1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
4 Copyright (C) 2008, 2009, 2010, 2011 Free Software Foundation, Inc.
5 Contributed by Martin Jambor <mjambor@suse.cz>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* This file implements Scalar Reduction of Aggregates (SRA). SRA is run
24 twice, once in the early stages of compilation (early SRA) and once in the
25 late stages (late SRA). The aim of both is to turn references to scalar
26 parts of aggregates into uses of independent scalar variables.
28 The two passes are nearly identical, the only difference is that early SRA
29 does not scalarize unions which are used as the result in a GIMPLE_RETURN
30 statement because together with inlining this can lead to weird type
33 Both passes operate in four stages:
35 1. The declarations that have properties which make them candidates for
36 scalarization are identified in function find_var_candidates(). The
37 candidates are stored in candidate_bitmap.
39 2. The function body is scanned. In the process, declarations which are
40 used in a manner that prevent their scalarization are removed from the
41 candidate bitmap. More importantly, for every access into an aggregate,
42 an access structure (struct access) is created by create_access() and
43 stored in a vector associated with the aggregate. Among other
44 information, the aggregate declaration, the offset and size of the access
45 and its type are stored in the structure.
47 On a related note, assign_link structures are created for every assign
48 statement between candidate aggregates and attached to the related
51 3. The vectors of accesses are analyzed. They are first sorted according to
52 their offset and size and then scanned for partially overlapping accesses
53 (i.e. those which overlap but one is not entirely within another). Such
54 an access disqualifies the whole aggregate from being scalarized.
56 If there is no such inhibiting overlap, a representative access structure
57 is chosen for every unique combination of offset and size. Afterwards,
58 the pass builds a set of trees from these structures, in which children
59 of an access are within their parent (in terms of offset and size).
61 Then accesses are propagated whenever possible (i.e. in cases when it
62 does not create a partially overlapping access) across assign_links from
63 the right hand side to the left hand side.
65 Then the set of trees for each declaration is traversed again and those
66 accesses which should be replaced by a scalar are identified.
68 4. The function is traversed again, and for every reference into an
69 aggregate that has some component which is about to be scalarized,
70 statements are amended and new statements are created as necessary.
71 Finally, if a parameter got scalarized, the scalar replacements are
72 initialized with values from respective parameter aggregates. */
76 #include "coretypes.h"
77 #include "alloc-pool.h"
82 #include "tree-flow.h"
84 #include "tree-pretty-print.h"
85 #include "statistics.h"
86 #include "tree-dump.h"
92 #include "tree-inline.h"
93 #include "gimple-pretty-print.h"
94 #include "ipa-inline.h"
96 /* Enumeration of all aggregate reductions we can do. */
97 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
98 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
99 SRA_MODE_INTRA }; /* late intraprocedural SRA */
101 /* Global variable describing which aggregate reduction we are performing at
103 static enum sra_mode sra_mode;
107 /* ACCESS represents each access to an aggregate variable (as a whole or a
108 part). It can also represent a group of accesses that refer to exactly the
109 same fragment of an aggregate (i.e. those that have exactly the same offset
110 and size). Such representatives for a single aggregate, once determined,
111 are linked in a linked list and have the group fields set.
113 Moreover, when doing intraprocedural SRA, a tree is built from those
114 representatives (by the means of first_child and next_sibling pointers), in
115 which all items in a subtree are "within" the root, i.e. their offset is
116 greater or equal to offset of the root and offset+size is smaller or equal
117 to offset+size of the root. Children of an access are sorted by offset.
119 Note that accesses to parts of vector and complex number types always
120 represented by an access to the whole complex number or a vector. It is a
121 duty of the modifying functions to replace them appropriately. */
125 /* Values returned by `get_ref_base_and_extent' for each component reference
126 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
127 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
128 HOST_WIDE_INT offset;
132 /* Expression. It is context dependent so do not use it to create new
133 expressions to access the original aggregate. See PR 42154 for a
139 /* The statement this access belongs to. */
142 /* Next group representative for this aggregate. */
143 struct access *next_grp;
145 /* Pointer to the group representative. Pointer to itself if the struct is
146 the representative. */
147 struct access *group_representative;
149 /* If this access has any children (in terms of the definition above), this
150 points to the first one. */
151 struct access *first_child;
153 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
154 described above. In IPA-SRA this is a pointer to the next access
155 belonging to the same group (having the same representative). */
156 struct access *next_sibling;
158 /* Pointers to the first and last element in the linked list of assign
160 struct assign_link *first_link, *last_link;
162 /* Pointer to the next access in the work queue. */
163 struct access *next_queued;
165 /* Replacement variable for this access "region." Never to be accessed
166 directly, always only by the means of get_access_replacement() and only
167 when grp_to_be_replaced flag is set. */
168 tree replacement_decl;
170 /* Is this particular access write access? */
173 /* Is this access an access to a non-addressable field? */
174 unsigned non_addressable : 1;
176 /* Is this access currently in the work queue? */
177 unsigned grp_queued : 1;
179 /* Does this group contain a write access? This flag is propagated down the
181 unsigned grp_write : 1;
183 /* Does this group contain a read access? This flag is propagated down the
185 unsigned grp_read : 1;
187 /* Does this group contain a read access that comes from an assignment
188 statement? This flag is propagated down the access tree. */
189 unsigned grp_assignment_read : 1;
191 /* Does this group contain a write access that comes from an assignment
192 statement? This flag is propagated down the access tree. */
193 unsigned grp_assignment_write : 1;
195 /* Does this group contain a read access through a scalar type? This flag is
196 not propagated in the access tree in any direction. */
197 unsigned grp_scalar_read : 1;
199 /* Does this group contain a write access through a scalar type? This flag
200 is not propagated in the access tree in any direction. */
201 unsigned grp_scalar_write : 1;
203 /* Is this access an artificial one created to scalarize some record
205 unsigned grp_total_scalarization : 1;
207 /* Other passes of the analysis use this bit to make function
208 analyze_access_subtree create scalar replacements for this group if
210 unsigned grp_hint : 1;
212 /* Is the subtree rooted in this access fully covered by scalar
214 unsigned grp_covered : 1;
216 /* If set to true, this access and all below it in an access tree must not be
218 unsigned grp_unscalarizable_region : 1;
220 /* Whether data have been written to parts of the aggregate covered by this
221 access which is not to be scalarized. This flag is propagated up in the
223 unsigned grp_unscalarized_data : 1;
225 /* Does this access and/or group contain a write access through a
227 unsigned grp_partial_lhs : 1;
229 /* Set when a scalar replacement should be created for this variable. We do
230 the decision and creation at different places because create_tmp_var
231 cannot be called from within FOR_EACH_REFERENCED_VAR. */
232 unsigned grp_to_be_replaced : 1;
234 /* Should TREE_NO_WARNING of a replacement be set? */
235 unsigned grp_no_warning : 1;
237 /* Is it possible that the group refers to data which might be (directly or
238 otherwise) modified? */
239 unsigned grp_maybe_modified : 1;
241 /* Set when this is a representative of a pointer to scalar (i.e. by
242 reference) parameter which we consider for turning into a plain scalar
243 (i.e. a by value parameter). */
244 unsigned grp_scalar_ptr : 1;
246 /* Set when we discover that this pointer is not safe to dereference in the
248 unsigned grp_not_necessarilly_dereferenced : 1;
251 typedef struct access *access_p;
253 DEF_VEC_P (access_p);
254 DEF_VEC_ALLOC_P (access_p, heap);
256 /* Alloc pool for allocating access structures. */
257 static alloc_pool access_pool;
259 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
260 are used to propagate subaccesses from rhs to lhs as long as they don't
261 conflict with what is already there. */
264 struct access *lacc, *racc;
265 struct assign_link *next;
268 /* Alloc pool for allocating assign link structures. */
269 static alloc_pool link_pool;
271 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
272 static struct pointer_map_t *base_access_vec;
274 /* Bitmap of candidates. */
275 static bitmap candidate_bitmap;
277 /* Bitmap of candidates which we should try to entirely scalarize away and
278 those which cannot be (because they are and need be used as a whole). */
279 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
281 /* Obstack for creation of fancy names. */
282 static struct obstack name_obstack;
284 /* Head of a linked list of accesses that need to have its subaccesses
285 propagated to their assignment counterparts. */
286 static struct access *work_queue_head;
288 /* Number of parameters of the analyzed function when doing early ipa SRA. */
289 static int func_param_count;
291 /* scan_function sets the following to true if it encounters a call to
292 __builtin_apply_args. */
293 static bool encountered_apply_args;
295 /* Set by scan_function when it finds a recursive call. */
296 static bool encountered_recursive_call;
298 /* Set by scan_function when it finds a recursive call with less actual
299 arguments than formal parameters.. */
300 static bool encountered_unchangable_recursive_call;
302 /* This is a table in which for each basic block and parameter there is a
303 distance (offset + size) in that parameter which is dereferenced and
304 accessed in that BB. */
305 static HOST_WIDE_INT *bb_dereferences;
306 /* Bitmap of BBs that can cause the function to "stop" progressing by
307 returning, throwing externally, looping infinitely or calling a function
308 which might abort etc.. */
309 static bitmap final_bbs;
311 /* Representative of no accesses at all. */
312 static struct access no_accesses_representant;
314 /* Predicate to test the special value. */
317 no_accesses_p (struct access *access)
319 return access == &no_accesses_representant;
322 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
323 representative fields are dumped, otherwise those which only describe the
324 individual access are. */
328 /* Number of processed aggregates is readily available in
329 analyze_all_variable_accesses and so is not stored here. */
331 /* Number of created scalar replacements. */
334 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
338 /* Number of statements created by generate_subtree_copies. */
341 /* Number of statements created by load_assign_lhs_subreplacements. */
344 /* Number of times sra_modify_assign has deleted a statement. */
347 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
348 RHS reparately due to type conversions or nonexistent matching
350 int separate_lhs_rhs_handling;
352 /* Number of parameters that were removed because they were unused. */
353 int deleted_unused_parameters;
355 /* Number of scalars passed as parameters by reference that have been
356 converted to be passed by value. */
357 int scalar_by_ref_to_by_val;
359 /* Number of aggregate parameters that were replaced by one or more of their
361 int aggregate_params_reduced;
363 /* Numbber of components created when splitting aggregate parameters. */
364 int param_reductions_created;
368 dump_access (FILE *f, struct access *access, bool grp)
370 fprintf (f, "access { ");
371 fprintf (f, "base = (%d)'", DECL_UID (access->base));
372 print_generic_expr (f, access->base, 0);
373 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
374 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
375 fprintf (f, ", expr = ");
376 print_generic_expr (f, access->expr, 0);
377 fprintf (f, ", type = ");
378 print_generic_expr (f, access->type, 0);
380 fprintf (f, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
381 "grp_assignment_write = %d, grp_scalar_read = %d, "
382 "grp_scalar_write = %d, grp_total_scalarization = %d, "
383 "grp_hint = %d, grp_covered = %d, "
384 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
385 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
386 "grp_maybe_modified = %d, "
387 "grp_not_necessarilly_dereferenced = %d\n",
388 access->grp_read, access->grp_write, access->grp_assignment_read,
389 access->grp_assignment_write, access->grp_scalar_read,
390 access->grp_scalar_write, access->grp_total_scalarization,
391 access->grp_hint, access->grp_covered,
392 access->grp_unscalarizable_region, access->grp_unscalarized_data,
393 access->grp_partial_lhs, access->grp_to_be_replaced,
394 access->grp_maybe_modified,
395 access->grp_not_necessarilly_dereferenced);
397 fprintf (f, ", write = %d, grp_total_scalarization = %d, "
398 "grp_partial_lhs = %d\n",
399 access->write, access->grp_total_scalarization,
400 access->grp_partial_lhs);
403 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
406 dump_access_tree_1 (FILE *f, struct access *access, int level)
412 for (i = 0; i < level; i++)
413 fputs ("* ", dump_file);
415 dump_access (f, access, true);
417 if (access->first_child)
418 dump_access_tree_1 (f, access->first_child, level + 1);
420 access = access->next_sibling;
425 /* Dump all access trees for a variable, given the pointer to the first root in
429 dump_access_tree (FILE *f, struct access *access)
431 for (; access; access = access->next_grp)
432 dump_access_tree_1 (f, access, 0);
435 /* Return true iff ACC is non-NULL and has subaccesses. */
438 access_has_children_p (struct access *acc)
440 return acc && acc->first_child;
443 /* Return true iff ACC is (partly) covered by at least one replacement. */
446 access_has_replacements_p (struct access *acc)
448 struct access *child;
449 if (acc->grp_to_be_replaced)
451 for (child = acc->first_child; child; child = child->next_sibling)
452 if (access_has_replacements_p (child))
457 /* Return a vector of pointers to accesses for the variable given in BASE or
458 NULL if there is none. */
460 static VEC (access_p, heap) *
461 get_base_access_vector (tree base)
465 slot = pointer_map_contains (base_access_vec, base);
469 return *(VEC (access_p, heap) **) slot;
472 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
473 in ACCESS. Return NULL if it cannot be found. */
475 static struct access *
476 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
479 while (access && (access->offset != offset || access->size != size))
481 struct access *child = access->first_child;
483 while (child && (child->offset + child->size <= offset))
484 child = child->next_sibling;
491 /* Return the first group representative for DECL or NULL if none exists. */
493 static struct access *
494 get_first_repr_for_decl (tree base)
496 VEC (access_p, heap) *access_vec;
498 access_vec = get_base_access_vector (base);
502 return VEC_index (access_p, access_vec, 0);
505 /* Find an access representative for the variable BASE and given OFFSET and
506 SIZE. Requires that access trees have already been built. Return NULL if
507 it cannot be found. */
509 static struct access *
510 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
513 struct access *access;
515 access = get_first_repr_for_decl (base);
516 while (access && (access->offset + access->size <= offset))
517 access = access->next_grp;
521 return find_access_in_subtree (access, offset, size);
524 /* Add LINK to the linked list of assign links of RACC. */
526 add_link_to_rhs (struct access *racc, struct assign_link *link)
528 gcc_assert (link->racc == racc);
530 if (!racc->first_link)
532 gcc_assert (!racc->last_link);
533 racc->first_link = link;
536 racc->last_link->next = link;
538 racc->last_link = link;
542 /* Move all link structures in their linked list in OLD_RACC to the linked list
545 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
547 if (!old_racc->first_link)
549 gcc_assert (!old_racc->last_link);
553 if (new_racc->first_link)
555 gcc_assert (!new_racc->last_link->next);
556 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
558 new_racc->last_link->next = old_racc->first_link;
559 new_racc->last_link = old_racc->last_link;
563 gcc_assert (!new_racc->last_link);
565 new_racc->first_link = old_racc->first_link;
566 new_racc->last_link = old_racc->last_link;
568 old_racc->first_link = old_racc->last_link = NULL;
571 /* Add ACCESS to the work queue (which is actually a stack). */
574 add_access_to_work_queue (struct access *access)
576 if (!access->grp_queued)
578 gcc_assert (!access->next_queued);
579 access->next_queued = work_queue_head;
580 access->grp_queued = 1;
581 work_queue_head = access;
585 /* Pop an access from the work queue, and return it, assuming there is one. */
587 static struct access *
588 pop_access_from_work_queue (void)
590 struct access *access = work_queue_head;
592 work_queue_head = access->next_queued;
593 access->next_queued = NULL;
594 access->grp_queued = 0;
599 /* Allocate necessary structures. */
602 sra_initialize (void)
604 candidate_bitmap = BITMAP_ALLOC (NULL);
605 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
606 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
607 gcc_obstack_init (&name_obstack);
608 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
609 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
610 base_access_vec = pointer_map_create ();
611 memset (&sra_stats, 0, sizeof (sra_stats));
612 encountered_apply_args = false;
613 encountered_recursive_call = false;
614 encountered_unchangable_recursive_call = false;
617 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
620 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
621 void *data ATTRIBUTE_UNUSED)
623 VEC (access_p, heap) *access_vec;
624 access_vec = (VEC (access_p, heap) *) *value;
625 VEC_free (access_p, heap, access_vec);
630 /* Deallocate all general structures. */
633 sra_deinitialize (void)
635 BITMAP_FREE (candidate_bitmap);
636 BITMAP_FREE (should_scalarize_away_bitmap);
637 BITMAP_FREE (cannot_scalarize_away_bitmap);
638 free_alloc_pool (access_pool);
639 free_alloc_pool (link_pool);
640 obstack_free (&name_obstack, NULL);
642 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
643 pointer_map_destroy (base_access_vec);
646 /* Remove DECL from candidates for SRA and write REASON to the dump file if
649 disqualify_candidate (tree decl, const char *reason)
651 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
653 if (dump_file && (dump_flags & TDF_DETAILS))
655 fprintf (dump_file, "! Disqualifying ");
656 print_generic_expr (dump_file, decl, 0);
657 fprintf (dump_file, " - %s\n", reason);
661 /* Return true iff the type contains a field or an element which does not allow
665 type_internals_preclude_sra_p (tree type, const char **msg)
670 switch (TREE_CODE (type))
674 case QUAL_UNION_TYPE:
675 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
676 if (TREE_CODE (fld) == FIELD_DECL)
678 tree ft = TREE_TYPE (fld);
680 if (TREE_THIS_VOLATILE (fld))
682 *msg = "volatile structure field";
685 if (!DECL_FIELD_OFFSET (fld))
687 *msg = "no structure field offset";
690 if (!DECL_SIZE (fld))
692 *msg = "zero structure field size";
695 if (!host_integerp (DECL_FIELD_OFFSET (fld), 1))
697 *msg = "structure field offset not fixed";
700 if (!host_integerp (DECL_SIZE (fld), 1))
702 *msg = "structure field size not fixed";
705 if (AGGREGATE_TYPE_P (ft)
706 && int_bit_position (fld) % BITS_PER_UNIT != 0)
708 *msg = "structure field is bit field";
712 if (AGGREGATE_TYPE_P (ft) && type_internals_preclude_sra_p (ft, msg))
719 et = TREE_TYPE (type);
721 if (TYPE_VOLATILE (et))
723 *msg = "element type is volatile";
727 if (AGGREGATE_TYPE_P (et) && type_internals_preclude_sra_p (et, msg))
737 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
738 base variable if it is. Return T if it is not an SSA_NAME. */
741 get_ssa_base_param (tree t)
743 if (TREE_CODE (t) == SSA_NAME)
745 if (SSA_NAME_IS_DEFAULT_DEF (t))
746 return SSA_NAME_VAR (t);
753 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
754 belongs to, unless the BB has already been marked as a potentially
758 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
760 basic_block bb = gimple_bb (stmt);
761 int idx, parm_index = 0;
764 if (bitmap_bit_p (final_bbs, bb->index))
767 for (parm = DECL_ARGUMENTS (current_function_decl);
768 parm && parm != base;
769 parm = DECL_CHAIN (parm))
772 gcc_assert (parm_index < func_param_count);
774 idx = bb->index * func_param_count + parm_index;
775 if (bb_dereferences[idx] < dist)
776 bb_dereferences[idx] = dist;
779 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
780 the three fields. Also add it to the vector of accesses corresponding to
781 the base. Finally, return the new access. */
783 static struct access *
784 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
786 VEC (access_p, heap) *vec;
787 struct access *access;
790 access = (struct access *) pool_alloc (access_pool);
791 memset (access, 0, sizeof (struct access));
793 access->offset = offset;
796 slot = pointer_map_contains (base_access_vec, base);
798 vec = (VEC (access_p, heap) *) *slot;
800 vec = VEC_alloc (access_p, heap, 32);
802 VEC_safe_push (access_p, heap, vec, access);
804 *((struct VEC (access_p,heap) **)
805 pointer_map_insert (base_access_vec, base)) = vec;
810 /* Create and insert access for EXPR. Return created access, or NULL if it is
813 static struct access *
814 create_access (tree expr, gimple stmt, bool write)
816 struct access *access;
817 HOST_WIDE_INT offset, size, max_size;
819 bool ptr, unscalarizable_region = false;
821 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
823 if (sra_mode == SRA_MODE_EARLY_IPA
824 && TREE_CODE (base) == MEM_REF)
826 base = get_ssa_base_param (TREE_OPERAND (base, 0));
834 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
837 if (sra_mode == SRA_MODE_EARLY_IPA)
839 if (size < 0 || size != max_size)
841 disqualify_candidate (base, "Encountered a variable sized access.");
844 if (TREE_CODE (expr) == COMPONENT_REF
845 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
847 disqualify_candidate (base, "Encountered a bit-field access.");
850 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
853 mark_parm_dereference (base, offset + size, stmt);
857 if (size != max_size)
860 unscalarizable_region = true;
864 disqualify_candidate (base, "Encountered an unconstrained access.");
869 access = create_access_1 (base, offset, size);
871 access->type = TREE_TYPE (expr);
872 access->write = write;
873 access->grp_unscalarizable_region = unscalarizable_region;
876 if (TREE_CODE (expr) == COMPONENT_REF
877 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
878 access->non_addressable = 1;
884 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
885 register types or (recursively) records with only these two kinds of fields.
886 It also returns false if any of these records contains a bit-field. */
889 type_consists_of_records_p (tree type)
893 if (TREE_CODE (type) != RECORD_TYPE)
896 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
897 if (TREE_CODE (fld) == FIELD_DECL)
899 tree ft = TREE_TYPE (fld);
901 if (DECL_BIT_FIELD (fld))
904 if (!is_gimple_reg_type (ft)
905 && !type_consists_of_records_p (ft))
912 /* Create total_scalarization accesses for all scalar type fields in DECL that
913 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
914 must be the top-most VAR_DECL representing the variable, OFFSET must be the
915 offset of DECL within BASE. REF must be the memory reference expression for
919 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
922 tree fld, decl_type = TREE_TYPE (decl);
924 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
925 if (TREE_CODE (fld) == FIELD_DECL)
927 HOST_WIDE_INT pos = offset + int_bit_position (fld);
928 tree ft = TREE_TYPE (fld);
929 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
932 if (is_gimple_reg_type (ft))
934 struct access *access;
937 size = tree_low_cst (DECL_SIZE (fld), 1);
938 access = create_access_1 (base, pos, size);
941 access->grp_total_scalarization = 1;
942 /* Accesses for intraprocedural SRA can have their stmt NULL. */
945 completely_scalarize_record (base, fld, pos, nref);
949 /* Create total_scalarization accesses for all scalar type fields in VAR and
950 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
951 type_consists_of_records_p. */
954 completely_scalarize_var (tree var)
956 HOST_WIDE_INT size = tree_low_cst (DECL_SIZE (var), 1);
957 struct access *access;
959 access = create_access_1 (var, 0, size);
961 access->type = TREE_TYPE (var);
962 access->grp_total_scalarization = 1;
964 completely_scalarize_record (var, var, 0, var);
967 /* Search the given tree for a declaration by skipping handled components and
968 exclude it from the candidates. */
971 disqualify_base_of_expr (tree t, const char *reason)
973 t = get_base_address (t);
974 if (sra_mode == SRA_MODE_EARLY_IPA
975 && TREE_CODE (t) == MEM_REF)
976 t = get_ssa_base_param (TREE_OPERAND (t, 0));
979 disqualify_candidate (t, reason);
982 /* Scan expression EXPR and create access structures for all accesses to
983 candidates for scalarization. Return the created access or NULL if none is
986 static struct access *
987 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
989 struct access *ret = NULL;
992 if (TREE_CODE (expr) == BIT_FIELD_REF
993 || TREE_CODE (expr) == IMAGPART_EXPR
994 || TREE_CODE (expr) == REALPART_EXPR)
996 expr = TREE_OPERAND (expr, 0);
1000 partial_ref = false;
1002 /* We need to dive through V_C_Es in order to get the size of its parameter
1003 and not the result type. Ada produces such statements. We are also
1004 capable of handling the topmost V_C_E but not any of those buried in other
1005 handled components. */
1006 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
1007 expr = TREE_OPERAND (expr, 0);
1009 if (contains_view_convert_expr_p (expr))
1011 disqualify_base_of_expr (expr, "V_C_E under a different handled "
1016 switch (TREE_CODE (expr))
1019 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
1020 && sra_mode != SRA_MODE_EARLY_IPA)
1028 case ARRAY_RANGE_REF:
1029 ret = create_access (expr, stmt, write);
1036 if (write && partial_ref && ret)
1037 ret->grp_partial_lhs = 1;
1042 /* Scan expression EXPR and create access structures for all accesses to
1043 candidates for scalarization. Return true if any access has been inserted.
1044 STMT must be the statement from which the expression is taken, WRITE must be
1045 true if the expression is a store and false otherwise. */
1048 build_access_from_expr (tree expr, gimple stmt, bool write)
1050 struct access *access;
1052 access = build_access_from_expr_1 (expr, stmt, write);
1055 /* This means the aggregate is accesses as a whole in a way other than an
1056 assign statement and thus cannot be removed even if we had a scalar
1057 replacement for everything. */
1058 if (cannot_scalarize_away_bitmap)
1059 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1065 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1066 modes in which it matters, return true iff they have been disqualified. RHS
1067 may be NULL, in that case ignore it. If we scalarize an aggregate in
1068 intra-SRA we may need to add statements after each statement. This is not
1069 possible if a statement unconditionally has to end the basic block. */
1071 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1073 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1074 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1076 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1078 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1084 /* Return true if EXP is a memory reference less aligned than ALIGN. This is
1085 invoked only on strict-alignment targets. */
1088 tree_non_aligned_mem_p (tree exp, unsigned int align)
1090 unsigned int exp_align;
1092 if (TREE_CODE (exp) == VIEW_CONVERT_EXPR)
1093 exp = TREE_OPERAND (exp, 0);
1095 if (TREE_CODE (exp) == SSA_NAME || is_gimple_min_invariant (exp))
1098 /* get_object_alignment will fall back to BITS_PER_UNIT if it cannot
1099 compute an explicit alignment. Pretend that dereferenced pointers
1100 are always aligned on strict-alignment targets. */
1101 if (TREE_CODE (exp) == MEM_REF || TREE_CODE (exp) == TARGET_MEM_REF)
1102 exp_align = get_object_or_type_alignment (exp);
1104 exp_align = get_object_alignment (exp);
1106 if (exp_align < align)
1112 /* Return true if EXP is a memory reference less aligned than what the access
1113 ACC would require. This is invoked only on strict-alignment targets. */
1116 tree_non_aligned_mem_for_access_p (tree exp, struct access *acc)
1118 unsigned int acc_align;
1120 /* The alignment of the access is that of its expression. However, it may
1121 have been artificially increased, e.g. by a local alignment promotion,
1122 so we cap it to the alignment of the type of the base, on the grounds
1123 that valid sub-accesses cannot be more aligned than that. */
1124 acc_align = get_object_alignment (acc->expr);
1125 if (acc->base && acc_align > TYPE_ALIGN (TREE_TYPE (acc->base)))
1126 acc_align = TYPE_ALIGN (TREE_TYPE (acc->base));
1128 return tree_non_aligned_mem_p (exp, acc_align);
1131 /* Scan expressions occuring in STMT, create access structures for all accesses
1132 to candidates for scalarization and remove those candidates which occur in
1133 statements or expressions that prevent them from being split apart. Return
1134 true if any access has been inserted. */
1137 build_accesses_from_assign (gimple stmt)
1140 struct access *lacc, *racc;
1142 if (!gimple_assign_single_p (stmt)
1143 /* Scope clobbers don't influence scalarization. */
1144 || gimple_clobber_p (stmt))
1147 lhs = gimple_assign_lhs (stmt);
1148 rhs = gimple_assign_rhs1 (stmt);
1150 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1153 racc = build_access_from_expr_1 (rhs, stmt, false);
1154 lacc = build_access_from_expr_1 (lhs, stmt, true);
1158 lacc->grp_assignment_write = 1;
1159 if (STRICT_ALIGNMENT && tree_non_aligned_mem_for_access_p (rhs, lacc))
1160 lacc->grp_unscalarizable_region = 1;
1165 racc->grp_assignment_read = 1;
1166 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1167 && !is_gimple_reg_type (racc->type))
1168 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1169 if (STRICT_ALIGNMENT && tree_non_aligned_mem_for_access_p (lhs, racc))
1170 racc->grp_unscalarizable_region = 1;
1174 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1175 && !lacc->grp_unscalarizable_region
1176 && !racc->grp_unscalarizable_region
1177 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1178 /* FIXME: Turn the following line into an assert after PR 40058 is
1180 && lacc->size == racc->size
1181 && useless_type_conversion_p (lacc->type, racc->type))
1183 struct assign_link *link;
1185 link = (struct assign_link *) pool_alloc (link_pool);
1186 memset (link, 0, sizeof (struct assign_link));
1191 add_link_to_rhs (racc, link);
1194 return lacc || racc;
1197 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1198 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1201 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1202 void *data ATTRIBUTE_UNUSED)
1204 op = get_base_address (op);
1207 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1212 /* Return true iff callsite CALL has at least as many actual arguments as there
1213 are formal parameters of the function currently processed by IPA-SRA. */
1216 callsite_has_enough_arguments_p (gimple call)
1218 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1221 /* Scan function and look for interesting expressions and create access
1222 structures for them. Return true iff any access is created. */
1225 scan_function (void)
1232 gimple_stmt_iterator gsi;
1233 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1235 gimple stmt = gsi_stmt (gsi);
1239 if (final_bbs && stmt_can_throw_external (stmt))
1240 bitmap_set_bit (final_bbs, bb->index);
1241 switch (gimple_code (stmt))
1244 t = gimple_return_retval (stmt);
1246 ret |= build_access_from_expr (t, stmt, false);
1248 bitmap_set_bit (final_bbs, bb->index);
1252 ret |= build_accesses_from_assign (stmt);
1256 for (i = 0; i < gimple_call_num_args (stmt); i++)
1257 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1260 if (sra_mode == SRA_MODE_EARLY_IPA)
1262 tree dest = gimple_call_fndecl (stmt);
1263 int flags = gimple_call_flags (stmt);
1267 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1268 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1269 encountered_apply_args = true;
1270 if (cgraph_get_node (dest)
1271 == cgraph_get_node (current_function_decl))
1273 encountered_recursive_call = true;
1274 if (!callsite_has_enough_arguments_p (stmt))
1275 encountered_unchangable_recursive_call = true;
1280 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1281 bitmap_set_bit (final_bbs, bb->index);
1284 t = gimple_call_lhs (stmt);
1285 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1286 ret |= build_access_from_expr (t, stmt, true);
1290 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1293 bitmap_set_bit (final_bbs, bb->index);
1295 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1297 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1298 ret |= build_access_from_expr (t, stmt, false);
1300 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1302 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1303 ret |= build_access_from_expr (t, stmt, true);
1316 /* Helper of QSORT function. There are pointers to accesses in the array. An
1317 access is considered smaller than another if it has smaller offset or if the
1318 offsets are the same but is size is bigger. */
1321 compare_access_positions (const void *a, const void *b)
1323 const access_p *fp1 = (const access_p *) a;
1324 const access_p *fp2 = (const access_p *) b;
1325 const access_p f1 = *fp1;
1326 const access_p f2 = *fp2;
1328 if (f1->offset != f2->offset)
1329 return f1->offset < f2->offset ? -1 : 1;
1331 if (f1->size == f2->size)
1333 if (f1->type == f2->type)
1335 /* Put any non-aggregate type before any aggregate type. */
1336 else if (!is_gimple_reg_type (f1->type)
1337 && is_gimple_reg_type (f2->type))
1339 else if (is_gimple_reg_type (f1->type)
1340 && !is_gimple_reg_type (f2->type))
1342 /* Put any complex or vector type before any other scalar type. */
1343 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1344 && TREE_CODE (f1->type) != VECTOR_TYPE
1345 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1346 || TREE_CODE (f2->type) == VECTOR_TYPE))
1348 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1349 || TREE_CODE (f1->type) == VECTOR_TYPE)
1350 && TREE_CODE (f2->type) != COMPLEX_TYPE
1351 && TREE_CODE (f2->type) != VECTOR_TYPE)
1353 /* Put the integral type with the bigger precision first. */
1354 else if (INTEGRAL_TYPE_P (f1->type)
1355 && INTEGRAL_TYPE_P (f2->type))
1356 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1357 /* Put any integral type with non-full precision last. */
1358 else if (INTEGRAL_TYPE_P (f1->type)
1359 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1360 != TYPE_PRECISION (f1->type)))
1362 else if (INTEGRAL_TYPE_P (f2->type)
1363 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1364 != TYPE_PRECISION (f2->type)))
1366 /* Stabilize the sort. */
1367 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1370 /* We want the bigger accesses first, thus the opposite operator in the next
1372 return f1->size > f2->size ? -1 : 1;
1376 /* Append a name of the declaration to the name obstack. A helper function for
1380 make_fancy_decl_name (tree decl)
1384 tree name = DECL_NAME (decl);
1386 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1387 IDENTIFIER_LENGTH (name));
1390 sprintf (buffer, "D%u", DECL_UID (decl));
1391 obstack_grow (&name_obstack, buffer, strlen (buffer));
1395 /* Helper for make_fancy_name. */
1398 make_fancy_name_1 (tree expr)
1405 make_fancy_decl_name (expr);
1409 switch (TREE_CODE (expr))
1412 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1413 obstack_1grow (&name_obstack, '$');
1414 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1418 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1419 obstack_1grow (&name_obstack, '$');
1420 /* Arrays with only one element may not have a constant as their
1422 index = TREE_OPERAND (expr, 1);
1423 if (TREE_CODE (index) != INTEGER_CST)
1425 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1426 obstack_grow (&name_obstack, buffer, strlen (buffer));
1430 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1434 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1435 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1437 obstack_1grow (&name_obstack, '$');
1438 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1439 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1440 obstack_grow (&name_obstack, buffer, strlen (buffer));
1447 gcc_unreachable (); /* we treat these as scalars. */
1454 /* Create a human readable name for replacement variable of ACCESS. */
1457 make_fancy_name (tree expr)
1459 make_fancy_name_1 (expr);
1460 obstack_1grow (&name_obstack, '\0');
1461 return XOBFINISH (&name_obstack, char *);
1464 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1465 EXP_TYPE at the given OFFSET. If BASE is something for which
1466 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1467 to insert new statements either before or below the current one as specified
1468 by INSERT_AFTER. This function is not capable of handling bitfields. */
1471 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1472 tree exp_type, gimple_stmt_iterator *gsi,
1475 tree prev_base = base;
1477 HOST_WIDE_INT base_offset;
1478 unsigned HOST_WIDE_INT misalign;
1481 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1483 base = get_addr_base_and_unit_offset (base, &base_offset);
1485 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1486 offset such as array[var_index]. */
1492 gcc_checking_assert (gsi);
1493 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1494 add_referenced_var (tmp);
1495 tmp = make_ssa_name (tmp, NULL);
1496 addr = build_fold_addr_expr (unshare_expr (prev_base));
1497 STRIP_USELESS_TYPE_CONVERSION (addr);
1498 stmt = gimple_build_assign (tmp, addr);
1499 gimple_set_location (stmt, loc);
1500 SSA_NAME_DEF_STMT (tmp) = stmt;
1502 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1504 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1507 off = build_int_cst (reference_alias_ptr_type (prev_base),
1508 offset / BITS_PER_UNIT);
1511 else if (TREE_CODE (base) == MEM_REF)
1513 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1514 base_offset + offset / BITS_PER_UNIT);
1515 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1516 base = unshare_expr (TREE_OPERAND (base, 0));
1520 off = build_int_cst (reference_alias_ptr_type (base),
1521 base_offset + offset / BITS_PER_UNIT);
1522 base = build_fold_addr_expr (unshare_expr (base));
1525 /* If prev_base were always an originally performed access
1526 we can extract more optimistic alignment information
1527 by looking at the access mode. That would constrain the
1528 alignment of base + base_offset which we would need to
1529 adjust according to offset. */
1530 align = get_pointer_alignment_1 (base, &misalign);
1532 && (TREE_CODE (prev_base) == MEM_REF
1533 || TREE_CODE (prev_base) == TARGET_MEM_REF))
1534 align = MAX (align, TYPE_ALIGN (TREE_TYPE (prev_base)));
1535 misalign += (double_int_sext (tree_to_double_int (off),
1536 TYPE_PRECISION (TREE_TYPE (off))).low
1538 misalign = misalign & (align - 1);
1540 align = (misalign & -misalign);
1541 if (align < TYPE_ALIGN (exp_type))
1542 exp_type = build_aligned_type (exp_type, align);
1544 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1547 DEF_VEC_ALLOC_P_STACK (tree);
1548 #define VEC_tree_stack_alloc(alloc) VEC_stack_alloc (tree, alloc)
1550 /* Construct a memory reference to a part of an aggregate BASE at the given
1551 OFFSET and of the type of MODEL. In case this is a chain of references
1552 to component, the function will replicate the chain of COMPONENT_REFs of
1553 the expression of MODEL to access it. GSI and INSERT_AFTER have the same
1554 meaning as in build_ref_for_offset. */
1557 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1558 struct access *model, gimple_stmt_iterator *gsi,
1561 tree type = model->type, t;
1562 VEC(tree,stack) *cr_stack = NULL;
1564 if (TREE_CODE (model->expr) == COMPONENT_REF)
1566 tree expr = model->expr;
1568 /* Create a stack of the COMPONENT_REFs so later we can walk them in
1569 order from inner to outer. */
1570 cr_stack = VEC_alloc (tree, stack, 6);
1573 tree field = TREE_OPERAND (expr, 1);
1574 tree cr_offset = component_ref_field_offset (expr);
1575 HOST_WIDE_INT bit_pos
1576 = tree_low_cst (cr_offset, 1) * BITS_PER_UNIT
1577 + TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (field));
1579 /* We can be called with a model different from the one associated
1580 with BASE so we need to avoid going up the chain too far. */
1581 if (offset - bit_pos < 0)
1585 VEC_safe_push (tree, stack, cr_stack, expr);
1587 expr = TREE_OPERAND (expr, 0);
1588 type = TREE_TYPE (expr);
1589 } while (TREE_CODE (expr) == COMPONENT_REF);
1592 t = build_ref_for_offset (loc, base, offset, type, gsi, insert_after);
1594 if (TREE_CODE (model->expr) == COMPONENT_REF)
1599 /* Now replicate the chain of COMPONENT_REFs from inner to outer. */
1600 FOR_EACH_VEC_ELT_REVERSE (tree, cr_stack, i, expr)
1602 tree field = TREE_OPERAND (expr, 1);
1603 t = fold_build3_loc (loc, COMPONENT_REF, TREE_TYPE (field), t, field,
1604 TREE_OPERAND (expr, 2));
1607 VEC_free (tree, stack, cr_stack);
1613 /* Construct a memory reference consisting of component_refs and array_refs to
1614 a part of an aggregate *RES (which is of type TYPE). The requested part
1615 should have type EXP_TYPE at be the given OFFSET. This function might not
1616 succeed, it returns true when it does and only then *RES points to something
1617 meaningful. This function should be used only to build expressions that we
1618 might need to present to user (e.g. in warnings). In all other situations,
1619 build_ref_for_model or build_ref_for_offset should be used instead. */
1622 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1628 tree tr_size, index, minidx;
1629 HOST_WIDE_INT el_size;
1631 if (offset == 0 && exp_type
1632 && types_compatible_p (exp_type, type))
1635 switch (TREE_CODE (type))
1638 case QUAL_UNION_TYPE:
1640 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1642 HOST_WIDE_INT pos, size;
1643 tree expr, *expr_ptr;
1645 if (TREE_CODE (fld) != FIELD_DECL)
1648 pos = int_bit_position (fld);
1649 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1650 tr_size = DECL_SIZE (fld);
1651 if (!tr_size || !host_integerp (tr_size, 1))
1653 size = tree_low_cst (tr_size, 1);
1659 else if (pos > offset || (pos + size) <= offset)
1662 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1665 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1666 offset - pos, exp_type))
1675 tr_size = TYPE_SIZE (TREE_TYPE (type));
1676 if (!tr_size || !host_integerp (tr_size, 1))
1678 el_size = tree_low_cst (tr_size, 1);
1680 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1681 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1683 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1684 if (!integer_zerop (minidx))
1685 index = int_const_binop (PLUS_EXPR, index, minidx);
1686 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1687 NULL_TREE, NULL_TREE);
1688 offset = offset % el_size;
1689 type = TREE_TYPE (type);
1704 /* Return true iff TYPE is stdarg va_list type. */
1707 is_va_list_type (tree type)
1709 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1712 /* Print message to dump file why a variable was rejected. */
1715 reject (tree var, const char *msg)
1717 if (dump_file && (dump_flags & TDF_DETAILS))
1719 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1720 print_generic_expr (dump_file, var, 0);
1721 fprintf (dump_file, "\n");
1725 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1726 those with type which is suitable for scalarization. */
1729 find_var_candidates (void)
1732 referenced_var_iterator rvi;
1736 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
1738 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1740 type = TREE_TYPE (var);
1742 if (!AGGREGATE_TYPE_P (type))
1744 reject (var, "not aggregate");
1747 if (needs_to_live_in_memory (var))
1749 reject (var, "needs to live in memory");
1752 if (TREE_THIS_VOLATILE (var))
1754 reject (var, "is volatile");
1757 if (!COMPLETE_TYPE_P (type))
1759 reject (var, "has incomplete type");
1762 if (!host_integerp (TYPE_SIZE (type), 1))
1764 reject (var, "type size not fixed");
1767 if (tree_low_cst (TYPE_SIZE (type), 1) == 0)
1769 reject (var, "type size is zero");
1772 if (type_internals_preclude_sra_p (type, &msg))
1777 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1778 we also want to schedule it rather late. Thus we ignore it in
1780 (sra_mode == SRA_MODE_EARLY_INTRA
1781 && is_va_list_type (type)))
1783 reject (var, "is va_list");
1787 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1789 if (dump_file && (dump_flags & TDF_DETAILS))
1791 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1792 print_generic_expr (dump_file, var, 0);
1793 fprintf (dump_file, "\n");
1801 /* Sort all accesses for the given variable, check for partial overlaps and
1802 return NULL if there are any. If there are none, pick a representative for
1803 each combination of offset and size and create a linked list out of them.
1804 Return the pointer to the first representative and make sure it is the first
1805 one in the vector of accesses. */
1807 static struct access *
1808 sort_and_splice_var_accesses (tree var)
1810 int i, j, access_count;
1811 struct access *res, **prev_acc_ptr = &res;
1812 VEC (access_p, heap) *access_vec;
1814 HOST_WIDE_INT low = -1, high = 0;
1816 access_vec = get_base_access_vector (var);
1819 access_count = VEC_length (access_p, access_vec);
1821 /* Sort by <OFFSET, SIZE>. */
1822 VEC_qsort (access_p, access_vec, compare_access_positions);
1825 while (i < access_count)
1827 struct access *access = VEC_index (access_p, access_vec, i);
1828 bool grp_write = access->write;
1829 bool grp_read = !access->write;
1830 bool grp_scalar_write = access->write
1831 && is_gimple_reg_type (access->type);
1832 bool grp_scalar_read = !access->write
1833 && is_gimple_reg_type (access->type);
1834 bool grp_assignment_read = access->grp_assignment_read;
1835 bool grp_assignment_write = access->grp_assignment_write;
1836 bool multiple_scalar_reads = false;
1837 bool total_scalarization = access->grp_total_scalarization;
1838 bool grp_partial_lhs = access->grp_partial_lhs;
1839 bool first_scalar = is_gimple_reg_type (access->type);
1840 bool unscalarizable_region = access->grp_unscalarizable_region;
1842 if (first || access->offset >= high)
1845 low = access->offset;
1846 high = access->offset + access->size;
1848 else if (access->offset > low && access->offset + access->size > high)
1851 gcc_assert (access->offset >= low
1852 && access->offset + access->size <= high);
1855 while (j < access_count)
1857 struct access *ac2 = VEC_index (access_p, access_vec, j);
1858 if (ac2->offset != access->offset || ac2->size != access->size)
1863 grp_scalar_write = (grp_scalar_write
1864 || is_gimple_reg_type (ac2->type));
1869 if (is_gimple_reg_type (ac2->type))
1871 if (grp_scalar_read)
1872 multiple_scalar_reads = true;
1874 grp_scalar_read = true;
1877 grp_assignment_read |= ac2->grp_assignment_read;
1878 grp_assignment_write |= ac2->grp_assignment_write;
1879 grp_partial_lhs |= ac2->grp_partial_lhs;
1880 unscalarizable_region |= ac2->grp_unscalarizable_region;
1881 total_scalarization |= ac2->grp_total_scalarization;
1882 relink_to_new_repr (access, ac2);
1884 /* If there are both aggregate-type and scalar-type accesses with
1885 this combination of size and offset, the comparison function
1886 should have put the scalars first. */
1887 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1888 ac2->group_representative = access;
1894 access->group_representative = access;
1895 access->grp_write = grp_write;
1896 access->grp_read = grp_read;
1897 access->grp_scalar_read = grp_scalar_read;
1898 access->grp_scalar_write = grp_scalar_write;
1899 access->grp_assignment_read = grp_assignment_read;
1900 access->grp_assignment_write = grp_assignment_write;
1901 access->grp_hint = multiple_scalar_reads || total_scalarization;
1902 access->grp_total_scalarization = total_scalarization;
1903 access->grp_partial_lhs = grp_partial_lhs;
1904 access->grp_unscalarizable_region = unscalarizable_region;
1905 if (access->first_link)
1906 add_access_to_work_queue (access);
1908 *prev_acc_ptr = access;
1909 prev_acc_ptr = &access->next_grp;
1912 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1916 /* Create a variable for the given ACCESS which determines the type, name and a
1917 few other properties. Return the variable declaration and store it also to
1918 ACCESS->replacement. */
1921 create_access_replacement (struct access *access, bool rename)
1925 repl = create_tmp_var (access->type, "SR");
1926 add_referenced_var (repl);
1928 mark_sym_for_renaming (repl);
1930 if (!access->grp_partial_lhs
1931 && (TREE_CODE (access->type) == COMPLEX_TYPE
1932 || TREE_CODE (access->type) == VECTOR_TYPE))
1933 DECL_GIMPLE_REG_P (repl) = 1;
1935 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1936 DECL_ARTIFICIAL (repl) = 1;
1937 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1939 if (DECL_NAME (access->base)
1940 && !DECL_IGNORED_P (access->base)
1941 && !DECL_ARTIFICIAL (access->base))
1943 char *pretty_name = make_fancy_name (access->expr);
1944 tree debug_expr = unshare_expr (access->expr), d;
1946 DECL_NAME (repl) = get_identifier (pretty_name);
1947 obstack_free (&name_obstack, pretty_name);
1949 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1950 as DECL_DEBUG_EXPR isn't considered when looking for still
1951 used SSA_NAMEs and thus they could be freed. All debug info
1952 generation cares is whether something is constant or variable
1953 and that get_ref_base_and_extent works properly on the
1955 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1956 switch (TREE_CODE (d))
1959 case ARRAY_RANGE_REF:
1960 if (TREE_OPERAND (d, 1)
1961 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1962 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1963 if (TREE_OPERAND (d, 3)
1964 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1965 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1968 if (TREE_OPERAND (d, 2)
1969 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1970 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1975 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1976 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1977 if (access->grp_no_warning)
1978 TREE_NO_WARNING (repl) = 1;
1980 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1983 TREE_NO_WARNING (repl) = 1;
1987 fprintf (dump_file, "Created a replacement for ");
1988 print_generic_expr (dump_file, access->base, 0);
1989 fprintf (dump_file, " offset: %u, size: %u: ",
1990 (unsigned) access->offset, (unsigned) access->size);
1991 print_generic_expr (dump_file, repl, 0);
1992 fprintf (dump_file, "\n");
1994 sra_stats.replacements++;
1999 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
2002 get_access_replacement (struct access *access)
2004 gcc_assert (access->grp_to_be_replaced);
2006 if (!access->replacement_decl)
2007 access->replacement_decl = create_access_replacement (access, true);
2008 return access->replacement_decl;
2011 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
2012 not mark it for renaming. */
2015 get_unrenamed_access_replacement (struct access *access)
2017 gcc_assert (!access->grp_to_be_replaced);
2019 if (!access->replacement_decl)
2020 access->replacement_decl = create_access_replacement (access, false);
2021 return access->replacement_decl;
2025 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
2026 linked list along the way. Stop when *ACCESS is NULL or the access pointed
2027 to it is not "within" the root. Return false iff some accesses partially
2031 build_access_subtree (struct access **access)
2033 struct access *root = *access, *last_child = NULL;
2034 HOST_WIDE_INT limit = root->offset + root->size;
2036 *access = (*access)->next_grp;
2037 while (*access && (*access)->offset + (*access)->size <= limit)
2040 root->first_child = *access;
2042 last_child->next_sibling = *access;
2043 last_child = *access;
2045 if (!build_access_subtree (access))
2049 if (*access && (*access)->offset < limit)
2055 /* Build a tree of access representatives, ACCESS is the pointer to the first
2056 one, others are linked in a list by the next_grp field. Return false iff
2057 some accesses partially overlap. */
2060 build_access_trees (struct access *access)
2064 struct access *root = access;
2066 if (!build_access_subtree (&access))
2068 root->next_grp = access;
2073 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2077 expr_with_var_bounded_array_refs_p (tree expr)
2079 while (handled_component_p (expr))
2081 if (TREE_CODE (expr) == ARRAY_REF
2082 && !host_integerp (array_ref_low_bound (expr), 0))
2084 expr = TREE_OPERAND (expr, 0);
2089 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2090 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2091 sorts of access flags appropriately along the way, notably always set
2092 grp_read and grp_assign_read according to MARK_READ and grp_write when
2095 Creating a replacement for a scalar access is considered beneficial if its
2096 grp_hint is set (this means we are either attempting total scalarization or
2097 there is more than one direct read access) or according to the following
2100 Access written to through a scalar type (once or more times)
2102 | Written to in an assignment statement
2104 | | Access read as scalar _once_
2106 | | | Read in an assignment statement
2108 | | | | Scalarize Comment
2109 -----------------------------------------------------------------------------
2110 0 0 0 0 No access for the scalar
2111 0 0 0 1 No access for the scalar
2112 0 0 1 0 No Single read - won't help
2113 0 0 1 1 No The same case
2114 0 1 0 0 No access for the scalar
2115 0 1 0 1 No access for the scalar
2116 0 1 1 0 Yes s = *g; return s.i;
2117 0 1 1 1 Yes The same case as above
2118 1 0 0 0 No Won't help
2119 1 0 0 1 Yes s.i = 1; *g = s;
2120 1 0 1 0 Yes s.i = 5; g = s.i;
2121 1 0 1 1 Yes The same case as above
2122 1 1 0 0 No Won't help.
2123 1 1 0 1 Yes s.i = 1; *g = s;
2124 1 1 1 0 Yes s = *g; return s.i;
2125 1 1 1 1 Yes Any of the above yeses */
2128 analyze_access_subtree (struct access *root, struct access *parent,
2129 bool allow_replacements)
2131 struct access *child;
2132 HOST_WIDE_INT limit = root->offset + root->size;
2133 HOST_WIDE_INT covered_to = root->offset;
2134 bool scalar = is_gimple_reg_type (root->type);
2135 bool hole = false, sth_created = false;
2139 if (parent->grp_read)
2141 if (parent->grp_assignment_read)
2142 root->grp_assignment_read = 1;
2143 if (parent->grp_write)
2144 root->grp_write = 1;
2145 if (parent->grp_assignment_write)
2146 root->grp_assignment_write = 1;
2147 if (parent->grp_total_scalarization)
2148 root->grp_total_scalarization = 1;
2151 if (root->grp_unscalarizable_region)
2152 allow_replacements = false;
2154 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2155 allow_replacements = false;
2157 for (child = root->first_child; child; child = child->next_sibling)
2159 hole |= covered_to < child->offset;
2160 sth_created |= analyze_access_subtree (child, root,
2161 allow_replacements && !scalar);
2163 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2164 root->grp_total_scalarization &= child->grp_total_scalarization;
2165 if (child->grp_covered)
2166 covered_to += child->size;
2171 if (allow_replacements && scalar && !root->first_child
2173 || ((root->grp_scalar_read || root->grp_assignment_read)
2174 && (root->grp_scalar_write || root->grp_assignment_write))))
2176 bool new_integer_type;
2177 /* Always create access replacements that cover the whole access.
2178 For integral types this means the precision has to match.
2179 Avoid assumptions based on the integral type kind, too. */
2180 if (INTEGRAL_TYPE_P (root->type)
2181 && (TREE_CODE (root->type) != INTEGER_TYPE
2182 || TYPE_PRECISION (root->type) != root->size)
2183 /* But leave bitfield accesses alone. */
2184 && (root->offset % BITS_PER_UNIT) == 0)
2186 tree rt = root->type;
2187 root->type = build_nonstandard_integer_type (root->size,
2188 TYPE_UNSIGNED (rt));
2189 root->expr = build_ref_for_offset (UNKNOWN_LOCATION,
2190 root->base, root->offset,
2191 root->type, NULL, false);
2192 new_integer_type = true;
2195 new_integer_type = false;
2197 if (dump_file && (dump_flags & TDF_DETAILS))
2199 fprintf (dump_file, "Marking ");
2200 print_generic_expr (dump_file, root->base, 0);
2201 fprintf (dump_file, " offset: %u, size: %u ",
2202 (unsigned) root->offset, (unsigned) root->size);
2203 fprintf (dump_file, " to be replaced%s.\n",
2204 new_integer_type ? " with an integer": "");
2207 root->grp_to_be_replaced = 1;
2213 if (covered_to < limit)
2216 root->grp_total_scalarization = 0;
2220 && (!hole || root->grp_total_scalarization))
2222 root->grp_covered = 1;
2225 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2226 root->grp_unscalarized_data = 1; /* not covered and written to */
2232 /* Analyze all access trees linked by next_grp by the means of
2233 analyze_access_subtree. */
2235 analyze_access_trees (struct access *access)
2241 if (analyze_access_subtree (access, NULL, true))
2243 access = access->next_grp;
2249 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2250 SIZE would conflict with an already existing one. If exactly such a child
2251 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2254 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2255 HOST_WIDE_INT size, struct access **exact_match)
2257 struct access *child;
2259 for (child = lacc->first_child; child; child = child->next_sibling)
2261 if (child->offset == norm_offset && child->size == size)
2263 *exact_match = child;
2267 if (child->offset < norm_offset + size
2268 && child->offset + child->size > norm_offset)
2275 /* Create a new child access of PARENT, with all properties just like MODEL
2276 except for its offset and with its grp_write false and grp_read true.
2277 Return the new access or NULL if it cannot be created. Note that this access
2278 is created long after all splicing and sorting, it's not located in any
2279 access vector and is automatically a representative of its group. */
2281 static struct access *
2282 create_artificial_child_access (struct access *parent, struct access *model,
2283 HOST_WIDE_INT new_offset)
2285 struct access *access;
2286 struct access **child;
2287 tree expr = parent->base;
2289 gcc_assert (!model->grp_unscalarizable_region);
2291 access = (struct access *) pool_alloc (access_pool);
2292 memset (access, 0, sizeof (struct access));
2293 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2296 access->grp_no_warning = true;
2297 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2298 new_offset, model, NULL, false);
2301 access->base = parent->base;
2302 access->expr = expr;
2303 access->offset = new_offset;
2304 access->size = model->size;
2305 access->type = model->type;
2306 access->grp_write = true;
2307 access->grp_read = false;
2309 child = &parent->first_child;
2310 while (*child && (*child)->offset < new_offset)
2311 child = &(*child)->next_sibling;
2313 access->next_sibling = *child;
2320 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2321 true if any new subaccess was created. Additionally, if RACC is a scalar
2322 access but LACC is not, change the type of the latter, if possible. */
2325 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2327 struct access *rchild;
2328 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2331 if (is_gimple_reg_type (lacc->type)
2332 || lacc->grp_unscalarizable_region
2333 || racc->grp_unscalarizable_region)
2336 if (is_gimple_reg_type (racc->type))
2338 if (!lacc->first_child && !racc->first_child)
2340 tree t = lacc->base;
2342 lacc->type = racc->type;
2343 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t),
2344 lacc->offset, racc->type))
2348 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2349 lacc->base, lacc->offset,
2351 lacc->grp_no_warning = true;
2357 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2359 struct access *new_acc = NULL;
2360 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2362 if (rchild->grp_unscalarizable_region)
2365 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2370 rchild->grp_hint = 1;
2371 new_acc->grp_hint |= new_acc->grp_read;
2372 if (rchild->first_child)
2373 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2378 rchild->grp_hint = 1;
2379 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2383 if (racc->first_child)
2384 propagate_subaccesses_across_link (new_acc, rchild);
2391 /* Propagate all subaccesses across assignment links. */
2394 propagate_all_subaccesses (void)
2396 while (work_queue_head)
2398 struct access *racc = pop_access_from_work_queue ();
2399 struct assign_link *link;
2401 gcc_assert (racc->first_link);
2403 for (link = racc->first_link; link; link = link->next)
2405 struct access *lacc = link->lacc;
2407 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2409 lacc = lacc->group_representative;
2410 if (propagate_subaccesses_across_link (lacc, racc)
2411 && lacc->first_link)
2412 add_access_to_work_queue (lacc);
2417 /* Go through all accesses collected throughout the (intraprocedural) analysis
2418 stage, exclude overlapping ones, identify representatives and build trees
2419 out of them, making decisions about scalarization on the way. Return true
2420 iff there are any to-be-scalarized variables after this stage. */
2423 analyze_all_variable_accesses (void)
2426 bitmap tmp = BITMAP_ALLOC (NULL);
2428 unsigned i, max_total_scalarization_size;
2430 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2431 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2433 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2434 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2435 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2437 tree var = referenced_var (i);
2439 if (TREE_CODE (var) == VAR_DECL
2440 && type_consists_of_records_p (TREE_TYPE (var)))
2442 if ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2443 <= max_total_scalarization_size)
2445 completely_scalarize_var (var);
2446 if (dump_file && (dump_flags & TDF_DETAILS))
2448 fprintf (dump_file, "Will attempt to totally scalarize ");
2449 print_generic_expr (dump_file, var, 0);
2450 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2453 else if (dump_file && (dump_flags & TDF_DETAILS))
2455 fprintf (dump_file, "Too big to totally scalarize: ");
2456 print_generic_expr (dump_file, var, 0);
2457 fprintf (dump_file, " (UID: %u)\n", DECL_UID (var));
2462 bitmap_copy (tmp, candidate_bitmap);
2463 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2465 tree var = referenced_var (i);
2466 struct access *access;
2468 access = sort_and_splice_var_accesses (var);
2469 if (!access || !build_access_trees (access))
2470 disqualify_candidate (var,
2471 "No or inhibitingly overlapping accesses.");
2474 propagate_all_subaccesses ();
2476 bitmap_copy (tmp, candidate_bitmap);
2477 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2479 tree var = referenced_var (i);
2480 struct access *access = get_first_repr_for_decl (var);
2482 if (analyze_access_trees (access))
2485 if (dump_file && (dump_flags & TDF_DETAILS))
2487 fprintf (dump_file, "\nAccess trees for ");
2488 print_generic_expr (dump_file, var, 0);
2489 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2490 dump_access_tree (dump_file, access);
2491 fprintf (dump_file, "\n");
2495 disqualify_candidate (var, "No scalar replacements to be created.");
2502 statistics_counter_event (cfun, "Scalarized aggregates", res);
2509 /* Generate statements copying scalar replacements of accesses within a subtree
2510 into or out of AGG. ACCESS, all its children, siblings and their children
2511 are to be processed. AGG is an aggregate type expression (can be a
2512 declaration but does not have to be, it can for example also be a mem_ref or
2513 a series of handled components). TOP_OFFSET is the offset of the processed
2514 subtree which has to be subtracted from offsets of individual accesses to
2515 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2516 replacements in the interval <start_offset, start_offset + chunk_size>,
2517 otherwise copy all. GSI is a statement iterator used to place the new
2518 statements. WRITE should be true when the statements should write from AGG
2519 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2520 statements will be added after the current statement in GSI, they will be
2521 added before the statement otherwise. */
2524 generate_subtree_copies (struct access *access, tree agg,
2525 HOST_WIDE_INT top_offset,
2526 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2527 gimple_stmt_iterator *gsi, bool write,
2528 bool insert_after, location_t loc)
2532 if (chunk_size && access->offset >= start_offset + chunk_size)
2535 if (access->grp_to_be_replaced
2537 || access->offset + access->size > start_offset))
2539 tree expr, repl = get_access_replacement (access);
2542 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2543 access, gsi, insert_after);
2547 if (access->grp_partial_lhs)
2548 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2550 insert_after ? GSI_NEW_STMT
2552 stmt = gimple_build_assign (repl, expr);
2556 TREE_NO_WARNING (repl) = 1;
2557 if (access->grp_partial_lhs)
2558 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2560 insert_after ? GSI_NEW_STMT
2562 stmt = gimple_build_assign (expr, repl);
2564 gimple_set_location (stmt, loc);
2567 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2569 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2571 sra_stats.subtree_copies++;
2574 if (access->first_child)
2575 generate_subtree_copies (access->first_child, agg, top_offset,
2576 start_offset, chunk_size, gsi,
2577 write, insert_after, loc);
2579 access = access->next_sibling;
2584 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2585 the root of the subtree to be processed. GSI is the statement iterator used
2586 for inserting statements which are added after the current statement if
2587 INSERT_AFTER is true or before it otherwise. */
2590 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2591 bool insert_after, location_t loc)
2594 struct access *child;
2596 if (access->grp_to_be_replaced)
2600 stmt = gimple_build_assign (get_access_replacement (access),
2601 build_zero_cst (access->type));
2603 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2605 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2607 gimple_set_location (stmt, loc);
2610 for (child = access->first_child; child; child = child->next_sibling)
2611 init_subtree_with_zero (child, gsi, insert_after, loc);
2614 /* Search for an access representative for the given expression EXPR and
2615 return it or NULL if it cannot be found. */
2617 static struct access *
2618 get_access_for_expr (tree expr)
2620 HOST_WIDE_INT offset, size, max_size;
2623 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2624 a different size than the size of its argument and we need the latter
2626 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2627 expr = TREE_OPERAND (expr, 0);
2629 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2630 if (max_size == -1 || !DECL_P (base))
2633 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2636 return get_var_base_offset_size_access (base, offset, max_size);
2639 /* Replace the expression EXPR with a scalar replacement if there is one and
2640 generate other statements to do type conversion or subtree copying if
2641 necessary. GSI is used to place newly created statements, WRITE is true if
2642 the expression is being written to (it is on a LHS of a statement or output
2643 in an assembly statement). */
2646 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2649 struct access *access;
2652 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2655 expr = &TREE_OPERAND (*expr, 0);
2660 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2661 expr = &TREE_OPERAND (*expr, 0);
2662 access = get_access_for_expr (*expr);
2665 type = TREE_TYPE (*expr);
2667 loc = gimple_location (gsi_stmt (*gsi));
2668 if (access->grp_to_be_replaced)
2670 tree repl = get_access_replacement (access);
2671 /* If we replace a non-register typed access simply use the original
2672 access expression to extract the scalar component afterwards.
2673 This happens if scalarizing a function return value or parameter
2674 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2675 gcc.c-torture/compile/20011217-1.c.
2677 We also want to use this when accessing a complex or vector which can
2678 be accessed as a different type too, potentially creating a need for
2679 type conversion (see PR42196) and when scalarized unions are involved
2680 in assembler statements (see PR42398). */
2681 if (!useless_type_conversion_p (type, access->type))
2685 ref = build_ref_for_model (loc, access->base, access->offset, access,
2692 if (access->grp_partial_lhs)
2693 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2694 false, GSI_NEW_STMT);
2695 stmt = gimple_build_assign (repl, ref);
2696 gimple_set_location (stmt, loc);
2697 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2703 if (access->grp_partial_lhs)
2704 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2705 true, GSI_SAME_STMT);
2706 stmt = gimple_build_assign (ref, repl);
2707 gimple_set_location (stmt, loc);
2708 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2716 if (access->first_child)
2718 HOST_WIDE_INT start_offset, chunk_size;
2720 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2721 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2723 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2724 start_offset = access->offset
2725 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2728 start_offset = chunk_size = 0;
2730 generate_subtree_copies (access->first_child, access->base, 0,
2731 start_offset, chunk_size, gsi, write, write,
2737 /* Where scalar replacements of the RHS have been written to when a replacement
2738 of a LHS of an assigments cannot be direclty loaded from a replacement of
2740 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2741 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2742 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2744 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2745 base aggregate if there are unscalarized data or directly to LHS of the
2746 statement that is pointed to by GSI otherwise. */
2748 static enum unscalarized_data_handling
2749 handle_unscalarized_data_in_subtree (struct access *top_racc,
2750 gimple_stmt_iterator *gsi)
2752 if (top_racc->grp_unscalarized_data)
2754 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2756 gimple_location (gsi_stmt (*gsi)));
2757 return SRA_UDH_RIGHT;
2761 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2762 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2763 0, 0, gsi, false, false,
2764 gimple_location (gsi_stmt (*gsi)));
2765 return SRA_UDH_LEFT;
2770 /* Try to generate statements to load all sub-replacements in an access subtree
2771 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2772 If that is not possible, refresh the TOP_RACC base aggregate and load the
2773 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2774 copied. NEW_GSI is stmt iterator used for statement insertions after the
2775 original assignment, OLD_GSI is used to insert statements before the
2776 assignment. *REFRESHED keeps the information whether we have needed to
2777 refresh replacements of the LHS and from which side of the assignments this
2781 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2782 HOST_WIDE_INT left_offset,
2783 gimple_stmt_iterator *old_gsi,
2784 gimple_stmt_iterator *new_gsi,
2785 enum unscalarized_data_handling *refreshed)
2787 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2788 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2790 if (lacc->grp_to_be_replaced)
2792 struct access *racc;
2793 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2797 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2798 if (racc && racc->grp_to_be_replaced)
2800 rhs = get_access_replacement (racc);
2801 if (!useless_type_conversion_p (lacc->type, racc->type))
2802 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2804 if (racc->grp_partial_lhs && lacc->grp_partial_lhs)
2805 rhs = force_gimple_operand_gsi (old_gsi, rhs, true, NULL_TREE,
2806 true, GSI_SAME_STMT);
2810 /* No suitable access on the right hand side, need to load from
2811 the aggregate. See if we have to update it first... */
2812 if (*refreshed == SRA_UDH_NONE)
2813 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2816 if (*refreshed == SRA_UDH_LEFT)
2817 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2820 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2822 if (lacc->grp_partial_lhs)
2823 rhs = force_gimple_operand_gsi (new_gsi, rhs, true, NULL_TREE,
2824 false, GSI_NEW_STMT);
2827 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2828 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2829 gimple_set_location (stmt, loc);
2831 sra_stats.subreplacements++;
2833 else if (*refreshed == SRA_UDH_NONE
2834 && lacc->grp_read && !lacc->grp_covered)
2835 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2838 if (lacc->first_child)
2839 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2840 old_gsi, new_gsi, refreshed);
2844 /* Result code for SRA assignment modification. */
2845 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2846 SRA_AM_MODIFIED, /* stmt changed but not
2848 SRA_AM_REMOVED }; /* stmt eliminated */
2850 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2851 to the assignment and GSI is the statement iterator pointing at it. Returns
2852 the same values as sra_modify_assign. */
2854 static enum assignment_mod_result
2855 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2857 tree lhs = gimple_assign_lhs (*stmt);
2861 acc = get_access_for_expr (lhs);
2865 if (gimple_clobber_p (*stmt))
2867 /* Remove clobbers of fully scalarized variables, otherwise
2869 if (acc->grp_covered)
2871 unlink_stmt_vdef (*stmt);
2872 gsi_remove (gsi, true);
2873 return SRA_AM_REMOVED;
2879 loc = gimple_location (*stmt);
2880 if (VEC_length (constructor_elt,
2881 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2883 /* I have never seen this code path trigger but if it can happen the
2884 following should handle it gracefully. */
2885 if (access_has_children_p (acc))
2886 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2888 return SRA_AM_MODIFIED;
2891 if (acc->grp_covered)
2893 init_subtree_with_zero (acc, gsi, false, loc);
2894 unlink_stmt_vdef (*stmt);
2895 gsi_remove (gsi, true);
2896 return SRA_AM_REMOVED;
2900 init_subtree_with_zero (acc, gsi, true, loc);
2901 return SRA_AM_MODIFIED;
2905 /* Create and return a new suitable default definition SSA_NAME for RACC which
2906 is an access describing an uninitialized part of an aggregate that is being
2910 get_repl_default_def_ssa_name (struct access *racc)
2914 decl = get_unrenamed_access_replacement (racc);
2916 repl = gimple_default_def (cfun, decl);
2919 repl = make_ssa_name (decl, gimple_build_nop ());
2920 set_default_def (decl, repl);
2926 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2930 contains_bitfld_comp_ref_p (const_tree ref)
2932 while (handled_component_p (ref))
2934 if (TREE_CODE (ref) == COMPONENT_REF
2935 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2937 ref = TREE_OPERAND (ref, 0);
2943 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2944 bit-field field declaration somewhere in it. */
2947 contains_vce_or_bfcref_p (const_tree ref)
2949 while (handled_component_p (ref))
2951 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2952 || (TREE_CODE (ref) == COMPONENT_REF
2953 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2955 ref = TREE_OPERAND (ref, 0);
2961 /* Examine both sides of the assignment statement pointed to by STMT, replace
2962 them with a scalare replacement if there is one and generate copying of
2963 replacements if scalarized aggregates have been used in the assignment. GSI
2964 is used to hold generated statements for type conversions and subtree
2967 static enum assignment_mod_result
2968 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2970 struct access *lacc, *racc;
2972 bool modify_this_stmt = false;
2973 bool force_gimple_rhs = false;
2975 gimple_stmt_iterator orig_gsi = *gsi;
2977 if (!gimple_assign_single_p (*stmt))
2979 lhs = gimple_assign_lhs (*stmt);
2980 rhs = gimple_assign_rhs1 (*stmt);
2982 if (TREE_CODE (rhs) == CONSTRUCTOR)
2983 return sra_modify_constructor_assign (stmt, gsi);
2985 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2986 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2987 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2989 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2991 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2993 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2996 lacc = get_access_for_expr (lhs);
2997 racc = get_access_for_expr (rhs);
3001 loc = gimple_location (*stmt);
3002 if (lacc && lacc->grp_to_be_replaced)
3004 lhs = get_access_replacement (lacc);
3005 gimple_assign_set_lhs (*stmt, lhs);
3006 modify_this_stmt = true;
3007 if (lacc->grp_partial_lhs)
3008 force_gimple_rhs = true;
3012 if (racc && racc->grp_to_be_replaced)
3014 rhs = get_access_replacement (racc);
3015 modify_this_stmt = true;
3016 if (racc->grp_partial_lhs)
3017 force_gimple_rhs = true;
3021 && !racc->grp_unscalarized_data
3022 && TREE_CODE (lhs) == SSA_NAME
3023 && !access_has_replacements_p (racc))
3025 rhs = get_repl_default_def_ssa_name (racc);
3026 modify_this_stmt = true;
3030 if (modify_this_stmt)
3032 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3034 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
3035 ??? This should move to fold_stmt which we simply should
3036 call after building a VIEW_CONVERT_EXPR here. */
3037 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
3038 && !contains_bitfld_comp_ref_p (lhs)
3039 && !access_has_children_p (lacc))
3041 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
3042 gimple_assign_set_lhs (*stmt, lhs);
3044 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
3045 && !contains_vce_or_bfcref_p (rhs)
3046 && !access_has_children_p (racc))
3047 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
3049 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3051 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
3053 if (is_gimple_reg_type (TREE_TYPE (lhs))
3054 && TREE_CODE (lhs) != SSA_NAME)
3055 force_gimple_rhs = true;
3060 /* From this point on, the function deals with assignments in between
3061 aggregates when at least one has scalar reductions of some of its
3062 components. There are three possible scenarios: Both the LHS and RHS have
3063 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
3065 In the first case, we would like to load the LHS components from RHS
3066 components whenever possible. If that is not possible, we would like to
3067 read it directly from the RHS (after updating it by storing in it its own
3068 components). If there are some necessary unscalarized data in the LHS,
3069 those will be loaded by the original assignment too. If neither of these
3070 cases happen, the original statement can be removed. Most of this is done
3071 by load_assign_lhs_subreplacements.
3073 In the second case, we would like to store all RHS scalarized components
3074 directly into LHS and if they cover the aggregate completely, remove the
3075 statement too. In the third case, we want the LHS components to be loaded
3076 directly from the RHS (DSE will remove the original statement if it
3079 This is a bit complex but manageable when types match and when unions do
3080 not cause confusion in a way that we cannot really load a component of LHS
3081 from the RHS or vice versa (the access representing this level can have
3082 subaccesses that are accessible only through a different union field at a
3083 higher level - different from the one used in the examined expression).
3086 Therefore, I specially handle a fourth case, happening when there is a
3087 specific type cast or it is impossible to locate a scalarized subaccess on
3088 the other side of the expression. If that happens, I simply "refresh" the
3089 RHS by storing in it is scalarized components leave the original statement
3090 there to do the copying and then load the scalar replacements of the LHS.
3091 This is what the first branch does. */
3093 if (modify_this_stmt
3094 || gimple_has_volatile_ops (*stmt)
3095 || contains_vce_or_bfcref_p (rhs)
3096 || contains_vce_or_bfcref_p (lhs))
3098 if (access_has_children_p (racc))
3099 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3100 gsi, false, false, loc);
3101 if (access_has_children_p (lacc))
3102 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
3103 gsi, true, true, loc);
3104 sra_stats.separate_lhs_rhs_handling++;
3106 /* This gimplification must be done after generate_subtree_copies,
3107 lest we insert the subtree copies in the middle of the gimplified
3109 if (force_gimple_rhs)
3110 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
3111 true, GSI_SAME_STMT);
3112 if (gimple_assign_rhs1 (*stmt) != rhs)
3114 modify_this_stmt = true;
3115 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
3116 gcc_assert (*stmt == gsi_stmt (orig_gsi));
3119 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3123 if (access_has_children_p (lacc) && access_has_children_p (racc))
3125 gimple_stmt_iterator orig_gsi = *gsi;
3126 enum unscalarized_data_handling refreshed;
3128 if (lacc->grp_read && !lacc->grp_covered)
3129 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
3131 refreshed = SRA_UDH_NONE;
3133 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
3134 &orig_gsi, gsi, &refreshed);
3135 if (refreshed != SRA_UDH_RIGHT)
3138 unlink_stmt_vdef (*stmt);
3139 gsi_remove (&orig_gsi, true);
3140 sra_stats.deleted++;
3141 return SRA_AM_REMOVED;
3146 if (access_has_children_p (racc)
3147 && !racc->grp_unscalarized_data)
3151 fprintf (dump_file, "Removing load: ");
3152 print_gimple_stmt (dump_file, *stmt, 0, 0);
3154 generate_subtree_copies (racc->first_child, lhs,
3155 racc->offset, 0, 0, gsi,
3157 gcc_assert (*stmt == gsi_stmt (*gsi));
3158 unlink_stmt_vdef (*stmt);
3159 gsi_remove (gsi, true);
3160 sra_stats.deleted++;
3161 return SRA_AM_REMOVED;
3163 /* Restore the aggregate RHS from its components so the
3164 prevailing aggregate copy does the right thing. */
3165 if (access_has_children_p (racc))
3166 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3167 gsi, false, false, loc);
3168 /* Re-load the components of the aggregate copy destination.
3169 But use the RHS aggregate to load from to expose more
3170 optimization opportunities. */
3171 if (access_has_children_p (lacc))
3172 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
3173 0, 0, gsi, true, true, loc);
3180 /* Traverse the function body and all modifications as decided in
3181 analyze_all_variable_accesses. Return true iff the CFG has been
3185 sra_modify_function_body (void)
3187 bool cfg_changed = false;
3192 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3193 while (!gsi_end_p (gsi))
3195 gimple stmt = gsi_stmt (gsi);
3196 enum assignment_mod_result assign_result;
3197 bool modified = false, deleted = false;
3201 switch (gimple_code (stmt))
3204 t = gimple_return_retval_ptr (stmt);
3205 if (*t != NULL_TREE)
3206 modified |= sra_modify_expr (t, &gsi, false);
3210 assign_result = sra_modify_assign (&stmt, &gsi);
3211 modified |= assign_result == SRA_AM_MODIFIED;
3212 deleted = assign_result == SRA_AM_REMOVED;
3216 /* Operands must be processed before the lhs. */
3217 for (i = 0; i < gimple_call_num_args (stmt); i++)
3219 t = gimple_call_arg_ptr (stmt, i);
3220 modified |= sra_modify_expr (t, &gsi, false);
3223 if (gimple_call_lhs (stmt))
3225 t = gimple_call_lhs_ptr (stmt);
3226 modified |= sra_modify_expr (t, &gsi, true);
3231 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3233 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3234 modified |= sra_modify_expr (t, &gsi, false);
3236 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3238 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3239 modified |= sra_modify_expr (t, &gsi, true);
3250 if (maybe_clean_eh_stmt (stmt)
3251 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3262 /* Generate statements initializing scalar replacements of parts of function
3266 initialize_parameter_reductions (void)
3268 gimple_stmt_iterator gsi;
3269 gimple_seq seq = NULL;
3272 for (parm = DECL_ARGUMENTS (current_function_decl);
3274 parm = DECL_CHAIN (parm))
3276 VEC (access_p, heap) *access_vec;
3277 struct access *access;
3279 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3281 access_vec = get_base_access_vector (parm);
3287 seq = gimple_seq_alloc ();
3288 gsi = gsi_start (seq);
3291 for (access = VEC_index (access_p, access_vec, 0);
3293 access = access->next_grp)
3294 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3295 EXPR_LOCATION (parm));
3299 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3302 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3303 it reveals there are components of some aggregates to be scalarized, it runs
3304 the required transformations. */
3306 perform_intra_sra (void)
3311 if (!find_var_candidates ())
3314 if (!scan_function ())
3317 if (!analyze_all_variable_accesses ())
3320 if (sra_modify_function_body ())
3321 ret = TODO_update_ssa | TODO_cleanup_cfg;
3323 ret = TODO_update_ssa;
3324 initialize_parameter_reductions ();
3326 statistics_counter_event (cfun, "Scalar replacements created",
3327 sra_stats.replacements);
3328 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3329 statistics_counter_event (cfun, "Subtree copy stmts",
3330 sra_stats.subtree_copies);
3331 statistics_counter_event (cfun, "Subreplacement stmts",
3332 sra_stats.subreplacements);
3333 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3334 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3335 sra_stats.separate_lhs_rhs_handling);
3338 sra_deinitialize ();
3342 /* Perform early intraprocedural SRA. */
3344 early_intra_sra (void)
3346 sra_mode = SRA_MODE_EARLY_INTRA;
3347 return perform_intra_sra ();
3350 /* Perform "late" intraprocedural SRA. */
3352 late_intra_sra (void)
3354 sra_mode = SRA_MODE_INTRA;
3355 return perform_intra_sra ();
3360 gate_intra_sra (void)
3362 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3366 struct gimple_opt_pass pass_sra_early =
3371 gate_intra_sra, /* gate */
3372 early_intra_sra, /* execute */
3375 0, /* static_pass_number */
3376 TV_TREE_SRA, /* tv_id */
3377 PROP_cfg | PROP_ssa, /* properties_required */
3378 0, /* properties_provided */
3379 0, /* properties_destroyed */
3380 0, /* todo_flags_start */
3383 | TODO_verify_ssa /* todo_flags_finish */
3387 struct gimple_opt_pass pass_sra =
3392 gate_intra_sra, /* gate */
3393 late_intra_sra, /* execute */
3396 0, /* static_pass_number */
3397 TV_TREE_SRA, /* tv_id */
3398 PROP_cfg | PROP_ssa, /* properties_required */
3399 0, /* properties_provided */
3400 0, /* properties_destroyed */
3401 TODO_update_address_taken, /* todo_flags_start */
3404 | TODO_verify_ssa /* todo_flags_finish */
3409 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3413 is_unused_scalar_param (tree parm)
3416 return (is_gimple_reg (parm)
3417 && (!(name = gimple_default_def (cfun, parm))
3418 || has_zero_uses (name)));
3421 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3422 examine whether there are any direct or otherwise infeasible ones. If so,
3423 return true, otherwise return false. PARM must be a gimple register with a
3424 non-NULL default definition. */
3427 ptr_parm_has_direct_uses (tree parm)
3429 imm_use_iterator ui;
3431 tree name = gimple_default_def (cfun, parm);
3434 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3437 use_operand_p use_p;
3439 if (is_gimple_debug (stmt))
3442 /* Valid uses include dereferences on the lhs and the rhs. */
3443 if (gimple_has_lhs (stmt))
3445 tree lhs = gimple_get_lhs (stmt);
3446 while (handled_component_p (lhs))
3447 lhs = TREE_OPERAND (lhs, 0);
3448 if (TREE_CODE (lhs) == MEM_REF
3449 && TREE_OPERAND (lhs, 0) == name
3450 && integer_zerop (TREE_OPERAND (lhs, 1))
3451 && types_compatible_p (TREE_TYPE (lhs),
3452 TREE_TYPE (TREE_TYPE (name)))
3453 && !TREE_THIS_VOLATILE (lhs))
3456 if (gimple_assign_single_p (stmt))
3458 tree rhs = gimple_assign_rhs1 (stmt);
3459 while (handled_component_p (rhs))
3460 rhs = TREE_OPERAND (rhs, 0);
3461 if (TREE_CODE (rhs) == MEM_REF
3462 && TREE_OPERAND (rhs, 0) == name
3463 && integer_zerop (TREE_OPERAND (rhs, 1))
3464 && types_compatible_p (TREE_TYPE (rhs),
3465 TREE_TYPE (TREE_TYPE (name)))
3466 && !TREE_THIS_VOLATILE (rhs))
3469 else if (is_gimple_call (stmt))
3472 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3474 tree arg = gimple_call_arg (stmt, i);
3475 while (handled_component_p (arg))
3476 arg = TREE_OPERAND (arg, 0);
3477 if (TREE_CODE (arg) == MEM_REF
3478 && TREE_OPERAND (arg, 0) == name
3479 && integer_zerop (TREE_OPERAND (arg, 1))
3480 && types_compatible_p (TREE_TYPE (arg),
3481 TREE_TYPE (TREE_TYPE (name)))
3482 && !TREE_THIS_VOLATILE (arg))
3487 /* If the number of valid uses does not match the number of
3488 uses in this stmt there is an unhandled use. */
3489 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3496 BREAK_FROM_IMM_USE_STMT (ui);
3502 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3503 them in candidate_bitmap. Note that these do not necessarily include
3504 parameter which are unused and thus can be removed. Return true iff any
3505 such candidate has been found. */
3508 find_param_candidates (void)
3515 for (parm = DECL_ARGUMENTS (current_function_decl);
3517 parm = DECL_CHAIN (parm))
3519 tree type = TREE_TYPE (parm);
3523 if (TREE_THIS_VOLATILE (parm)
3524 || TREE_ADDRESSABLE (parm)
3525 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3528 if (is_unused_scalar_param (parm))
3534 if (POINTER_TYPE_P (type))
3536 type = TREE_TYPE (type);
3538 if (TREE_CODE (type) == FUNCTION_TYPE
3539 || TYPE_VOLATILE (type)
3540 || (TREE_CODE (type) == ARRAY_TYPE
3541 && TYPE_NONALIASED_COMPONENT (type))
3542 || !is_gimple_reg (parm)
3543 || is_va_list_type (type)
3544 || ptr_parm_has_direct_uses (parm))
3547 else if (!AGGREGATE_TYPE_P (type))
3550 if (!COMPLETE_TYPE_P (type)
3551 || !host_integerp (TYPE_SIZE (type), 1)
3552 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3553 || (AGGREGATE_TYPE_P (type)
3554 && type_internals_preclude_sra_p (type, &msg)))
3557 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3559 if (dump_file && (dump_flags & TDF_DETAILS))
3561 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3562 print_generic_expr (dump_file, parm, 0);
3563 fprintf (dump_file, "\n");
3567 func_param_count = count;
3571 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3575 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3578 struct access *repr = (struct access *) data;
3580 repr->grp_maybe_modified = 1;
3584 /* Analyze what representatives (in linked lists accessible from
3585 REPRESENTATIVES) can be modified by side effects of statements in the
3586 current function. */
3589 analyze_modified_params (VEC (access_p, heap) *representatives)
3593 for (i = 0; i < func_param_count; i++)
3595 struct access *repr;
3597 for (repr = VEC_index (access_p, representatives, i);
3599 repr = repr->next_grp)
3601 struct access *access;
3605 if (no_accesses_p (repr))
3607 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3608 || repr->grp_maybe_modified)
3611 ao_ref_init (&ar, repr->expr);
3612 visited = BITMAP_ALLOC (NULL);
3613 for (access = repr; access; access = access->next_sibling)
3615 /* All accesses are read ones, otherwise grp_maybe_modified would
3616 be trivially set. */
3617 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3618 mark_maybe_modified, repr, &visited);
3619 if (repr->grp_maybe_modified)
3622 BITMAP_FREE (visited);
3627 /* Propagate distances in bb_dereferences in the opposite direction than the
3628 control flow edges, in each step storing the maximum of the current value
3629 and the minimum of all successors. These steps are repeated until the table
3630 stabilizes. Note that BBs which might terminate the functions (according to
3631 final_bbs bitmap) never updated in this way. */
3634 propagate_dereference_distances (void)
3636 VEC (basic_block, heap) *queue;
3639 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3640 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3643 VEC_quick_push (basic_block, queue, bb);
3647 while (!VEC_empty (basic_block, queue))
3651 bool change = false;
3654 bb = VEC_pop (basic_block, queue);
3657 if (bitmap_bit_p (final_bbs, bb->index))
3660 for (i = 0; i < func_param_count; i++)
3662 int idx = bb->index * func_param_count + i;
3664 HOST_WIDE_INT inh = 0;
3666 FOR_EACH_EDGE (e, ei, bb->succs)
3668 int succ_idx = e->dest->index * func_param_count + i;
3670 if (e->src == EXIT_BLOCK_PTR)
3676 inh = bb_dereferences [succ_idx];
3678 else if (bb_dereferences [succ_idx] < inh)
3679 inh = bb_dereferences [succ_idx];
3682 if (!first && bb_dereferences[idx] < inh)
3684 bb_dereferences[idx] = inh;
3689 if (change && !bitmap_bit_p (final_bbs, bb->index))
3690 FOR_EACH_EDGE (e, ei, bb->preds)
3695 e->src->aux = e->src;
3696 VEC_quick_push (basic_block, queue, e->src);
3700 VEC_free (basic_block, heap, queue);
3703 /* Dump a dereferences TABLE with heading STR to file F. */
3706 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3710 fprintf (dump_file, str);
3711 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3713 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3714 if (bb != EXIT_BLOCK_PTR)
3717 for (i = 0; i < func_param_count; i++)
3719 int idx = bb->index * func_param_count + i;
3720 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3725 fprintf (dump_file, "\n");
3728 /* Determine what (parts of) parameters passed by reference that are not
3729 assigned to are not certainly dereferenced in this function and thus the
3730 dereferencing cannot be safely moved to the caller without potentially
3731 introducing a segfault. Mark such REPRESENTATIVES as
3732 grp_not_necessarilly_dereferenced.
3734 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3735 part is calculated rather than simple booleans are calculated for each
3736 pointer parameter to handle cases when only a fraction of the whole
3737 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3740 The maximum dereference distances for each pointer parameter and BB are
3741 already stored in bb_dereference. This routine simply propagates these
3742 values upwards by propagate_dereference_distances and then compares the
3743 distances of individual parameters in the ENTRY BB to the equivalent
3744 distances of each representative of a (fraction of a) parameter. */
3747 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3751 if (dump_file && (dump_flags & TDF_DETAILS))
3752 dump_dereferences_table (dump_file,
3753 "Dereference table before propagation:\n",
3756 propagate_dereference_distances ();
3758 if (dump_file && (dump_flags & TDF_DETAILS))
3759 dump_dereferences_table (dump_file,
3760 "Dereference table after propagation:\n",
3763 for (i = 0; i < func_param_count; i++)
3765 struct access *repr = VEC_index (access_p, representatives, i);
3766 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3768 if (!repr || no_accesses_p (repr))
3773 if ((repr->offset + repr->size) > bb_dereferences[idx])
3774 repr->grp_not_necessarilly_dereferenced = 1;
3775 repr = repr->next_grp;
3781 /* Return the representative access for the parameter declaration PARM if it is
3782 a scalar passed by reference which is not written to and the pointer value
3783 is not used directly. Thus, if it is legal to dereference it in the caller
3784 and we can rule out modifications through aliases, such parameter should be
3785 turned into one passed by value. Return NULL otherwise. */
3787 static struct access *
3788 unmodified_by_ref_scalar_representative (tree parm)
3790 int i, access_count;
3791 struct access *repr;
3792 VEC (access_p, heap) *access_vec;
3794 access_vec = get_base_access_vector (parm);
3795 gcc_assert (access_vec);
3796 repr = VEC_index (access_p, access_vec, 0);
3799 repr->group_representative = repr;
3801 access_count = VEC_length (access_p, access_vec);
3802 for (i = 1; i < access_count; i++)
3804 struct access *access = VEC_index (access_p, access_vec, i);
3807 access->group_representative = repr;
3808 access->next_sibling = repr->next_sibling;
3809 repr->next_sibling = access;
3813 repr->grp_scalar_ptr = 1;
3817 /* Return true iff this access precludes IPA-SRA of the parameter it is
3821 access_precludes_ipa_sra_p (struct access *access)
3823 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3824 is incompatible assign in a call statement (and possibly even in asm
3825 statements). This can be relaxed by using a new temporary but only for
3826 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3827 intraprocedural SRA we deal with this by keeping the old aggregate around,
3828 something we cannot do in IPA-SRA.) */
3830 && (is_gimple_call (access->stmt)
3831 || gimple_code (access->stmt) == GIMPLE_ASM))
3834 if (STRICT_ALIGNMENT
3835 && tree_non_aligned_mem_p (access->expr, TYPE_ALIGN (access->type)))
3842 /* Sort collected accesses for parameter PARM, identify representatives for
3843 each accessed region and link them together. Return NULL if there are
3844 different but overlapping accesses, return the special ptr value meaning
3845 there are no accesses for this parameter if that is the case and return the
3846 first representative otherwise. Set *RO_GRP if there is a group of accesses
3847 with only read (i.e. no write) accesses. */
3849 static struct access *
3850 splice_param_accesses (tree parm, bool *ro_grp)
3852 int i, j, access_count, group_count;
3853 int agg_size, total_size = 0;
3854 struct access *access, *res, **prev_acc_ptr = &res;
3855 VEC (access_p, heap) *access_vec;
3857 access_vec = get_base_access_vector (parm);
3859 return &no_accesses_representant;
3860 access_count = VEC_length (access_p, access_vec);
3862 VEC_qsort (access_p, access_vec, compare_access_positions);
3867 while (i < access_count)
3871 access = VEC_index (access_p, access_vec, i);
3872 modification = access->write;
3873 if (access_precludes_ipa_sra_p (access))
3875 a1_alias_type = reference_alias_ptr_type (access->expr);
3877 /* Access is about to become group representative unless we find some
3878 nasty overlap which would preclude us from breaking this parameter
3882 while (j < access_count)
3884 struct access *ac2 = VEC_index (access_p, access_vec, j);
3885 if (ac2->offset != access->offset)
3887 /* All or nothing law for parameters. */
3888 if (access->offset + access->size > ac2->offset)
3893 else if (ac2->size != access->size)
3896 if (access_precludes_ipa_sra_p (ac2)
3897 || (ac2->type != access->type
3898 && (TREE_ADDRESSABLE (ac2->type)
3899 || TREE_ADDRESSABLE (access->type)))
3900 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3903 modification |= ac2->write;
3904 ac2->group_representative = access;
3905 ac2->next_sibling = access->next_sibling;
3906 access->next_sibling = ac2;
3911 access->grp_maybe_modified = modification;
3914 *prev_acc_ptr = access;
3915 prev_acc_ptr = &access->next_grp;
3916 total_size += access->size;
3920 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3921 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3923 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3924 if (total_size >= agg_size)
3927 gcc_assert (group_count > 0);
3931 /* Decide whether parameters with representative accesses given by REPR should
3932 be reduced into components. */
3935 decide_one_param_reduction (struct access *repr)
3937 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3942 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3943 gcc_assert (cur_parm_size > 0);
3945 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3948 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3953 agg_size = cur_parm_size;
3959 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3960 print_generic_expr (dump_file, parm, 0);
3961 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3962 for (acc = repr; acc; acc = acc->next_grp)
3963 dump_access (dump_file, acc, true);
3967 new_param_count = 0;
3969 for (; repr; repr = repr->next_grp)
3971 gcc_assert (parm == repr->base);
3973 /* Taking the address of a non-addressable field is verboten. */
3974 if (by_ref && repr->non_addressable)
3977 /* Do not decompose a non-BLKmode param in a way that would
3978 create BLKmode params. Especially for by-reference passing
3979 (thus, pointer-type param) this is hardly worthwhile. */
3980 if (DECL_MODE (parm) != BLKmode
3981 && TYPE_MODE (repr->type) == BLKmode)
3984 if (!by_ref || (!repr->grp_maybe_modified
3985 && !repr->grp_not_necessarilly_dereferenced))
3986 total_size += repr->size;
3988 total_size += cur_parm_size;
3993 gcc_assert (new_param_count > 0);
3995 if (optimize_function_for_size_p (cfun))
3996 parm_size_limit = cur_parm_size;
3998 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
4001 if (total_size < agg_size
4002 && total_size <= parm_size_limit)
4005 fprintf (dump_file, " ....will be split into %i components\n",
4007 return new_param_count;
4013 /* The order of the following enums is important, we need to do extra work for
4014 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
4015 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
4016 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
4018 /* Identify representatives of all accesses to all candidate parameters for
4019 IPA-SRA. Return result based on what representatives have been found. */
4021 static enum ipa_splicing_result
4022 splice_all_param_accesses (VEC (access_p, heap) **representatives)
4024 enum ipa_splicing_result result = NO_GOOD_ACCESS;
4026 struct access *repr;
4028 *representatives = VEC_alloc (access_p, heap, func_param_count);
4030 for (parm = DECL_ARGUMENTS (current_function_decl);
4032 parm = DECL_CHAIN (parm))
4034 if (is_unused_scalar_param (parm))
4036 VEC_quick_push (access_p, *representatives,
4037 &no_accesses_representant);
4038 if (result == NO_GOOD_ACCESS)
4039 result = UNUSED_PARAMS;
4041 else if (POINTER_TYPE_P (TREE_TYPE (parm))
4042 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
4043 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4045 repr = unmodified_by_ref_scalar_representative (parm);
4046 VEC_quick_push (access_p, *representatives, repr);
4048 result = UNMODIF_BY_REF_ACCESSES;
4050 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4052 bool ro_grp = false;
4053 repr = splice_param_accesses (parm, &ro_grp);
4054 VEC_quick_push (access_p, *representatives, repr);
4056 if (repr && !no_accesses_p (repr))
4058 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4061 result = UNMODIF_BY_REF_ACCESSES;
4062 else if (result < MODIF_BY_REF_ACCESSES)
4063 result = MODIF_BY_REF_ACCESSES;
4065 else if (result < BY_VAL_ACCESSES)
4066 result = BY_VAL_ACCESSES;
4068 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
4069 result = UNUSED_PARAMS;
4072 VEC_quick_push (access_p, *representatives, NULL);
4075 if (result == NO_GOOD_ACCESS)
4077 VEC_free (access_p, heap, *representatives);
4078 *representatives = NULL;
4079 return NO_GOOD_ACCESS;
4085 /* Return the index of BASE in PARMS. Abort if it is not found. */
4088 get_param_index (tree base, VEC(tree, heap) *parms)
4092 len = VEC_length (tree, parms);
4093 for (i = 0; i < len; i++)
4094 if (VEC_index (tree, parms, i) == base)
4099 /* Convert the decisions made at the representative level into compact
4100 parameter adjustments. REPRESENTATIVES are pointers to first
4101 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4102 final number of adjustments. */
4104 static ipa_parm_adjustment_vec
4105 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
4106 int adjustments_count)
4108 VEC (tree, heap) *parms;
4109 ipa_parm_adjustment_vec adjustments;
4113 gcc_assert (adjustments_count > 0);
4114 parms = ipa_get_vector_of_formal_parms (current_function_decl);
4115 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
4116 parm = DECL_ARGUMENTS (current_function_decl);
4117 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
4119 struct access *repr = VEC_index (access_p, representatives, i);
4121 if (!repr || no_accesses_p (repr))
4123 struct ipa_parm_adjustment *adj;
4125 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4126 memset (adj, 0, sizeof (*adj));
4127 adj->base_index = get_param_index (parm, parms);
4130 adj->copy_param = 1;
4132 adj->remove_param = 1;
4136 struct ipa_parm_adjustment *adj;
4137 int index = get_param_index (parm, parms);
4139 for (; repr; repr = repr->next_grp)
4141 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4142 memset (adj, 0, sizeof (*adj));
4143 gcc_assert (repr->base == parm);
4144 adj->base_index = index;
4145 adj->base = repr->base;
4146 adj->type = repr->type;
4147 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
4148 adj->offset = repr->offset;
4149 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
4150 && (repr->grp_maybe_modified
4151 || repr->grp_not_necessarilly_dereferenced));
4156 VEC_free (tree, heap, parms);
4160 /* Analyze the collected accesses and produce a plan what to do with the
4161 parameters in the form of adjustments, NULL meaning nothing. */
4163 static ipa_parm_adjustment_vec
4164 analyze_all_param_acesses (void)
4166 enum ipa_splicing_result repr_state;
4167 bool proceed = false;
4168 int i, adjustments_count = 0;
4169 VEC (access_p, heap) *representatives;
4170 ipa_parm_adjustment_vec adjustments;
4172 repr_state = splice_all_param_accesses (&representatives);
4173 if (repr_state == NO_GOOD_ACCESS)
4176 /* If there are any parameters passed by reference which are not modified
4177 directly, we need to check whether they can be modified indirectly. */
4178 if (repr_state == UNMODIF_BY_REF_ACCESSES)
4180 analyze_caller_dereference_legality (representatives);
4181 analyze_modified_params (representatives);
4184 for (i = 0; i < func_param_count; i++)
4186 struct access *repr = VEC_index (access_p, representatives, i);
4188 if (repr && !no_accesses_p (repr))
4190 if (repr->grp_scalar_ptr)
4192 adjustments_count++;
4193 if (repr->grp_not_necessarilly_dereferenced
4194 || repr->grp_maybe_modified)
4195 VEC_replace (access_p, representatives, i, NULL);
4199 sra_stats.scalar_by_ref_to_by_val++;
4204 int new_components = decide_one_param_reduction (repr);
4206 if (new_components == 0)
4208 VEC_replace (access_p, representatives, i, NULL);
4209 adjustments_count++;
4213 adjustments_count += new_components;
4214 sra_stats.aggregate_params_reduced++;
4215 sra_stats.param_reductions_created += new_components;
4222 if (no_accesses_p (repr))
4225 sra_stats.deleted_unused_parameters++;
4227 adjustments_count++;
4231 if (!proceed && dump_file)
4232 fprintf (dump_file, "NOT proceeding to change params.\n");
4235 adjustments = turn_representatives_into_adjustments (representatives,
4240 VEC_free (access_p, heap, representatives);
4244 /* If a parameter replacement identified by ADJ does not yet exist in the form
4245 of declaration, create it and record it, otherwise return the previously
4249 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4252 if (!adj->new_ssa_base)
4254 char *pretty_name = make_fancy_name (adj->base);
4256 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4257 DECL_NAME (repl) = get_identifier (pretty_name);
4258 obstack_free (&name_obstack, pretty_name);
4260 add_referenced_var (repl);
4261 adj->new_ssa_base = repl;
4264 repl = adj->new_ssa_base;
4268 /* Find the first adjustment for a particular parameter BASE in a vector of
4269 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4272 static struct ipa_parm_adjustment *
4273 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4277 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4278 for (i = 0; i < len; i++)
4280 struct ipa_parm_adjustment *adj;
4282 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4283 if (!adj->copy_param && adj->base == base)
4290 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4291 removed because its value is not used, replace the SSA_NAME with a one
4292 relating to a created VAR_DECL together all of its uses and return true.
4293 ADJUSTMENTS is a pointer to an adjustments vector. */
4296 replace_removed_params_ssa_names (gimple stmt,
4297 ipa_parm_adjustment_vec adjustments)
4299 struct ipa_parm_adjustment *adj;
4300 tree lhs, decl, repl, name;
4302 if (gimple_code (stmt) == GIMPLE_PHI)
4303 lhs = gimple_phi_result (stmt);
4304 else if (is_gimple_assign (stmt))
4305 lhs = gimple_assign_lhs (stmt);
4306 else if (is_gimple_call (stmt))
4307 lhs = gimple_call_lhs (stmt);
4311 if (TREE_CODE (lhs) != SSA_NAME)
4313 decl = SSA_NAME_VAR (lhs);
4314 if (TREE_CODE (decl) != PARM_DECL)
4317 adj = get_adjustment_for_base (adjustments, decl);
4321 repl = get_replaced_param_substitute (adj);
4322 name = make_ssa_name (repl, stmt);
4326 fprintf (dump_file, "replacing an SSA name of a removed param ");
4327 print_generic_expr (dump_file, lhs, 0);
4328 fprintf (dump_file, " with ");
4329 print_generic_expr (dump_file, name, 0);
4330 fprintf (dump_file, "\n");
4333 if (is_gimple_assign (stmt))
4334 gimple_assign_set_lhs (stmt, name);
4335 else if (is_gimple_call (stmt))
4336 gimple_call_set_lhs (stmt, name);
4338 gimple_phi_set_result (stmt, name);
4340 replace_uses_by (lhs, name);
4341 release_ssa_name (lhs);
4345 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4346 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4347 specifies whether the function should care about type incompatibility the
4348 current and new expressions. If it is false, the function will leave
4349 incompatibility issues to the caller. Return true iff the expression
4353 sra_ipa_modify_expr (tree *expr, bool convert,
4354 ipa_parm_adjustment_vec adjustments)
4357 struct ipa_parm_adjustment *adj, *cand = NULL;
4358 HOST_WIDE_INT offset, size, max_size;
4361 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4363 if (TREE_CODE (*expr) == BIT_FIELD_REF
4364 || TREE_CODE (*expr) == IMAGPART_EXPR
4365 || TREE_CODE (*expr) == REALPART_EXPR)
4367 expr = &TREE_OPERAND (*expr, 0);
4371 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4372 if (!base || size == -1 || max_size == -1)
4375 if (TREE_CODE (base) == MEM_REF)
4377 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4378 base = TREE_OPERAND (base, 0);
4381 base = get_ssa_base_param (base);
4382 if (!base || TREE_CODE (base) != PARM_DECL)
4385 for (i = 0; i < len; i++)
4387 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4389 if (adj->base == base &&
4390 (adj->offset == offset || adj->remove_param))
4396 if (!cand || cand->copy_param || cand->remove_param)
4400 src = build_simple_mem_ref (cand->reduction);
4402 src = cand->reduction;
4404 if (dump_file && (dump_flags & TDF_DETAILS))
4406 fprintf (dump_file, "About to replace expr ");
4407 print_generic_expr (dump_file, *expr, 0);
4408 fprintf (dump_file, " with ");
4409 print_generic_expr (dump_file, src, 0);
4410 fprintf (dump_file, "\n");
4413 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4415 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4423 /* If the statement pointed to by STMT_PTR contains any expressions that need
4424 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4425 potential type incompatibilities (GSI is used to accommodate conversion
4426 statements and must point to the statement). Return true iff the statement
4430 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4431 ipa_parm_adjustment_vec adjustments)
4433 gimple stmt = *stmt_ptr;
4434 tree *lhs_p, *rhs_p;
4437 if (!gimple_assign_single_p (stmt))
4440 rhs_p = gimple_assign_rhs1_ptr (stmt);
4441 lhs_p = gimple_assign_lhs_ptr (stmt);
4443 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4444 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4447 tree new_rhs = NULL_TREE;
4449 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4451 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4453 /* V_C_Es of constructors can cause trouble (PR 42714). */
4454 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4455 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4457 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4460 new_rhs = fold_build1_loc (gimple_location (stmt),
4461 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4464 else if (REFERENCE_CLASS_P (*rhs_p)
4465 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4466 && !is_gimple_reg (*lhs_p))
4467 /* This can happen when an assignment in between two single field
4468 structures is turned into an assignment in between two pointers to
4469 scalars (PR 42237). */
4474 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4475 true, GSI_SAME_STMT);
4477 gimple_assign_set_rhs_from_tree (gsi, tmp);
4486 /* Traverse the function body and all modifications as described in
4487 ADJUSTMENTS. Return true iff the CFG has been changed. */
4490 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4492 bool cfg_changed = false;
4497 gimple_stmt_iterator gsi;
4499 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4500 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4502 gsi = gsi_start_bb (bb);
4503 while (!gsi_end_p (gsi))
4505 gimple stmt = gsi_stmt (gsi);
4506 bool modified = false;
4510 switch (gimple_code (stmt))
4513 t = gimple_return_retval_ptr (stmt);
4514 if (*t != NULL_TREE)
4515 modified |= sra_ipa_modify_expr (t, true, adjustments);
4519 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4520 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4524 /* Operands must be processed before the lhs. */
4525 for (i = 0; i < gimple_call_num_args (stmt); i++)
4527 t = gimple_call_arg_ptr (stmt, i);
4528 modified |= sra_ipa_modify_expr (t, true, adjustments);
4531 if (gimple_call_lhs (stmt))
4533 t = gimple_call_lhs_ptr (stmt);
4534 modified |= sra_ipa_modify_expr (t, false, adjustments);
4535 modified |= replace_removed_params_ssa_names (stmt,
4541 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4543 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4544 modified |= sra_ipa_modify_expr (t, true, adjustments);
4546 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4548 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4549 modified |= sra_ipa_modify_expr (t, false, adjustments);
4560 if (maybe_clean_eh_stmt (stmt)
4561 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4571 /* Call gimple_debug_bind_reset_value on all debug statements describing
4572 gimple register parameters that are being removed or replaced. */
4575 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4578 gimple_stmt_iterator *gsip = NULL, gsi;
4580 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR))
4582 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
4585 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4586 for (i = 0; i < len; i++)
4588 struct ipa_parm_adjustment *adj;
4589 imm_use_iterator ui;
4590 gimple stmt, def_temp;
4591 tree name, vexpr, copy = NULL_TREE;
4592 use_operand_p use_p;
4594 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4595 if (adj->copy_param || !is_gimple_reg (adj->base))
4597 name = gimple_default_def (cfun, adj->base);
4600 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4602 /* All other users must have been removed by
4603 ipa_sra_modify_function_body. */
4604 gcc_assert (is_gimple_debug (stmt));
4605 if (vexpr == NULL && gsip != NULL)
4607 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4608 vexpr = make_node (DEBUG_EXPR_DECL);
4609 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4611 DECL_ARTIFICIAL (vexpr) = 1;
4612 TREE_TYPE (vexpr) = TREE_TYPE (name);
4613 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4614 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4618 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4619 SET_USE (use_p, vexpr);
4622 gimple_debug_bind_reset_value (stmt);
4625 /* Create a VAR_DECL for debug info purposes. */
4626 if (!DECL_IGNORED_P (adj->base))
4628 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4629 VAR_DECL, DECL_NAME (adj->base),
4630 TREE_TYPE (adj->base));
4631 if (DECL_PT_UID_SET_P (adj->base))
4632 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4633 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4634 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4635 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4636 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4637 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4638 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4639 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4640 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4641 SET_DECL_RTL (copy, 0);
4642 TREE_USED (copy) = 1;
4643 DECL_CONTEXT (copy) = current_function_decl;
4644 add_referenced_var (copy);
4645 add_local_decl (cfun, copy);
4647 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4648 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4650 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4652 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4654 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4656 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4658 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4663 /* Return false iff all callers have at least as many actual arguments as there
4664 are formal parameters in the current function. */
4667 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4668 void *data ATTRIBUTE_UNUSED)
4670 struct cgraph_edge *cs;
4671 for (cs = node->callers; cs; cs = cs->next_caller)
4672 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4678 /* Convert all callers of NODE. */
4681 convert_callers_for_node (struct cgraph_node *node,
4684 ipa_parm_adjustment_vec adjustments = (ipa_parm_adjustment_vec)data;
4685 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4686 struct cgraph_edge *cs;
4688 for (cs = node->callers; cs; cs = cs->next_caller)
4690 current_function_decl = cs->caller->decl;
4691 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4694 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4695 cs->caller->uid, cs->callee->uid,
4696 cgraph_node_name (cs->caller),
4697 cgraph_node_name (cs->callee));
4699 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4704 for (cs = node->callers; cs; cs = cs->next_caller)
4705 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4706 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4707 compute_inline_parameters (cs->caller, true);
4708 BITMAP_FREE (recomputed_callers);
4713 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4716 convert_callers (struct cgraph_node *node, tree old_decl,
4717 ipa_parm_adjustment_vec adjustments)
4719 tree old_cur_fndecl = current_function_decl;
4720 basic_block this_block;
4722 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4723 adjustments, false);
4725 current_function_decl = old_cur_fndecl;
4727 if (!encountered_recursive_call)
4730 FOR_EACH_BB (this_block)
4732 gimple_stmt_iterator gsi;
4734 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4736 gimple stmt = gsi_stmt (gsi);
4738 if (gimple_code (stmt) != GIMPLE_CALL)
4740 call_fndecl = gimple_call_fndecl (stmt);
4741 if (call_fndecl == old_decl)
4744 fprintf (dump_file, "Adjusting recursive call");
4745 gimple_call_set_fndecl (stmt, node->decl);
4746 ipa_modify_call_arguments (NULL, stmt, adjustments);
4754 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4755 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4758 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4760 struct cgraph_node *new_node;
4762 VEC (cgraph_edge_p, heap) * redirect_callers = collect_callers_of_node (node);
4764 rebuild_cgraph_edges ();
4765 free_dominance_info (CDI_DOMINATORS);
4767 current_function_decl = NULL_TREE;
4769 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4770 false, NULL, NULL, "isra");
4771 current_function_decl = new_node->decl;
4772 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4774 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4775 cfg_changed = ipa_sra_modify_function_body (adjustments);
4776 sra_ipa_reset_debug_stmts (adjustments);
4777 convert_callers (new_node, node->decl, adjustments);
4778 cgraph_make_node_local (new_node);
4782 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4783 attributes, return true otherwise. NODE is the cgraph node of the current
4787 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4789 if (!cgraph_node_can_be_local_p (node))
4792 fprintf (dump_file, "Function not local to this compilation unit.\n");
4796 if (!node->local.can_change_signature)
4799 fprintf (dump_file, "Function can not change signature.\n");
4803 if (!tree_versionable_function_p (node->decl))
4806 fprintf (dump_file, "Function is not versionable.\n");
4810 if (DECL_VIRTUAL_P (current_function_decl))
4813 fprintf (dump_file, "Function is a virtual method.\n");
4817 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4818 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4821 fprintf (dump_file, "Function too big to be made truly local.\n");
4829 "Function has no callers in this compilation unit.\n");
4836 fprintf (dump_file, "Function uses stdarg. \n");
4840 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4846 /* Perform early interprocedural SRA. */
4849 ipa_early_sra (void)
4851 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4852 ipa_parm_adjustment_vec adjustments;
4855 if (!ipa_sra_preliminary_function_checks (node))
4859 sra_mode = SRA_MODE_EARLY_IPA;
4861 if (!find_param_candidates ())
4864 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4868 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4872 fprintf (dump_file, "There are callers with insufficient number of "
4877 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4879 * last_basic_block_for_function (cfun));
4880 final_bbs = BITMAP_ALLOC (NULL);
4883 if (encountered_apply_args)
4886 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4890 if (encountered_unchangable_recursive_call)
4893 fprintf (dump_file, "Function calls itself with insufficient "
4894 "number of arguments.\n");
4898 adjustments = analyze_all_param_acesses ();
4902 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4904 if (modify_function (node, adjustments))
4905 ret = TODO_update_ssa | TODO_cleanup_cfg;
4907 ret = TODO_update_ssa;
4908 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4910 statistics_counter_event (cfun, "Unused parameters deleted",
4911 sra_stats.deleted_unused_parameters);
4912 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4913 sra_stats.scalar_by_ref_to_by_val);
4914 statistics_counter_event (cfun, "Aggregate parameters broken up",
4915 sra_stats.aggregate_params_reduced);
4916 statistics_counter_event (cfun, "Aggregate parameter components created",
4917 sra_stats.param_reductions_created);
4920 BITMAP_FREE (final_bbs);
4921 free (bb_dereferences);
4923 sra_deinitialize ();
4927 /* Return if early ipa sra shall be performed. */
4929 ipa_early_sra_gate (void)
4931 return flag_ipa_sra && dbg_cnt (eipa_sra);
4934 struct gimple_opt_pass pass_early_ipa_sra =
4938 "eipa_sra", /* name */
4939 ipa_early_sra_gate, /* gate */
4940 ipa_early_sra, /* execute */
4943 0, /* static_pass_number */
4944 TV_IPA_SRA, /* tv_id */
4945 0, /* properties_required */
4946 0, /* properties_provided */
4947 0, /* properties_destroyed */
4948 0, /* todo_flags_start */
4949 TODO_dump_cgraph /* todo_flags_finish */