220 != TYPE_PRECISION (f1->type)))
1222 else if (INTEGRAL_TYPE_P (f2->type)
1223 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1224 != TYPE_PRECISION (f2->type)))
1226 /* Stabilize the sort. */
1227 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1230 /* We want the bigger accesses first, thus the opposite operator in the next
1232 return f1->size > f2->size ? -1 : 1;
1236 /* Append a name of the declaration to the name obstack. A helper function for
1240 make_fancy_decl_name (tree decl)
1244 tree name = DECL_NAME (decl);
1246 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1247 IDENTIFIER_LENGTH (name));
1250 sprintf (buffer, "D%u", DECL_UID (decl));
1251 obstack_grow (&name_obstack, buffer, strlen (buffer));
1255 /* Helper for make_fancy_name. */
1258 make_fancy_name_1 (tree expr)
1265 make_fancy_decl_name (expr);
1269 switch (TREE_CODE (expr))
1272 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1273 obstack_1grow (&name_obstack, '$');
1274 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1278 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1279 obstack_1grow (&name_obstack, '$');
1280 /* Arrays with only one element may not have a constant as their
1282 index = TREE_OPERAND (expr, 1);
1283 if (TREE_CODE (index) != INTEGER_CST)
1285 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1286 obstack_grow (&name_obstack, buffer, strlen (buffer));
1293 gcc_unreachable (); /* we treat these as scalars. */
1300 /* Create a human readable name for replacement variable of ACCESS. */
1303 make_fancy_name (tree expr)
1305 make_fancy_name_1 (expr);
1306 obstack_1grow (&name_obstack, '\0');
1307 return XOBFINISH (&name_obstack, char *);
1310 /* Helper function for build_ref_for_offset. */
1313 build_ref_for_offset_1 (tree *res, tree type, HOST_WIDE_INT offset,
1319 tree tr_size, index, minidx;
1320 HOST_WIDE_INT el_size;
1322 if (offset == 0 && exp_type
1323 && types_compatible_p (exp_type, type))
1326 switch (TREE_CODE (type))
1329 case QUAL_UNION_TYPE:
1331 for (fld = TYPE_FIELDS (type); fld; fld = TREE_CHAIN (fld))
1333 HOST_WIDE_INT pos, size;
1334 tree expr, *expr_ptr;
1336 if (TREE_CODE (fld) != FIELD_DECL)
1339 pos = int_bit_position (fld);
1340 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1341 tr_size = DECL_SIZE (fld);
1342 if (!tr_size || !host_integerp (tr_size, 1))
1344 size = tree_low_cst (tr_size, 1);
1350 else if (pos > offset || (pos + size) <= offset)
1355 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1361 if (build_ref_for_offset_1 (expr_ptr, TREE_TYPE (fld),
1362 offset - pos, exp_type))
1372 tr_size = TYPE_SIZE (TREE_TYPE (type));
1373 if (!tr_size || !host_integerp (tr_size, 1))
1375 el_size = tree_low_cst (tr_size, 1);
1377 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1378 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1382 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1383 if (!integer_zerop (minidx))
1384 index = int_const_binop (PLUS_EXPR, index, minidx, 0);
1385 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1386 NULL_TREE, NULL_TREE);
1388 offset = offset % el_size;
1389 type = TREE_TYPE (type);
1404 /* Construct an expression that would reference a part of aggregate *EXPR of
1405 type TYPE at the given OFFSET of the type EXP_TYPE. If EXPR is NULL, the
1406 function only determines whether it can build such a reference without
1407 actually doing it, otherwise, the tree it points to is unshared first and
1408 then used as a base for furhter sub-references.
1410 FIXME: Eventually this should be replaced with
1411 maybe_fold_offset_to_reference() from tree-ssa-ccp.c but that requires a
1412 minor rewrite of fold_stmt.
1416 build_ref_for_offset (tree *expr, tree type, HOST_WIDE_INT offset,
1417 tree exp_type, bool allow_ptr)
1419 location_t loc = expr ? EXPR_LOCATION (*expr) : UNKNOWN_LOCATION;
1422 *expr = unshare_expr (*expr);
1424 if (allow_ptr && POINTER_TYPE_P (type))
1426 type = TREE_TYPE (type);
1428 *expr = fold_build1_loc (loc, INDIRECT_REF, type, *expr);
1431 return build_ref_for_offset_1 (expr, type, offset, exp_type);
1434 /* Return true iff TYPE is stdarg va_list type. */
1437 is_va_list_type (tree type)
1439 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1442 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1443 those with type which is suitable for scalarization. */
1446 find_var_candidates (void)
1449 referenced_var_iterator rvi;
1452 FOR_EACH_REFERENCED_VAR (var, rvi)
1454 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1456 type = TREE_TYPE (var);
1458 if (!AGGREGATE_TYPE_P (type)
1459 || needs_to_live_in_memory (var)
1460 || TREE_THIS_VOLATILE (var)
1461 || !COMPLETE_TYPE_P (type)
1462 || !host_integerp (TYPE_SIZE (type), 1)
1463 || tree_low_cst (TYPE_SIZE (type), 1) == 0
1464 || type_internals_preclude_sra_p (type)
1465 /* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1466 we also want to schedule it rather late. Thus we ignore it in
1468 || (sra_mode == SRA_MODE_EARLY_INTRA
1469 && is_va_list_type (type)))
1472 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1474 if (dump_file && (dump_flags & TDF_DETAILS))
1476 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1477 print_generic_expr (dump_file, var, 0);
1478 fprintf (dump_file, "\n");
1486 /* Sort all accesses for the given variable, check for partial overlaps and
1487 return NULL if there are any. If there are none, pick a representative for
1488 each combination of offset and size and create a linked list out of them.
1489 Return the pointer to the first representative and make sure it is the first
1490 one in the vector of accesses. */
1492 static struct access *
1493 sort_and_splice_var_accesses (tree var)
1495 int i, j, access_count;
1496 struct access *res, **prev_acc_ptr = &res;
1497 VEC (access_p, heap) *access_vec;
1499 HOST_WIDE_INT low = -1, high = 0;
1501 access_vec = get_base_access_vector (var);
1504 access_count = VEC_length (access_p, access_vec);
1506 /* Sort by <OFFSET, SIZE>. */
1507 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
1508 compare_access_positions);
1511 while (i < access_count)
1513 struct access *access = VEC_index (access_p, access_vec, i);
1514 bool grp_write = access->write;
1515 bool grp_read = !access->write;
1516 bool grp_assignment_read = access->grp_assignment_read;
1517 bool multiple_reads = false;
1518 bool total_scalarization = access->total_scalarization;
1519 bool grp_partial_lhs = access->grp_partial_lhs;
1520 bool first_scalar = is_gimple_reg_type (access->type);
1521 bool unscalarizable_region = access->grp_unscalarizable_region;
1523 if (first || access->offset >= high)
1526 low = access->offset;
1527 high = access->offset + access->size;
1529 else if (access->offset > low && access->offset + access->size > high)
1532 gcc_assert (access->offset >= low
1533 && access->offset + access->size <= high);
1536 while (j < access_count)
1538 struct access *ac2 = VEC_index (access_p, access_vec, j);
1539 if (ac2->offset != access->offset || ac2->size != access->size)
1546 multiple_reads = true;
1550 grp_assignment_read |= ac2->grp_assignment_read;
1551 grp_partial_lhs |= ac2->grp_partial_lhs;
1552 unscalarizable_region |= ac2->grp_unscalarizable_region;
1553 total_scalarization |= ac2->total_scalarization;
1554 relink_to_new_repr (access, ac2);
1556 /* If there are both aggregate-type and scalar-type accesses with
1557 this combination of size and offset, the comparison function
1558 should have put the scalars first. */
1559 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1560 ac2->group_representative = access;
1566 access->group_representative = access;
1567 access->grp_write = grp_write;
1568 access->grp_read = grp_read;
1569 access->grp_assignment_read = grp_assignment_read;
1570 access->grp_hint = multiple_reads || total_scalarization;
1571 access->grp_partial_lhs = grp_partial_lhs;
1572 access->grp_unscalarizable_region = unscalarizable_region;
1573 if (access->first_link)
1574 add_access_to_work_queue (access);
1576 *prev_acc_ptr = access;
1577 prev_acc_ptr = &access->next_grp;
1580 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1584 /* Create a variable for the given ACCESS which determines the type, name and a
1585 few other properties. Return the variable declaration and store it also to
1586 ACCESS->replacement. */
1589 create_access_replacement (struct access *access, bool rename)
1593 repl = create_tmp_var (access->type, "SR");
1595 add_referenced_var (repl);
1597 mark_sym_for_renaming (repl);
1599 if (!access->grp_partial_lhs
1600 && (TREE_CODE (access->type) == COMPLEX_TYPE
1601 || TREE_CODE (access->type) == VECTOR_TYPE))
1602 DECL_GIMPLE_REG_P (repl) = 1;
1604 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1605 DECL_ARTIFICIAL (repl) = 1;
1606 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1608 if (DECL_NAME (access->base)
1609 && !DECL_IGNORED_P (access->base)
1610 && !DECL_ARTIFICIAL (access->base))
1612 char *pretty_name = make_fancy_name (access->expr);
1613 tree debug_expr = unshare_expr (access->expr), d;
1615 DECL_NAME (repl) = get_identifier (pretty_name);
1616 obstack_free (&name_obstack, pretty_name);
1618 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1619 as DECL_DEBUG_EXPR isn't considered when looking for still
1620 used SSA_NAMEs and thus they could be freed. All debug info
1621 generation cares is whether something is constant or variable
1622 and that get_ref_base_and_extent works properly on the
1624 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1625 switch (TREE_CODE (d))
1628 case ARRAY_RANGE_REF:
1629 if (TREE_OPERAND (d, 1)
1630 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1631 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1632 if (TREE_OPERAND (d, 3)
1633 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1634 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1637 if (TREE_OPERAND (d, 2)
1638 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1639 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1644 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1645 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1646 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1649 TREE_NO_WARNING (repl) = 1;
1653 fprintf (dump_file, "Created a replacement for ");
1654 print_generic_expr (dump_file, access->base, 0);
1655 fprintf (dump_file, " offset: %u, size: %u: ",
1656 (unsigned) access->offset, (unsigned) access->size);
1657 print_generic_expr (dump_file, repl, 0);
1658 fprintf (dump_file, "\n");
1660 sra_stats.replacements++;
1665 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1668 get_access_replacement (struct access *access)
1670 gcc_assert (access->grp_to_be_replaced);
1672 if (!access->replacement_decl)
1673 access->replacement_decl = create_access_replacement (access, true);
1674 return access->replacement_decl;
1677 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1678 not mark it for renaming. */
1681 get_unrenamed_access_replacement (struct access *access)
1683 gcc_assert (!access->grp_to_be_replaced);
1685 if (!access->replacement_decl)
1686 access->replacement_decl = create_access_replacement (access, false);
1687 return access->replacement_decl;
1691 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1692 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1693 to it is not "within" the root. */
1696 build_access_subtree (struct access **access)
1698 struct access *root = *access, *last_child = NULL;
1699 HOST_WIDE_INT limit = root->offset + root->size;
1701 *access = (*access)->next_grp;
1702 while (*access && (*access)->offset + (*access)->size <= limit)
1705 root->first_child = *access;
1707 last_child->next_sibling = *access;
1708 last_child = *access;
1710 build_access_subtree (access);
1714 /* Build a tree of access representatives, ACCESS is the pointer to the first
1715 one, others are linked in a list by the next_grp field. Decide about scalar
1716 replacements on the way, return true iff any are to be created. */
1719 build_access_trees (struct access *access)
1723 struct access *root = access;
1725 build_access_subtree (&access);
1726 root->next_grp = access;
1730 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1734 expr_with_var_bounded_array_refs_p (tree expr)
1736 while (handled_component_p (expr))
1738 if (TREE_CODE (expr) == ARRAY_REF
1739 && !host_integerp (array_ref_low_bound (expr), 0))
1741 expr = TREE_OPERAND (expr, 0);
1746 enum mark_read_status { SRA_MR_NOT_READ, SRA_MR_READ, SRA_MR_ASSIGN_READ};
1748 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1749 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1750 sorts of access flags appropriately along the way, notably always set
1751 grp_read and grp_assign_read according to MARK_READ and grp_write when
1752 MARK_WRITE is true. */
1755 analyze_access_subtree (struct access *root, bool allow_replacements,
1756 enum mark_read_status mark_read, bool mark_write)
1758 struct access *child;
1759 HOST_WIDE_INT limit = root->offset + root->size;
1760 HOST_WIDE_INT covered_to = root->offset;
1761 bool scalar = is_gimple_reg_type (root->type);
1762 bool hole = false, sth_created = false;
1763 bool direct_read = root->grp_read;
1765 if (mark_read == SRA_MR_ASSIGN_READ)
1768 root->grp_assignment_read = 1;
1770 if (mark_read == SRA_MR_READ)
1772 else if (root->grp_assignment_read)
1773 mark_read = SRA_MR_ASSIGN_READ;
1774 else if (root->grp_read)
1775 mark_read = SRA_MR_READ;
1778 root->grp_write = true;
1779 else if (root->grp_write)
1782 if (root->grp_unscalarizable_region)
1783 allow_replacements = false;
1785 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
1786 allow_replacements = false;
1788 for (child = root->first_child; child; child = child->next_sibling)
1790 if (!hole && child->offset < covered_to)
1793 covered_to += child->size;
1795 sth_created |= analyze_access_subtree (child, allow_replacements,
1796 mark_read, mark_write);
1798 root->grp_unscalarized_data |= child->grp_unscalarized_data;
1799 hole |= !child->grp_covered;
1802 if (allow_replacements && scalar && !root->first_child
1804 || (root->grp_write && (direct_read || root->grp_assignment_read)))
1805 /* We must not ICE later on when trying to build an access to the
1806 original data within the aggregate even when it is impossible to do in
1807 a defined way like in the PR 42703 testcase. Therefore we check
1808 pre-emptively here that we will be able to do that. */
1809 && build_ref_for_offset (NULL, TREE_TYPE (root->base), root->offset,
1812 if (dump_file && (dump_flags & TDF_DETAILS))
1814 fprintf (dump_file, "Marking ");
1815 print_generic_expr (dump_file, root->base, 0);
1816 fprintf (dump_file, " offset: %u, size: %u: ",
1817 (unsigned) root->offset, (unsigned) root->size);
1818 fprintf (dump_file, " to be replaced.\n");
1821 root->grp_to_be_replaced = 1;
1825 else if (covered_to < limit)
1828 if (sth_created && !hole)
1830 root->grp_covered = 1;
1833 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
1834 root->grp_unscalarized_data = 1; /* not covered and written to */
1840 /* Analyze all access trees linked by next_grp by the means of
1841 analyze_access_subtree. */
1843 analyze_access_trees (struct access *access)
1849 if (analyze_access_subtree (access, true, SRA_MR_NOT_READ, false))
1851 access = access->next_grp;
1857 /* Return true iff a potential new child of LACC at offset OFFSET and with size
1858 SIZE would conflict with an already existing one. If exactly such a child
1859 already exists in LACC, store a pointer to it in EXACT_MATCH. */
1862 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
1863 HOST_WIDE_INT size, struct access **exact_match)
1865 struct access *child;
1867 for (child = lacc->first_child; child; child = child->next_sibling)
1869 if (child->offset == norm_offset && child->size == size)
1871 *exact_match = child;
1875 if (child->offset < norm_offset + size
1876 && child->offset + child->size > norm_offset)
1883 /* Create a new child access of PARENT, with all properties just like MODEL
1884 except for its offset and with its grp_write false and grp_read true.
1885 Return the new access or NULL if it cannot be created. Note that this access
1886 is created long after all splicing and sorting, it's not located in any
1887 access vector and is automatically a representative of its group. */
1889 static struct access *
1890 create_artificial_child_access (struct access *parent, struct access *model,
1891 HOST_WIDE_INT new_offset)
1893 struct access *access;
1894 struct access **child;
1895 tree expr = parent->base;;
1897 gcc_assert (!model->grp_unscalarizable_region);
1899 if (!build_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
1900 model->type, false))
1903 access = (struct access *) pool_alloc (access_pool);
1904 memset (access, 0, sizeof (struct access));
1905 access->base = parent->base;
1906 access->expr = expr;
1907 access->offset = new_offset;
1908 access->size = model->size;
1909 access->type = model->type;
1910 access->grp_write = true;
1911 access->grp_read = false;
1913 child = &parent->first_child;
1914 while (*child && (*child)->offset < new_offset)
1915 child = &(*child)->next_sibling;
1917 access->next_sibling = *child;
1924 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
1925 true if any new subaccess was created. Additionally, if RACC is a scalar
1926 access but LACC is not, change the type of the latter, if possible. */
1929 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
1931 struct access *rchild;
1932 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
1935 if (is_gimple_reg_type (lacc->type)
1936 || lacc->grp_unscalarizable_region
1937 || racc->grp_unscalarizable_region)
1940 if (!lacc->first_child && !racc->first_child
1941 && is_gimple_reg_type (racc->type))
1943 tree t = lacc->base;
1945 if (build_ref_for_offset (&t, TREE_TYPE (t), lacc->offset, racc->type,
1949 lacc->type = racc->type;
1954 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
1956 struct access *new_acc = NULL;
1957 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
1959 if (rchild->grp_unscalarizable_region)
1962 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
1967 rchild->grp_hint = 1;
1968 new_acc->grp_hint |= new_acc->grp_read;
1969 if (rchild->first_child)
1970 ret |= propagate_subaccesses_across_link (new_acc, rchild);
1975 /* If a (part of) a union field is on the RHS of an assignment, it can
1976 have sub-accesses which do not make sense on the LHS (PR 40351).
1977 Check that this is not the case. */
1978 if (!build_ref_for_offset (NULL, TREE_TYPE (lacc->base), norm_offset,
1979 rchild->type, false))
1982 rchild->grp_hint = 1;
1983 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
1987 if (racc->first_child)
1988 propagate_subaccesses_across_link (new_acc, rchild);
1995 /* Propagate all subaccesses across assignment links. */
1998 propagate_all_subaccesses (void)
2000 while (work_queue_head)
2002 struct access *racc = pop_access_from_work_queue ();
2003 struct assign_link *link;
2005 gcc_assert (racc->first_link);
2007 for (link = racc->first_link; link; link = link->next)
2009 struct access *lacc = link->lacc;
2011 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2013 lacc = lacc->group_representative;
2014 if (propagate_subaccesses_across_link (lacc, racc)
2015 && lacc->first_link)
2016 add_access_to_work_queue (lacc);
2021 /* Go through all accesses collected throughout the (intraprocedural) analysis
2022 stage, exclude overlapping ones, identify representatives and build trees
2023 out of them, making decisions about scalarization on the way. Return true
2024 iff there are any to-be-scalarized variables after this stage. */
2027 analyze_all_variable_accesses (void)
2030 bitmap tmp = BITMAP_ALLOC (NULL);
2032 unsigned i, max_total_scalarization_size;
2034 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2035 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2037 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2038 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2039 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2041 tree var = referenced_var (i);
2043 if (TREE_CODE (var) == VAR_DECL
2044 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2045 <= max_total_scalarization_size)
2046 && type_consists_of_records_p (TREE_TYPE (var)))
2048 completely_scalarize_record (var, var, 0);
2049 if (dump_file && (dump_flags & TDF_DETAILS))
2051 fprintf (dump_file, "Will attempt to totally scalarize ");
2052 print_generic_expr (dump_file, var, 0);
2053 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2058 bitmap_copy (tmp, candidate_bitmap);
2059 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2061 tree var = referenced_var (i);
2062 struct access *access;
2064 access = sort_and_splice_var_accesses (var);
2066 build_access_trees (access);
2068 disqualify_candidate (var,
2069 "No or inhibitingly overlapping accesses.");
2072 propagate_all_subaccesses ();
2074 bitmap_copy (tmp, candidate_bitmap);
2075 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2077 tree var = referenced_var (i);
2078 struct access *access = get_first_repr_for_decl (var);
2080 if (analyze_access_trees (access))
2083 if (dump_file && (dump_flags & TDF_DETAILS))
2085 fprintf (dump_file, "\nAccess trees for ");
2086 print_generic_expr (dump_file, var, 0);
2087 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2088 dump_access_tree (dump_file, access);
2089 fprintf (dump_file, "\n");
2093 disqualify_candidate (var, "No scalar replacements to be created.");
2100 statistics_counter_event (cfun, "Scalarized aggregates", res);
2107 /* Return true iff a reference statement into aggregate AGG can be built for
2108 every single to-be-replaced accesses that is a child of ACCESS, its sibling
2109 or a child of its sibling. TOP_OFFSET is the offset from the processed
2110 access subtree that has to be subtracted from offset of each access. */
2113 ref_expr_for_all_replacements_p (struct access *access, tree agg,
2114 HOST_WIDE_INT top_offset)
2118 if (access->grp_to_be_replaced
2119 && !build_ref_for_offset (NULL, TREE_TYPE (agg),
2120 access->offset - top_offset,
2121 access->type, false))
2124 if (access->first_child
2125 && !ref_expr_for_all_replacements_p (access->first_child, agg,
2129 access = access->next_sibling;
2136 /* Generate statements copying scalar replacements of accesses within a subtree
2137 into or out of AGG. ACCESS is the first child of the root of the subtree to
2138 be processed. AGG is an aggregate type expression (can be a declaration but
2139 does not have to be, it can for example also be an indirect_ref).
2140 TOP_OFFSET is the offset of the processed subtree which has to be subtracted
2141 from offsets of individual accesses to get corresponding offsets for AGG.
2142 If CHUNK_SIZE is non-null, copy only replacements in the interval
2143 <start_offset, start_offset + chunk_size>, otherwise copy all. GSI is a
2144 statement iterator used to place the new statements. WRITE should be true
2145 when the statements should write from AGG to the replacement and false if
2146 vice versa. if INSERT_AFTER is true, new statements will be added after the
2147 current statement in GSI, they will be added before the statement
2151 generate_subtree_copies (struct access *access, tree agg,
2152 HOST_WIDE_INT top_offset,
2153 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2154 gimple_stmt_iterator *gsi, bool write,
2161 if (chunk_size && access->offset >= start_offset + chunk_size)
2164 if (access->grp_to_be_replaced
2166 || access->offset + access->size > start_offset))
2168 tree repl = get_access_replacement (access);
2172 ref_found = build_ref_for_offset (&expr, TREE_TYPE (agg),
2173 access->offset - top_offset,
2174 access->type, false);
2175 gcc_assert (ref_found);
2179 if (access->grp_partial_lhs)
2180 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2182 insert_after ? GSI_NEW_STMT
2184 stmt = gimple_build_assign (repl, expr);
2188 TREE_NO_WARNING (repl) = 1;
2189 if (access->grp_partial_lhs)
2190 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2192 insert_after ? GSI_NEW_STMT
2194 stmt = gimple_build_assign (expr, repl);
2198 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2200 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2202 sra_stats.subtree_copies++;
2205 if (access->first_child)
2206 generate_subtree_copies (access->first_child, agg, top_offset,
2207 start_offset, chunk_size, gsi,
2208 write, insert_after);
2210 access = access->next_sibling;
2215 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2216 the root of the subtree to be processed. GSI is the statement iterator used
2217 for inserting statements which are added after the current statement if
2218 INSERT_AFTER is true or before it otherwise. */
2221 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2225 struct access *child;
2227 if (access->grp_to_be_replaced)
2231 stmt = gimple_build_assign (get_access_replacement (access),
2232 fold_convert (access->type,
2233 integer_zero_node));
2235 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2237 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2241 for (child = access->first_child; child; child = child->next_sibling)
2242 init_subtree_with_zero (child, gsi, insert_after);
2245 /* Search for an access representative for the given expression EXPR and
2246 return it or NULL if it cannot be found. */
2248 static struct access *
2249 get_access_for_expr (tree expr)
2251 HOST_WIDE_INT offset, size, max_size;
2254 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2255 a different size than the size of its argument and we need the latter
2257 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2258 expr = TREE_OPERAND (expr, 0);
2260 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2261 if (max_size == -1 || !DECL_P (base))
2264 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2267 return get_var_base_offset_size_access (base, offset, max_size);
2270 /* Replace the expression EXPR with a scalar replacement if there is one and
2271 generate other statements to do type conversion or subtree copying if
2272 necessary. GSI is used to place newly created statements, WRITE is true if
2273 the expression is being written to (it is on a LHS of a statement or output
2274 in an assembly statement). */
2277 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2279 struct access *access;
2282 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2285 expr = &TREE_OPERAND (*expr, 0);
2290 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2291 expr = &TREE_OPERAND (*expr, 0);
2292 access = get_access_for_expr (*expr);
2295 type = TREE_TYPE (*expr);
2297 if (access->grp_to_be_replaced)
2299 tree repl = get_access_replacement (access);
2300 /* If we replace a non-register typed access simply use the original
2301 access expression to extract the scalar component afterwards.
2302 This happens if scalarizing a function return value or parameter
2303 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2304 gcc.c-torture/compile/20011217-1.c.
2306 We also want to use this when accessing a complex or vector which can
2307 be accessed as a different type too, potentially creating a need for
2308 type conversion (see PR42196) and when scalarized unions are involved
2309 in assembler statements (see PR42398). */
2310 if (!useless_type_conversion_p (type, access->type))
2312 tree ref = access->base;
2315 ok = build_ref_for_offset (&ref, TREE_TYPE (ref),
2316 access->offset, access->type, false);
2323 if (access->grp_partial_lhs)
2324 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2325 false, GSI_NEW_STMT);
2326 stmt = gimple_build_assign (repl, ref);
2327 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2333 if (access->grp_partial_lhs)
2334 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2335 true, GSI_SAME_STMT);
2336 stmt = gimple_build_assign (ref, repl);
2337 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2345 if (access->first_child)
2347 HOST_WIDE_INT start_offset, chunk_size;
2349 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2350 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2352 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2353 start_offset = access->offset
2354 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2357 start_offset = chunk_size = 0;
2359 generate_subtree_copies (access->first_child, access->base, 0,
2360 start_offset, chunk_size, gsi, write, write);
2365 /* Where scalar replacements of the RHS have been written to when a replacement
2366 of a LHS of an assigments cannot be direclty loaded from a replacement of
2368 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2369 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2370 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2372 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2373 base aggregate if there are unscalarized data or directly to LHS
2376 static enum unscalarized_data_handling
2377 handle_unscalarized_data_in_subtree (struct access *top_racc, tree lhs,
2378 gimple_stmt_iterator *gsi)
2380 if (top_racc->grp_unscalarized_data)
2382 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2384 return SRA_UDH_RIGHT;
2388 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2389 0, 0, gsi, false, false);
2390 return SRA_UDH_LEFT;
2395 /* Try to generate statements to load all sub-replacements in an access
2396 (sub)tree (LACC is the first child) from scalar replacements in the TOP_RACC
2397 (sub)tree. If that is not possible, refresh the TOP_RACC base aggregate and
2398 load the accesses from it. LEFT_OFFSET is the offset of the left whole
2399 subtree being copied, RIGHT_OFFSET is the same thing for the right subtree.
2400 GSI is stmt iterator used for statement insertions. *REFRESHED is true iff
2401 the rhs top aggregate has already been refreshed by contents of its scalar
2402 reductions and is set to true if this function has to do it. */
2405 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2406 HOST_WIDE_INT left_offset,
2407 HOST_WIDE_INT right_offset,
2408 gimple_stmt_iterator *old_gsi,
2409 gimple_stmt_iterator *new_gsi,
2410 enum unscalarized_data_handling *refreshed,
2413 location_t loc = EXPR_LOCATION (lacc->expr);
2416 if (lacc->grp_to_be_replaced)
2418 struct access *racc;
2419 HOST_WIDE_INT offset = lacc->offset - left_offset + right_offset;
2423 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2424 if (racc && racc->grp_to_be_replaced)
2426 rhs = get_access_replacement (racc);
2427 if (!useless_type_conversion_p (lacc->type, racc->type))
2428 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2432 /* No suitable access on the right hand side, need to load from
2433 the aggregate. See if we have to update it first... */
2434 if (*refreshed == SRA_UDH_NONE)
2435 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2438 if (*refreshed == SRA_UDH_LEFT)
2443 repl_found = build_ref_for_offset (&rhs, TREE_TYPE (rhs),
2444 lacc->offset, lacc->type,
2446 gcc_assert (repl_found);
2452 rhs = top_racc->base;
2453 repl_found = build_ref_for_offset (&rhs,
2454 TREE_TYPE (top_racc->base),
2455 offset, lacc->type, false);
2456 gcc_assert (repl_found);
2460 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2461 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2463 sra_stats.subreplacements++;
2465 else if (*refreshed == SRA_UDH_NONE
2466 && lacc->grp_read && !lacc->grp_covered)
2467 *refreshed = handle_unscalarized_data_in_subtree (top_racc, lhs,
2470 if (lacc->first_child)
2471 load_assign_lhs_subreplacements (lacc->first_child, top_racc,
2472 left_offset, right_offset,
2473 old_gsi, new_gsi, refreshed, lhs);
2474 lacc = lacc->next_sibling;
2479 /* Result code for SRA assignment modification. */
2480 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2481 SRA_AM_MODIFIED, /* stmt changed but not
2483 SRA_AM_REMOVED }; /* stmt eliminated */
2485 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2486 to the assignment and GSI is the statement iterator pointing at it. Returns
2487 the same values as sra_modify_assign. */
2489 static enum assignment_mod_result
2490 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2492 tree lhs = gimple_assign_lhs (*stmt);
2495 acc = get_access_for_expr (lhs);
2499 if (VEC_length (constructor_elt,
2500 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2502 /* I have never seen this code path trigger but if it can happen the
2503 following should handle it gracefully. */
2504 if (access_has_children_p (acc))
2505 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2507 return SRA_AM_MODIFIED;
2510 if (acc->grp_covered)
2512 init_subtree_with_zero (acc, gsi, false);
2513 unlink_stmt_vdef (*stmt);
2514 gsi_remove (gsi, true);
2515 return SRA_AM_REMOVED;
2519 init_subtree_with_zero (acc, gsi, true);
2520 return SRA_AM_MODIFIED;
2524 /* Create a new suitable default definition SSA_NAME and replace all uses of
2528 replace_uses_with_default_def_ssa_name (tree ssa)
2530 tree repl, decl = SSA_NAME_VAR (ssa);
2531 if (TREE_CODE (decl) == PARM_DECL)
2533 tree tmp = create_tmp_var (TREE_TYPE (decl), "SR");
2534 if (TREE_CODE (TREE_TYPE (tmp)) == COMPLEX_TYPE
2535 || TREE_CODE (TREE_TYPE (tmp)) == VECTOR_TYPE)
2536 DECL_GIMPLE_REG_P (tmp) = 1;
2539 add_referenced_var (tmp);
2540 repl = make_ssa_name (tmp, gimple_build_nop ());
2541 set_default_def (tmp, repl);
2545 repl = gimple_default_def (cfun, decl);
2548 repl = make_ssa_name (decl, gimple_build_nop ());
2549 set_default_def (decl, repl);
2553 replace_uses_by (ssa, repl);
2556 repl = gimple_default_def (cfun, decl);
2559 repl = make_ssa_name (decl, gimple_build_nop ());
2560 set_default_def (decl, repl);
2563 replace_uses_by (ssa, repl);
2566 /* Examine both sides of the assignment statement pointed to by STMT, replace
2567 them with a scalare replacement if there is one and generate copying of
2568 replacements if scalarized aggregates have been used in the assignment. GSI
2569 is used to hold generated statements for type conversions and subtree
2572 static enum assignment_mod_result
2573 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2575 struct access *lacc, *racc;
2577 bool modify_this_stmt = false;
2578 bool force_gimple_rhs = false;
2579 location_t loc = gimple_location (*stmt);
2580 gimple_stmt_iterator orig_gsi = *gsi;
2582 if (!gimple_assign_single_p (*stmt))
2584 lhs = gimple_assign_lhs (*stmt);
2585 rhs = gimple_assign_rhs1 (*stmt);
2587 if (TREE_CODE (rhs) == CONSTRUCTOR)
2588 return sra_modify_constructor_assign (stmt, gsi);
2590 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2591 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2592 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2594 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2596 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2598 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2601 lacc = get_access_for_expr (lhs);
2602 racc = get_access_for_expr (rhs);
2606 if (lacc && lacc->grp_to_be_replaced)
2608 lhs = get_access_replacement (lacc);
2609 gimple_assign_set_lhs (*stmt, lhs);
2610 modify_this_stmt = true;
2611 if (lacc->grp_partial_lhs)
2612 force_gimple_rhs = true;
2616 if (racc && racc->grp_to_be_replaced)
2618 rhs = get_access_replacement (racc);
2619 modify_this_stmt = true;
2620 if (racc->grp_partial_lhs)
2621 force_gimple_rhs = true;
2625 if (modify_this_stmt)
2627 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2629 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2630 ??? This should move to fold_stmt which we simply should
2631 call after building a VIEW_CONVERT_EXPR here. */
2632 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2633 && !access_has_children_p (lacc))
2636 if (build_ref_for_offset (&expr, TREE_TYPE (lhs), 0,
2637 TREE_TYPE (rhs), false))
2640 gimple_assign_set_lhs (*stmt, expr);
2643 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2644 && !access_has_children_p (racc))
2647 if (build_ref_for_offset (&expr, TREE_TYPE (rhs), 0,
2648 TREE_TYPE (lhs), false))
2651 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2653 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs), rhs);
2654 if (is_gimple_reg_type (TREE_TYPE (lhs))
2655 && TREE_CODE (lhs) != SSA_NAME)
2656 force_gimple_rhs = true;
2661 /* From this point on, the function deals with assignments in between
2662 aggregates when at least one has scalar reductions of some of its
2663 components. There are three possible scenarios: Both the LHS and RHS have
2664 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2666 In the first case, we would like to load the LHS components from RHS
2667 components whenever possible. If that is not possible, we would like to
2668 read it directly from the RHS (after updating it by storing in it its own
2669 components). If there are some necessary unscalarized data in the LHS,
2670 those will be loaded by the original assignment too. If neither of these
2671 cases happen, the original statement can be removed. Most of this is done
2672 by load_assign_lhs_subreplacements.
2674 In the second case, we would like to store all RHS scalarized components
2675 directly into LHS and if they cover the aggregate completely, remove the
2676 statement too. In the third case, we want the LHS components to be loaded
2677 directly from the RHS (DSE will remove the original statement if it
2680 This is a bit complex but manageable when types match and when unions do
2681 not cause confusion in a way that we cannot really load a component of LHS
2682 from the RHS or vice versa (the access representing this level can have
2683 subaccesses that are accessible only through a different union field at a
2684 higher level - different from the one used in the examined expression).
2687 Therefore, I specially handle a fourth case, happening when there is a
2688 specific type cast or it is impossible to locate a scalarized subaccess on
2689 the other side of the expression. If that happens, I simply "refresh" the
2690 RHS by storing in it is scalarized components leave the original statement
2691 there to do the copying and then load the scalar replacements of the LHS.
2692 This is what the first branch does. */
2694 if (gimple_has_volatile_ops (*stmt)
2695 || contains_view_convert_expr_p (rhs)
2696 || contains_view_convert_expr_p (lhs)
2697 || (access_has_children_p (racc)
2698 && !ref_expr_for_all_replacements_p (racc, lhs, racc->offset))
2699 || (access_has_children_p (lacc)
2700 && !ref_expr_for_all_replacements_p (lacc, rhs, lacc->offset)))
2702 if (access_has_children_p (racc))
2703 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2705 if (access_has_children_p (lacc))
2706 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2708 sra_stats.separate_lhs_rhs_handling++;
2712 if (access_has_children_p (lacc) && access_has_children_p (racc))
2714 gimple_stmt_iterator orig_gsi = *gsi;
2715 enum unscalarized_data_handling refreshed;
2717 if (lacc->grp_read && !lacc->grp_covered)
2718 refreshed = handle_unscalarized_data_in_subtree (racc, lhs, gsi);
2720 refreshed = SRA_UDH_NONE;
2722 load_assign_lhs_subreplacements (lacc->first_child, racc,
2723 lacc->offset, racc->offset,
2724 &orig_gsi, gsi, &refreshed, lhs);
2725 if (refreshed != SRA_UDH_RIGHT)
2727 if (*stmt == gsi_stmt (*gsi))
2730 unlink_stmt_vdef (*stmt);
2731 gsi_remove (&orig_gsi, true);
2732 sra_stats.deleted++;
2733 return SRA_AM_REMOVED;
2740 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2742 if (racc->first_child)
2743 generate_subtree_copies (racc->first_child, lhs,
2744 racc->offset, 0, 0, gsi,
2746 gcc_assert (*stmt == gsi_stmt (*gsi));
2747 if (TREE_CODE (lhs) == SSA_NAME)
2748 replace_uses_with_default_def_ssa_name (lhs);
2750 unlink_stmt_vdef (*stmt);
2751 gsi_remove (gsi, true);
2752 sra_stats.deleted++;
2753 return SRA_AM_REMOVED;
2755 else if (racc->first_child)
2756 generate_subtree_copies (racc->first_child, lhs,
2757 racc->offset, 0, 0, gsi, false, true);
2759 if (access_has_children_p (lacc))
2760 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2761 0, 0, gsi, true, true);
2765 /* This gimplification must be done after generate_subtree_copies, lest we
2766 insert the subtree copies in the middle of the gimplified sequence. */
2767 if (force_gimple_rhs)
2768 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2769 true, GSI_SAME_STMT);
2770 if (gimple_assign_rhs1 (*stmt) != rhs)
2772 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2773 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2776 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2779 /* Traverse the function body and all modifications as decided in
2780 analyze_all_variable_accesses. */
2783 sra_modify_function_body (void)
2789 gimple_stmt_iterator gsi = gsi_start_bb (bb);
2790 while (!gsi_end_p (gsi))
2792 gimple stmt = gsi_stmt (gsi);
2793 enum assignment_mod_result assign_result;
2794 bool modified = false, deleted = false;
2798 switch (gimple_code (stmt))
2801 t = gimple_return_retval_ptr (stmt);
2802 if (*t != NULL_TREE)
2803 modified |= sra_modify_expr (t, &gsi, false);
2807 assign_result = sra_modify_assign (&stmt, &gsi);
2808 modified |= assign_result == SRA_AM_MODIFIED;
2809 deleted = assign_result == SRA_AM_REMOVED;
2813 /* Operands must be processed before the lhs. */
2814 for (i = 0; i < gimple_call_num_args (stmt); i++)
2816 t = gimple_call_arg_ptr (stmt, i);
2817 modified |= sra_modify_expr (t, &gsi, false);
2820 if (gimple_call_lhs (stmt))
2822 t = gimple_call_lhs_ptr (stmt);
2823 modified |= sra_modify_expr (t, &gsi, true);
2828 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
2830 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
2831 modified |= sra_modify_expr (t, &gsi, false);
2833 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
2835 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
2836 modified |= sra_modify_expr (t, &gsi, true);
2847 maybe_clean_eh_stmt (stmt);
2855 /* Generate statements initializing scalar replacements of parts of function
2859 initialize_parameter_reductions (void)
2861 gimple_stmt_iterator gsi;
2862 gimple_seq seq = NULL;
2865 for (parm = DECL_ARGUMENTS (current_function_decl);
2867 parm = TREE_CHAIN (parm))
2869 VEC (access_p, heap) *access_vec;
2870 struct access *access;
2872 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
2874 access_vec = get_base_access_vector (parm);
2880 seq = gimple_seq_alloc ();
2881 gsi = gsi_start (seq);
2884 for (access = VEC_index (access_p, access_vec, 0);
2886 access = access->next_grp)
2887 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true);
2891 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
2894 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
2895 it reveals there are components of some aggregates to be scalarized, it runs
2896 the required transformations. */
2898 perform_intra_sra (void)
2903 if (!find_var_candidates ())
2906 if (!scan_function ())
2909 if (!analyze_all_variable_accesses ())
2912 sra_modify_function_body ();
2913 initialize_parameter_reductions ();
2915 statistics_counter_event (cfun, "Scalar replacements created",
2916 sra_stats.replacements);
2917 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
2918 statistics_counter_event (cfun, "Subtree copy stmts",
2919 sra_stats.subtree_copies);
2920 statistics_counter_event (cfun, "Subreplacement stmts",
2921 sra_stats.subreplacements);
2922 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
2923 statistics_counter_event (cfun, "Separate LHS and RHS handling",
2924 sra_stats.separate_lhs_rhs_handling);
2926 ret = TODO_update_ssa;
2929 sra_deinitialize ();
2933 /* Perform early intraprocedural SRA. */
2935 early_intra_sra (void)
2937 sra_mode = SRA_MODE_EARLY_INTRA;
2938 return perform_intra_sra ();
2941 /* Perform "late" intraprocedural SRA. */
2943 late_intra_sra (void)
2945 sra_mode = SRA_MODE_INTRA;
2946 return perform_intra_sra ();
2951 gate_intra_sra (void)
2953 return flag_tree_sra != 0;
2957 struct gimple_opt_pass pass_sra_early =
2962 gate_intra_sra, /* gate */
2963 early_intra_sra, /* execute */
2966 0, /* static_pass_number */
2967 TV_TREE_SRA, /* tv_id */
2968 PROP_cfg | PROP_ssa, /* properties_required */
2969 0, /* properties_provided */
2970 0, /* properties_destroyed */
2971 0, /* todo_flags_start */
2975 | TODO_verify_ssa /* todo_flags_finish */
2979 struct gimple_opt_pass pass_sra =
2984 gate_intra_sra, /* gate */
2985 late_intra_sra, /* execute */
2988 0, /* static_pass_number */
2989 TV_TREE_SRA, /* tv_id */
2990 PROP_cfg | PROP_ssa, /* properties_required */
2991 0, /* properties_provided */
2992 0, /* properties_destroyed */
2993 TODO_update_address_taken, /* todo_flags_start */
2997 | TODO_verify_ssa /* todo_flags_finish */
3002 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3006 is_unused_scalar_param (tree parm)
3009 return (is_gimple_reg (parm)
3010 && (!(name = gimple_default_def (cfun, parm))
3011 || has_zero_uses (name)));
3014 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3015 examine whether there are any direct or otherwise infeasible ones. If so,
3016 return true, otherwise return false. PARM must be a gimple register with a
3017 non-NULL default definition. */
3020 ptr_parm_has_direct_uses (tree parm)
3022 imm_use_iterator ui;
3024 tree name = gimple_default_def (cfun, parm);
3027 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3030 use_operand_p use_p;
3032 if (is_gimple_debug (stmt))
3035 /* Valid uses include dereferences on the lhs and the rhs. */
3036 if (gimple_has_lhs (stmt))
3038 tree lhs = gimple_get_lhs (stmt);
3039 while (handled_component_p (lhs))
3040 lhs = TREE_OPERAND (lhs, 0);
3041 if (INDIRECT_REF_P (lhs)
3042 && TREE_OPERAND (lhs, 0) == name)
3045 if (gimple_assign_single_p (stmt))
3047 tree rhs = gimple_assign_rhs1 (stmt);
3048 while (handled_component_p (rhs))
3049 rhs = TREE_OPERAND (rhs, 0);
3050 if (INDIRECT_REF_P (rhs)
3051 && TREE_OPERAND (rhs, 0) == name)
3054 else if (is_gimple_call (stmt))
3057 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3059 tree arg = gimple_call_arg (stmt, i);
3060 while (handled_component_p (arg))
3061 arg = TREE_OPERAND (arg, 0);
3062 if (INDIRECT_REF_P (arg)
3063 && TREE_OPERAND (arg, 0) == name)
3068 /* If the number of valid uses does not match the number of
3069 uses in this stmt there is an unhandled use. */
3070 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3077 BREAK_FROM_IMM_USE_STMT (ui);
3083 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3084 them in candidate_bitmap. Note that these do not necessarily include
3085 parameter which are unused and thus can be removed. Return true iff any
3086 such candidate has been found. */
3089 find_param_candidates (void)
3095 for (parm = DECL_ARGUMENTS (current_function_decl);
3097 parm = TREE_CHAIN (parm))
3099 tree type = TREE_TYPE (parm);
3103 if (TREE_THIS_VOLATILE (parm)
3104 || TREE_ADDRESSABLE (parm)
3105 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3108 if (is_unused_scalar_param (parm))
3114 if (POINTER_TYPE_P (type))
3116 type = TREE_TYPE (type);
3118 if (TREE_CODE (type) == FUNCTION_TYPE
3119 || TYPE_VOLATILE (type)
3120 || !is_gimple_reg (parm)
3121 || is_va_list_type (type)
3122 || ptr_parm_has_direct_uses (parm))
3125 else if (!AGGREGATE_TYPE_P (type))
3128 if (!COMPLETE_TYPE_P (type)
3129 || !host_integerp (TYPE_SIZE (type), 1)
3130 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3131 || (AGGREGATE_TYPE_P (type)
3132 && type_internals_preclude_sra_p (type)))
3135 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3137 if (dump_file && (dump_flags & TDF_DETAILS))
3139 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3140 print_generic_expr (dump_file, parm, 0);
3141 fprintf (dump_file, "\n");
3145 func_param_count = count;
3149 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3153 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3156 struct access *repr = (struct access *) data;
3158 repr->grp_maybe_modified = 1;
3162 /* Analyze what representatives (in linked lists accessible from
3163 REPRESENTATIVES) can be modified by side effects of statements in the
3164 current function. */
3167 analyze_modified_params (VEC (access_p, heap) *representatives)
3171 for (i = 0; i < func_param_count; i++)
3173 struct access *repr;
3175 for (repr = VEC_index (access_p, representatives, i);
3177 repr = repr->next_grp)
3179 struct access *access;
3183 if (no_accesses_p (repr))
3185 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3186 || repr->grp_maybe_modified)
3189 ao_ref_init (&ar, repr->expr);
3190 visited = BITMAP_ALLOC (NULL);
3191 for (access = repr; access; access = access->next_sibling)
3193 /* All accesses are read ones, otherwise grp_maybe_modified would
3194 be trivially set. */
3195 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3196 mark_maybe_modified, repr, &visited);
3197 if (repr->grp_maybe_modified)
3200 BITMAP_FREE (visited);
3205 /* Propagate distances in bb_dereferences in the opposite direction than the
3206 control flow edges, in each step storing the maximum of the current value
3207 and the minimum of all successors. These steps are repeated until the table
3208 stabilizes. Note that BBs which might terminate the functions (according to
3209 final_bbs bitmap) never updated in this way. */
3212 propagate_dereference_distances (void)
3214 VEC (basic_block, heap) *queue;
3217 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3218 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3221 VEC_quick_push (basic_block, queue, bb);
3225 while (!VEC_empty (basic_block, queue))
3229 bool change = false;
3232 bb = VEC_pop (basic_block, queue);
3235 if (bitmap_bit_p (final_bbs, bb->index))
3238 for (i = 0; i < func_param_count; i++)
3240 int idx = bb->index * func_param_count + i;
3242 HOST_WIDE_INT inh = 0;
3244 FOR_EACH_EDGE (e, ei, bb->succs)
3246 int succ_idx = e->dest->index * func_param_count + i;
3248 if (e->src == EXIT_BLOCK_PTR)
3254 inh = bb_dereferences [succ_idx];
3256 else if (bb_dereferences [succ_idx] < inh)
3257 inh = bb_dereferences [succ_idx];
3260 if (!first && bb_dereferences[idx] < inh)
3262 bb_dereferences[idx] = inh;
3267 if (change && !bitmap_bit_p (final_bbs, bb->index))
3268 FOR_EACH_EDGE (e, ei, bb->preds)
3273 e->src->aux = e->src;
3274 VEC_quick_push (basic_block, queue, e->src);
3278 VEC_free (basic_block, heap, queue);
3281 /* Dump a dereferences TABLE with heading STR to file F. */
3284 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3288 fprintf (dump_file, str);
3289 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3291 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3292 if (bb != EXIT_BLOCK_PTR)
3295 for (i = 0; i < func_param_count; i++)
3297 int idx = bb->index * func_param_count + i;
3298 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3303 fprintf (dump_file, "\n");
3306 /* Determine what (parts of) parameters passed by reference that are not
3307 assigned to are not certainly dereferenced in this function and thus the
3308 dereferencing cannot be safely moved to the caller without potentially
3309 introducing a segfault. Mark such REPRESENTATIVES as
3310 grp_not_necessarilly_dereferenced.
3312 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3313 part is calculated rather than simple booleans are calculated for each
3314 pointer parameter to handle cases when only a fraction of the whole
3315 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3318 The maximum dereference distances for each pointer parameter and BB are
3319 already stored in bb_dereference. This routine simply propagates these
3320 values upwards by propagate_dereference_distances and then compares the
3321 distances of individual parameters in the ENTRY BB to the equivalent
3322 distances of each representative of a (fraction of a) parameter. */
3325 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3329 if (dump_file && (dump_flags & TDF_DETAILS))
3330 dump_dereferences_table (dump_file,
3331 "Dereference table before propagation:\n",
3334 propagate_dereference_distances ();
3336 if (dump_file && (dump_flags & TDF_DETAILS))
3337 dump_dereferences_table (dump_file,
3338 "Dereference table after propagation:\n",
3341 for (i = 0; i < func_param_count; i++)
3343 struct access *repr = VEC_index (access_p, representatives, i);
3344 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3346 if (!repr || no_accesses_p (repr))
3351 if ((repr->offset + repr->size) > bb_dereferences[idx])
3352 repr->grp_not_necessarilly_dereferenced = 1;
3353 repr = repr->next_grp;
3359 /* Return the representative access for the parameter declaration PARM if it is
3360 a scalar passed by reference which is not written to and the pointer value
3361 is not used directly. Thus, if it is legal to dereference it in the caller
3362 and we can rule out modifications through aliases, such parameter should be
3363 turned into one passed by value. Return NULL otherwise. */
3365 static struct access *
3366 unmodified_by_ref_scalar_representative (tree parm)
3368 int i, access_count;
3369 struct access *repr;
3370 VEC (access_p, heap) *access_vec;
3372 access_vec = get_base_access_vector (parm);
3373 gcc_assert (access_vec);
3374 repr = VEC_index (access_p, access_vec, 0);
3377 repr->group_representative = repr;
3379 access_count = VEC_length (access_p, access_vec);
3380 for (i = 1; i < access_count; i++)
3382 struct access *access = VEC_index (access_p, access_vec, i);
3385 access->group_representative = repr;
3386 access->next_sibling = repr->next_sibling;
3387 repr->next_sibling = access;
3391 repr->grp_scalar_ptr = 1;
3395 /* Return true iff this access precludes IPA-SRA of the parameter it is
3399 access_precludes_ipa_sra_p (struct access *access)
3401 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3402 is incompatible assign in a call statement (and possibly even in asm
3403 statements). This can be relaxed by using a new temporary but only for
3404 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3405 intraprocedural SRA we deal with this by keeping the old aggregate around,
3406 something we cannot do in IPA-SRA.) */
3408 && (is_gimple_call (access->stmt)
3409 || gimple_code (access->stmt) == GIMPLE_ASM))
3416 /* Sort collected accesses for parameter PARM, identify representatives for
3417 each accessed region and link them together. Return NULL if there are
3418 different but overlapping accesses, return the special ptr value meaning
3419 there are no accesses for this parameter if that is the case and return the
3420 first representative otherwise. Set *RO_GRP if there is a group of accesses
3421 with only read (i.e. no write) accesses. */
3423 static struct access *
3424 splice_param_accesses (tree parm, bool *ro_grp)
3426 int i, j, access_count, group_count;
3427 int agg_size, total_size = 0;
3428 struct access *access, *res, **prev_acc_ptr = &res;
3429 VEC (access_p, heap) *access_vec;
3431 access_vec = get_base_access_vector (parm);
3433 return &no_accesses_representant;
3434 access_count = VEC_length (access_p, access_vec);
3436 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
3437 compare_access_positions);
3442 while (i < access_count)
3445 access = VEC_index (access_p, access_vec, i);
3446 modification = access->write;
3447 if (access_precludes_ipa_sra_p (access))
3450 /* Access is about to become group representative unless we find some
3451 nasty overlap which would preclude us from breaking this parameter
3455 while (j < access_count)
3457 struct access *ac2 = VEC_index (access_p, access_vec, j);
3458 if (ac2->offset != access->offset)
3460 /* All or nothing law for parameters. */
3461 if (access->offset + access->size > ac2->offset)
3466 else if (ac2->size != access->size)
3469 if (access_precludes_ipa_sra_p (ac2))
3472 modification |= ac2->write;
3473 ac2->group_representative = access;
3474 ac2->next_sibling = access->next_sibling;
3475 access->next_sibling = ac2;
3480 access->grp_maybe_modified = modification;
3483 *prev_acc_ptr = access;
3484 prev_acc_ptr = &access->next_grp;
3485 total_size += access->size;
3489 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3490 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3492 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3493 if (total_size >= agg_size)
3496 gcc_assert (group_count > 0);
3500 /* Decide whether parameters with representative accesses given by REPR should
3501 be reduced into components. */
3504 decide_one_param_reduction (struct access *repr)
3506 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3511 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3512 gcc_assert (cur_parm_size > 0);
3514 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3517 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3522 agg_size = cur_parm_size;
3528 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3529 print_generic_expr (dump_file, parm, 0);
3530 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3531 for (acc = repr; acc; acc = acc->next_grp)
3532 dump_access (dump_file, acc, true);
3536 new_param_count = 0;
3538 for (; repr; repr = repr->next_grp)
3540 gcc_assert (parm == repr->base);
3543 if (!by_ref || (!repr->grp_maybe_modified
3544 && !repr->grp_not_necessarilly_dereferenced))
3545 total_size += repr->size;
3547 total_size += cur_parm_size;
3550 gcc_assert (new_param_count > 0);
3552 if (optimize_function_for_size_p (cfun))
3553 parm_size_limit = cur_parm_size;
3555 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3558 if (total_size < agg_size
3559 && total_size <= parm_size_limit)
3562 fprintf (dump_file, " ....will be split into %i components\n",
3564 return new_param_count;
3570 /* The order of the following enums is important, we need to do extra work for
3571 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3572 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3573 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3575 /* Identify representatives of all accesses to all candidate parameters for
3576 IPA-SRA. Return result based on what representatives have been found. */
3578 static enum ipa_splicing_result
3579 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3581 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3583 struct access *repr;
3585 *representatives = VEC_alloc (access_p, heap, func_param_count);
3587 for (parm = DECL_ARGUMENTS (current_function_decl);
3589 parm = TREE_CHAIN (parm))
3591 if (is_unused_scalar_param (parm))
3593 VEC_quick_push (access_p, *representatives,
3594 &no_accesses_representant);
3595 if (result == NO_GOOD_ACCESS)
3596 result = UNUSED_PARAMS;
3598 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3599 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3600 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3602 repr = unmodified_by_ref_scalar_representative (parm);
3603 VEC_quick_push (access_p, *representatives, repr);
3605 result = UNMODIF_BY_REF_ACCESSES;
3607 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3609 bool ro_grp = false;
3610 repr = splice_param_accesses (parm, &ro_grp);
3611 VEC_quick_push (access_p, *representatives, repr);
3613 if (repr && !no_accesses_p (repr))
3615 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3618 result = UNMODIF_BY_REF_ACCESSES;
3619 else if (result < MODIF_BY_REF_ACCESSES)
3620 result = MODIF_BY_REF_ACCESSES;
3622 else if (result < BY_VAL_ACCESSES)
3623 result = BY_VAL_ACCESSES;
3625 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3626 result = UNUSED_PARAMS;
3629 VEC_quick_push (access_p, *representatives, NULL);
3632 if (result == NO_GOOD_ACCESS)
3634 VEC_free (access_p, heap, *representatives);
3635 *representatives = NULL;
3636 return NO_GOOD_ACCESS;
3642 /* Return the index of BASE in PARMS. Abort if it is not found. */
3645 get_param_index (tree base, VEC(tree, heap) *parms)
3649 len = VEC_length (tree, parms);
3650 for (i = 0; i < len; i++)
3651 if (VEC_index (tree, parms, i) == base)
3656 /* Convert the decisions made at the representative level into compact
3657 parameter adjustments. REPRESENTATIVES are pointers to first
3658 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3659 final number of adjustments. */
3661 static ipa_parm_adjustment_vec
3662 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3663 int adjustments_count)
3665 VEC (tree, heap) *parms;
3666 ipa_parm_adjustment_vec adjustments;
3670 gcc_assert (adjustments_count > 0);
3671 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3672 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3673 parm = DECL_ARGUMENTS (current_function_decl);
3674 for (i = 0; i < func_param_count; i++, parm = TREE_CHAIN (parm))
3676 struct access *repr = VEC_index (access_p, representatives, i);
3678 if (!repr || no_accesses_p (repr))
3680 struct ipa_parm_adjustment *adj;
3682 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3683 memset (adj, 0, sizeof (*adj));
3684 adj->base_index = get_param_index (parm, parms);
3687 adj->copy_param = 1;
3689 adj->remove_param = 1;
3693 struct ipa_parm_adjustment *adj;
3694 int index = get_param_index (parm, parms);
3696 for (; repr; repr = repr->next_grp)
3698 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3699 memset (adj, 0, sizeof (*adj));
3700 gcc_assert (repr->base == parm);
3701 adj->base_index = index;
3702 adj->base = repr->base;
3703 adj->type = repr->type;
3704 adj->offset = repr->offset;
3705 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3706 && (repr->grp_maybe_modified
3707 || repr->grp_not_necessarilly_dereferenced));
3712 VEC_free (tree, heap, parms);
3716 /* Analyze the collected accesses and produce a plan what to do with the
3717 parameters in the form of adjustments, NULL meaning nothing. */
3719 static ipa_parm_adjustment_vec
3720 analyze_all_param_acesses (void)
3722 enum ipa_splicing_result repr_state;
3723 bool proceed = false;
3724 int i, adjustments_count = 0;
3725 VEC (access_p, heap) *representatives;
3726 ipa_parm_adjustment_vec adjustments;
3728 repr_state = splice_all_param_accesses (&representatives);
3729 if (repr_state == NO_GOOD_ACCESS)
3732 /* If there are any parameters passed by reference which are not modified
3733 directly, we need to check whether they can be modified indirectly. */
3734 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3736 analyze_caller_dereference_legality (representatives);
3737 analyze_modified_params (representatives);
3740 for (i = 0; i < func_param_count; i++)
3742 struct access *repr = VEC_index (access_p, representatives, i);
3744 if (repr && !no_accesses_p (repr))
3746 if (repr->grp_scalar_ptr)
3748 adjustments_count++;
3749 if (repr->grp_not_necessarilly_dereferenced
3750 || repr->grp_maybe_modified)
3751 VEC_replace (access_p, representatives, i, NULL);
3755 sra_stats.scalar_by_ref_to_by_val++;
3760 int new_components = decide_one_param_reduction (repr);
3762 if (new_components == 0)
3764 VEC_replace (access_p, representatives, i, NULL);
3765 adjustments_count++;
3769 adjustments_count += new_components;
3770 sra_stats.aggregate_params_reduced++;
3771 sra_stats.param_reductions_created += new_components;
3778 if (no_accesses_p (repr))
3781 sra_stats.deleted_unused_parameters++;
3783 adjustments_count++;
3787 if (!proceed && dump_file)
3788 fprintf (dump_file, "NOT proceeding to change params.\n");
3791 adjustments = turn_representatives_into_adjustments (representatives,
3796 VEC_free (access_p, heap, representatives);
3800 /* If a parameter replacement identified by ADJ does not yet exist in the form
3801 of declaration, create it and record it, otherwise return the previously
3805 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
3808 if (!adj->new_ssa_base)
3810 char *pretty_name = make_fancy_name (adj->base);
3812 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
3813 DECL_NAME (repl) = get_identifier (pretty_name);
3814 obstack_free (&name_obstack, pretty_name);
3817 add_referenced_var (repl);
3818 adj->new_ssa_base = repl;
3821 repl = adj->new_ssa_base;
3825 /* Find the first adjustment for a particular parameter BASE in a vector of
3826 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3829 static struct ipa_parm_adjustment *
3830 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
3834 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3835 for (i = 0; i < len; i++)
3837 struct ipa_parm_adjustment *adj;
3839 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3840 if (!adj->copy_param && adj->base == base)
3847 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3848 removed because its value is not used, replace the SSA_NAME with a one
3849 relating to a created VAR_DECL together all of its uses and return true.
3850 ADJUSTMENTS is a pointer to an adjustments vector. */
3853 replace_removed_params_ssa_names (gimple stmt,
3854 ipa_parm_adjustment_vec adjustments)
3856 struct ipa_parm_adjustment *adj;
3857 tree lhs, decl, repl, name;
3859 if (gimple_code (stmt) == GIMPLE_PHI)
3860 lhs = gimple_phi_result (stmt);
3861 else if (is_gimple_assign (stmt))
3862 lhs = gimple_assign_lhs (stmt);
3863 else if (is_gimple_call (stmt))
3864 lhs = gimple_call_lhs (stmt);
3868 if (TREE_CODE (lhs) != SSA_NAME)
3870 decl = SSA_NAME_VAR (lhs);
3871 if (TREE_CODE (decl) != PARM_DECL)
3874 adj = get_adjustment_for_base (adjustments, decl);
3878 repl = get_replaced_param_substitute (adj);
3879 name = make_ssa_name (repl, stmt);
3883 fprintf (dump_file, "replacing an SSA name of a removed param ");
3884 print_generic_expr (dump_file, lhs, 0);
3885 fprintf (dump_file, " with ");
3886 print_generic_expr (dump_file, name, 0);
3887 fprintf (dump_file, "\n");
3890 if (is_gimple_assign (stmt))
3891 gimple_assign_set_lhs (stmt, name);
3892 else if (is_gimple_call (stmt))
3893 gimple_call_set_lhs (stmt, name);
3895 gimple_phi_set_result (stmt, name);
3897 replace_uses_by (lhs, name);
3901 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
3902 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
3903 specifies whether the function should care about type incompatibility the
3904 current and new expressions. If it is false, the function will leave
3905 incompatibility issues to the caller. Return true iff the expression
3909 sra_ipa_modify_expr (tree *expr, bool convert,
3910 ipa_parm_adjustment_vec adjustments)
3913 struct ipa_parm_adjustment *adj, *cand = NULL;
3914 HOST_WIDE_INT offset, size, max_size;
3917 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3919 if (TREE_CODE (*expr) == BIT_FIELD_REF
3920 || TREE_CODE (*expr) == IMAGPART_EXPR
3921 || TREE_CODE (*expr) == REALPART_EXPR)
3923 expr = &TREE_OPERAND (*expr, 0);
3927 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
3928 if (!base || size == -1 || max_size == -1)
3931 if (INDIRECT_REF_P (base))
3932 base = TREE_OPERAND (base, 0);
3934 base = get_ssa_base_param (base);
3935 if (!base || TREE_CODE (base) != PARM_DECL)
3938 for (i = 0; i < len; i++)
3940 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3942 if (adj->base == base &&
3943 (adj->offset == offset || adj->remove_param))
3949 if (!cand || cand->copy_param || cand->remove_param)
3955 src = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (cand->reduction)),
3957 folded = gimple_fold_indirect_ref (src);
3962 src = cand->reduction;
3964 if (dump_file && (dump_flags & TDF_DETAILS))
3966 fprintf (dump_file, "About to replace expr ");
3967 print_generic_expr (dump_file, *expr, 0);
3968 fprintf (dump_file, " with ");
3969 print_generic_expr (dump_file, src, 0);
3970 fprintf (dump_file, "\n");
3973 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
3975 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
3983 /* If the statement pointed to by STMT_PTR contains any expressions that need
3984 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
3985 potential type incompatibilities (GSI is used to accommodate conversion
3986 statements and must point to the statement). Return true iff the statement
3990 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
3991 ipa_parm_adjustment_vec adjustments)
3993 gimple stmt = *stmt_ptr;
3994 tree *lhs_p, *rhs_p;
3997 if (!gimple_assign_single_p (stmt))
4000 rhs_p = gimple_assign_rhs1_ptr (stmt);
4001 lhs_p = gimple_assign_lhs_ptr (stmt);
4003 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4004 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4007 tree new_rhs = NULL_TREE;
4009 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4011 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4013 /* V_C_Es of constructors can cause trouble (PR 42714). */
4014 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4015 *rhs_p = fold_convert (TREE_TYPE (*lhs_p), integer_zero_node);
4017 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4020 new_rhs = fold_build1_loc (gimple_location (stmt),
4021 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4024 else if (REFERENCE_CLASS_P (*rhs_p)
4025 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4026 && !is_gimple_reg (*lhs_p))
4027 /* This can happen when an assignment in between two single field
4028 structures is turned into an assignment in between two pointers to
4029 scalars (PR 42237). */
4034 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4035 true, GSI_SAME_STMT);
4037 gimple_assign_set_rhs_from_tree (gsi, tmp);
4046 /* Traverse the function body and all modifications as described in
4050 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4056 gimple_stmt_iterator gsi;
4057 bool bb_changed = false;
4059 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4060 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4062 gsi = gsi_start_bb (bb);
4063 while (!gsi_end_p (gsi))
4065 gimple stmt = gsi_stmt (gsi);
4066 bool modified = false;
4070 switch (gimple_code (stmt))
4073 t = gimple_return_retval_ptr (stmt);
4074 if (*t != NULL_TREE)
4075 modified |= sra_ipa_modify_expr (t, true, adjustments);
4079 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4080 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4084 /* Operands must be processed before the lhs. */
4085 for (i = 0; i < gimple_call_num_args (stmt); i++)
4087 t = gimple_call_arg_ptr (stmt, i);
4088 modified |= sra_ipa_modify_expr (t, true, adjustments);
4091 if (gimple_call_lhs (stmt))
4093 t = gimple_call_lhs_ptr (stmt);
4094 modified |= sra_ipa_modify_expr (t, false, adjustments);
4095 modified |= replace_removed_params_ssa_names (stmt,
4101 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4103 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4104 modified |= sra_ipa_modify_expr (t, true, adjustments);
4106 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4108 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4109 modified |= sra_ipa_modify_expr (t, false, adjustments);
4121 maybe_clean_eh_stmt (stmt);
4126 gimple_purge_dead_eh_edges (bb);
4130 /* Call gimple_debug_bind_reset_value on all debug statements describing
4131 gimple register parameters that are being removed or replaced. */
4134 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4138 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4139 for (i = 0; i < len; i++)
4141 struct ipa_parm_adjustment *adj;
4142 imm_use_iterator ui;
4146 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4147 if (adj->copy_param || !is_gimple_reg (adj->base))
4149 name = gimple_default_def (cfun, adj->base);
4152 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4154 /* All other users must have been removed by
4155 ipa_sra_modify_function_body. */
4156 gcc_assert (is_gimple_debug (stmt));
4157 gimple_debug_bind_reset_value (stmt);
4163 /* Return true iff all callers have at least as many actual arguments as there
4164 are formal parameters in the current function. */
4167 all_callers_have_enough_arguments_p (struct cgraph_node *node)
4169 struct cgraph_edge *cs;
4170 for (cs = node->callers; cs; cs = cs->next_caller)
4171 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4178 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4181 convert_callers (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4183 tree old_cur_fndecl = current_function_decl;
4184 struct cgraph_edge *cs;
4185 basic_block this_block;
4186 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4188 for (cs = node->callers; cs; cs = cs->next_caller)
4190 current_function_decl = cs->caller->decl;
4191 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4194 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4195 cs->caller->uid, cs->callee->uid,
4196 cgraph_node_name (cs->caller),
4197 cgraph_node_name (cs->callee));
4199 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4204 for (cs = node->callers; cs; cs = cs->next_caller)
4205 if (!bitmap_bit_p (recomputed_callers, cs->caller->uid))
4207 compute_inline_parameters (cs->caller);
4208 bitmap_set_bit (recomputed_callers, cs->caller->uid);
4210 BITMAP_FREE (recomputed_callers);
4212 current_function_decl = old_cur_fndecl;
4214 if (!encountered_recursive_call)
4217 FOR_EACH_BB (this_block)
4219 gimple_stmt_iterator gsi;
4221 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4223 gimple stmt = gsi_stmt (gsi);
4225 if (gimple_code (stmt) != GIMPLE_CALL)
4227 call_fndecl = gimple_call_fndecl (stmt);
4228 if (call_fndecl && cgraph_get_node (call_fndecl) == node)
4231 fprintf (dump_file, "Adjusting recursive call");
4232 ipa_modify_call_arguments (NULL, stmt, adjustments);
4240 /* Create an abstract origin declaration for OLD_DECL and make it an abstract
4241 origin of the provided decl so that there are preserved parameters for debug
4245 create_abstract_origin (tree old_decl)
4247 if (!DECL_ABSTRACT_ORIGIN (old_decl))
4249 tree new_decl = copy_node (old_decl);
4251 DECL_ABSTRACT (new_decl) = 1;
4252 SET_DECL_ASSEMBLER_NAME (new_decl, NULL_TREE);
4253 SET_DECL_RTL (new_decl, NULL);
4254 DECL_STRUCT_FUNCTION (new_decl) = NULL;
4255 DECL_ARTIFICIAL (old_decl) = 1;
4256 DECL_ABSTRACT_ORIGIN (old_decl) = new_decl;
4260 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4261 as given in ADJUSTMENTS. */
4264 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4266 struct cgraph_node *alias;
4267 for (alias = node->same_body; alias; alias = alias->next)
4268 ipa_modify_formal_parameters (alias->decl, adjustments, "ISRA");
4269 /* current_function_decl must be handled last, after same_body aliases,
4270 as following functions will use what it computed. */
4271 create_abstract_origin (current_function_decl);
4272 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4273 ipa_sra_modify_function_body (adjustments);
4274 sra_ipa_reset_debug_stmts (adjustments);
4275 convert_callers (node, adjustments);
4276 cgraph_make_node_local (node);
4280 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4281 attributes, return true otherwise. NODE is the cgraph node of the current
4285 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4287 if (!cgraph_node_can_be_local_p (node))
4290 fprintf (dump_file, "Function not local to this compilation unit.\n");
4294 if (DECL_VIRTUAL_P (current_function_decl))
4297 fprintf (dump_file, "Function is a virtual method.\n");
4301 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4302 && node->global.size >= MAX_INLINE_INSNS_AUTO)
4305 fprintf (dump_file, "Function too big to be made truly local.\n");
4313 "Function has no callers in this compilation unit.\n");
4320 fprintf (dump_file, "Function uses stdarg. \n");
4324 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4330 /* Perform early interprocedural SRA. */
4333 ipa_early_sra (void)
4335 struct cgraph_node *node = cgraph_node (current_function_decl);
4336 ipa_parm_adjustment_vec adjustments;
4339 if (!ipa_sra_preliminary_function_checks (node))
4343 sra_mode = SRA_MODE_EARLY_IPA;
4345 if (!find_param_candidates ())
4348 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4352 if (!all_callers_have_enough_arguments_p (node))
4355 fprintf (dump_file, "There are callers with insufficient number of "
4360 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4362 * last_basic_block_for_function (cfun));
4363 final_bbs = BITMAP_ALLOC (NULL);
4366 if (encountered_apply_args)
4369 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4373 if (encountered_unchangable_recursive_call)
4376 fprintf (dump_file, "Function calls itself with insufficient "
4377 "number of arguments.\n");
4381 adjustments = analyze_all_param_acesses ();
4385 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4387 modify_function (node, adjustments);
4388 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4389 ret = TODO_update_ssa;
4391 statistics_counter_event (cfun, "Unused parameters deleted",
4392 sra_stats.deleted_unused_parameters);
4393 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4394 sra_stats.scalar_by_ref_to_by_val);
4395 statistics_counter_event (cfun, "Aggregate parameters broken up",
4396 sra_stats.aggregate_params_reduced);
4397 statistics_counter_event (cfun, "Aggregate parameter components created",
4398 sra_stats.param_reductions_created);
4401 BITMAP_FREE (final_bbs);
4402 free (bb_dereferences);
4404 sra_deinitialize ();
4408 /* Return if early ipa sra shall be performed. */
4410 ipa_early_sra_gate (void)
4412 return flag_ipa_sra;
4415 struct gimple_opt_pass pass_early_ipa_sra =
4419 "eipa_sra", /* name */
4420 ipa_early_sra_gate, /* gate */
4421 ipa_early_sra, /* execute */
4424 0, /* static_pass_number */
4425 TV_IPA_SRA, /* tv_id */
4426 0, /* properties_required */
4427 0, /* properties_provided */
4428 0, /* properties_destroyed */
4429 0, /* todo_flags_start */
4430 TODO_dump_func | TODO_dump_cgraph /* todo_flags_finish */