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 ??? But it is not at all clear that prev_base is an access
1531 that was in the IL that way, so be conservative for now. */
1532 align = get_pointer_alignment_1 (base, &misalign);
1533 misalign += (double_int_sext (tree_to_double_int (off),
1534 TYPE_PRECISION (TREE_TYPE (off))).low
1536 misalign = misalign & (align - 1);
1538 align = (misalign & -misalign);
1539 if (align < TYPE_ALIGN (exp_type))
1540 exp_type = build_aligned_type (exp_type, align);
1542 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1545 DEF_VEC_ALLOC_P_STACK (tree);
1546 #define VEC_tree_stack_alloc(alloc) VEC_stack_alloc (tree, alloc)
1548 /* Construct a memory reference to a part of an aggregate BASE at the given
1549 OFFSET and of the type of MODEL. In case this is a chain of references
1550 to component, the function will replicate the chain of COMPONENT_REFs of
1551 the expression of MODEL to access it. GSI and INSERT_AFTER have the same
1552 meaning as in build_ref_for_offset. */
1555 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1556 struct access *model, gimple_stmt_iterator *gsi,
1559 tree type = model->type, t;
1560 VEC(tree,stack) *cr_stack = NULL;
1562 if (TREE_CODE (model->expr) == COMPONENT_REF)
1564 tree expr = model->expr;
1566 /* Create a stack of the COMPONENT_REFs so later we can walk them in
1567 order from inner to outer. */
1568 cr_stack = VEC_alloc (tree, stack, 6);
1571 tree field = TREE_OPERAND (expr, 1);
1572 tree cr_offset = component_ref_field_offset (expr);
1573 HOST_WIDE_INT bit_pos
1574 = tree_low_cst (cr_offset, 1) * BITS_PER_UNIT
1575 + TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (field));
1577 /* We can be called with a model different from the one associated
1578 with BASE so we need to avoid going up the chain too far. */
1579 if (offset - bit_pos < 0)
1583 VEC_safe_push (tree, stack, cr_stack, expr);
1585 expr = TREE_OPERAND (expr, 0);
1586 type = TREE_TYPE (expr);
1587 } while (TREE_CODE (expr) == COMPONENT_REF);
1590 t = build_ref_for_offset (loc, base, offset, type, gsi, insert_after);
1592 if (TREE_CODE (model->expr) == COMPONENT_REF)
1597 /* Now replicate the chain of COMPONENT_REFs from inner to outer. */
1598 FOR_EACH_VEC_ELT_REVERSE (tree, cr_stack, i, expr)
1600 tree field = TREE_OPERAND (expr, 1);
1601 t = fold_build3_loc (loc, COMPONENT_REF, TREE_TYPE (field), t, field,
1602 TREE_OPERAND (expr, 2));
1605 VEC_free (tree, stack, cr_stack);
1611 /* Construct a memory reference consisting of component_refs and array_refs to
1612 a part of an aggregate *RES (which is of type TYPE). The requested part
1613 should have type EXP_TYPE at be the given OFFSET. This function might not
1614 succeed, it returns true when it does and only then *RES points to something
1615 meaningful. This function should be used only to build expressions that we
1616 might need to present to user (e.g. in warnings). In all other situations,
1617 build_ref_for_model or build_ref_for_offset should be used instead. */
1620 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1626 tree tr_size, index, minidx;
1627 HOST_WIDE_INT el_size;
1629 if (offset == 0 && exp_type
1630 && types_compatible_p (exp_type, type))
1633 switch (TREE_CODE (type))
1636 case QUAL_UNION_TYPE:
1638 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1640 HOST_WIDE_INT pos, size;
1641 tree expr, *expr_ptr;
1643 if (TREE_CODE (fld) != FIELD_DECL)
1646 pos = int_bit_position (fld);
1647 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1648 tr_size = DECL_SIZE (fld);
1649 if (!tr_size || !host_integerp (tr_size, 1))
1651 size = tree_low_cst (tr_size, 1);
1657 else if (pos > offset || (pos + size) <= offset)
1660 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1663 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1664 offset - pos, exp_type))
1673 tr_size = TYPE_SIZE (TREE_TYPE (type));
1674 if (!tr_size || !host_integerp (tr_size, 1))
1676 el_size = tree_low_cst (tr_size, 1);
1678 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1679 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1681 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1682 if (!integer_zerop (minidx))
1683 index = int_const_binop (PLUS_EXPR, index, minidx);
1684 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1685 NULL_TREE, NULL_TREE);
1686 offset = offset % el_size;
1687 type = TREE_TYPE (type);
1702 /* Return true iff TYPE is stdarg va_list type. */
1705 is_va_list_type (tree type)
1707 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1710 /* Print message to dump file why a variable was rejected. */
1713 reject (tree var, const char *msg)
1715 if (dump_file && (dump_flags & TDF_DETAILS))
1717 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1718 print_generic_expr (dump_file, var, 0);
1719 fprintf (dump_file, "\n");
1723 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1724 those with type which is suitable for scalarization. */
1727 find_var_candidates (void)
1730 referenced_var_iterator rvi;
1734 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
1736 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1738 type = TREE_TYPE (var);
1740 if (!AGGREGATE_TYPE_P (type))
1742 reject (var, "not aggregate");
1745 if (needs_to_live_in_memory (var))
1747 reject (var, "needs to live in memory");
1750 if (TREE_THIS_VOLATILE (var))
1752 reject (var, "is volatile");
1755 if (!COMPLETE_TYPE_P (type))
1757 reject (var, "has incomplete type");
1760 if (!host_integerp (TYPE_SIZE (type), 1))
1762 reject (var, "type size not fixed");
1765 if (tree_low_cst (TYPE_SIZE (type), 1) == 0)
1767 reject (var, "type size is zero");
1770 if (type_internals_preclude_sra_p (type, &msg))
1775 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1776 we also want to schedule it rather late. Thus we ignore it in
1778 (sra_mode == SRA_MODE_EARLY_INTRA
1779 && is_va_list_type (type)))
1781 reject (var, "is va_list");
1785 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1787 if (dump_file && (dump_flags & TDF_DETAILS))
1789 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1790 print_generic_expr (dump_file, var, 0);
1791 fprintf (dump_file, "\n");
1799 /* Sort all accesses for the given variable, check for partial overlaps and
1800 return NULL if there are any. If there are none, pick a representative for
1801 each combination of offset and size and create a linked list out of them.
1802 Return the pointer to the first representative and make sure it is the first
1803 one in the vector of accesses. */
1805 static struct access *
1806 sort_and_splice_var_accesses (tree var)
1808 int i, j, access_count;
1809 struct access *res, **prev_acc_ptr = &res;
1810 VEC (access_p, heap) *access_vec;
1812 HOST_WIDE_INT low = -1, high = 0;
1814 access_vec = get_base_access_vector (var);
1817 access_count = VEC_length (access_p, access_vec);
1819 /* Sort by <OFFSET, SIZE>. */
1820 VEC_qsort (access_p, access_vec, compare_access_positions);
1823 while (i < access_count)
1825 struct access *access = VEC_index (access_p, access_vec, i);
1826 bool grp_write = access->write;
1827 bool grp_read = !access->write;
1828 bool grp_scalar_write = access->write
1829 && is_gimple_reg_type (access->type);
1830 bool grp_scalar_read = !access->write
1831 && is_gimple_reg_type (access->type);
1832 bool grp_assignment_read = access->grp_assignment_read;
1833 bool grp_assignment_write = access->grp_assignment_write;
1834 bool multiple_scalar_reads = false;
1835 bool total_scalarization = access->grp_total_scalarization;
1836 bool grp_partial_lhs = access->grp_partial_lhs;
1837 bool first_scalar = is_gimple_reg_type (access->type);
1838 bool unscalarizable_region = access->grp_unscalarizable_region;
1840 if (first || access->offset >= high)
1843 low = access->offset;
1844 high = access->offset + access->size;
1846 else if (access->offset > low && access->offset + access->size > high)
1849 gcc_assert (access->offset >= low
1850 && access->offset + access->size <= high);
1853 while (j < access_count)
1855 struct access *ac2 = VEC_index (access_p, access_vec, j);
1856 if (ac2->offset != access->offset || ac2->size != access->size)
1861 grp_scalar_write = (grp_scalar_write
1862 || is_gimple_reg_type (ac2->type));
1867 if (is_gimple_reg_type (ac2->type))
1869 if (grp_scalar_read)
1870 multiple_scalar_reads = true;
1872 grp_scalar_read = true;
1875 grp_assignment_read |= ac2->grp_assignment_read;
1876 grp_assignment_write |= ac2->grp_assignment_write;
1877 grp_partial_lhs |= ac2->grp_partial_lhs;
1878 unscalarizable_region |= ac2->grp_unscalarizable_region;
1879 total_scalarization |= ac2->grp_total_scalarization;
1880 relink_to_new_repr (access, ac2);
1882 /* If there are both aggregate-type and scalar-type accesses with
1883 this combination of size and offset, the comparison function
1884 should have put the scalars first. */
1885 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1886 ac2->group_representative = access;
1892 access->group_representative = access;
1893 access->grp_write = grp_write;
1894 access->grp_read = grp_read;
1895 access->grp_scalar_read = grp_scalar_read;
1896 access->grp_scalar_write = grp_scalar_write;
1897 access->grp_assignment_read = grp_assignment_read;
1898 access->grp_assignment_write = grp_assignment_write;
1899 access->grp_hint = multiple_scalar_reads || total_scalarization;
1900 access->grp_total_scalarization = total_scalarization;
1901 access->grp_partial_lhs = grp_partial_lhs;
1902 access->grp_unscalarizable_region = unscalarizable_region;
1903 if (access->first_link)
1904 add_access_to_work_queue (access);
1906 *prev_acc_ptr = access;
1907 prev_acc_ptr = &access->next_grp;
1910 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1914 /* Create a variable for the given ACCESS which determines the type, name and a
1915 few other properties. Return the variable declaration and store it also to
1916 ACCESS->replacement. */
1919 create_access_replacement (struct access *access, bool rename)
1923 repl = create_tmp_var (access->type, "SR");
1924 add_referenced_var (repl);
1926 mark_sym_for_renaming (repl);
1928 if (!access->grp_partial_lhs
1929 && (TREE_CODE (access->type) == COMPLEX_TYPE
1930 || TREE_CODE (access->type) == VECTOR_TYPE))
1931 DECL_GIMPLE_REG_P (repl) = 1;
1933 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1934 DECL_ARTIFICIAL (repl) = 1;
1935 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1937 if (DECL_NAME (access->base)
1938 && !DECL_IGNORED_P (access->base)
1939 && !DECL_ARTIFICIAL (access->base))
1941 char *pretty_name = make_fancy_name (access->expr);
1942 tree debug_expr = unshare_expr (access->expr), d;
1944 DECL_NAME (repl) = get_identifier (pretty_name);
1945 obstack_free (&name_obstack, pretty_name);
1947 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1948 as DECL_DEBUG_EXPR isn't considered when looking for still
1949 used SSA_NAMEs and thus they could be freed. All debug info
1950 generation cares is whether something is constant or variable
1951 and that get_ref_base_and_extent works properly on the
1953 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1954 switch (TREE_CODE (d))
1957 case ARRAY_RANGE_REF:
1958 if (TREE_OPERAND (d, 1)
1959 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1960 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1961 if (TREE_OPERAND (d, 3)
1962 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1963 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1966 if (TREE_OPERAND (d, 2)
1967 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1968 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1973 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1974 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1975 if (access->grp_no_warning)
1976 TREE_NO_WARNING (repl) = 1;
1978 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1981 TREE_NO_WARNING (repl) = 1;
1985 fprintf (dump_file, "Created a replacement for ");
1986 print_generic_expr (dump_file, access->base, 0);
1987 fprintf (dump_file, " offset: %u, size: %u: ",
1988 (unsigned) access->offset, (unsigned) access->size);
1989 print_generic_expr (dump_file, repl, 0);
1990 fprintf (dump_file, "\n");
1992 sra_stats.replacements++;
1997 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
2000 get_access_replacement (struct access *access)
2002 gcc_assert (access->grp_to_be_replaced);
2004 if (!access->replacement_decl)
2005 access->replacement_decl = create_access_replacement (access, true);
2006 return access->replacement_decl;
2009 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
2010 not mark it for renaming. */
2013 get_unrenamed_access_replacement (struct access *access)
2015 gcc_assert (!access->grp_to_be_replaced);
2017 if (!access->replacement_decl)
2018 access->replacement_decl = create_access_replacement (access, false);
2019 return access->replacement_decl;
2023 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
2024 linked list along the way. Stop when *ACCESS is NULL or the access pointed
2025 to it is not "within" the root. Return false iff some accesses partially
2029 build_access_subtree (struct access **access)
2031 struct access *root = *access, *last_child = NULL;
2032 HOST_WIDE_INT limit = root->offset + root->size;
2034 *access = (*access)->next_grp;
2035 while (*access && (*access)->offset + (*access)->size <= limit)
2038 root->first_child = *access;
2040 last_child->next_sibling = *access;
2041 last_child = *access;
2043 if (!build_access_subtree (access))
2047 if (*access && (*access)->offset < limit)
2053 /* Build a tree of access representatives, ACCESS is the pointer to the first
2054 one, others are linked in a list by the next_grp field. Return false iff
2055 some accesses partially overlap. */
2058 build_access_trees (struct access *access)
2062 struct access *root = access;
2064 if (!build_access_subtree (&access))
2066 root->next_grp = access;
2071 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2075 expr_with_var_bounded_array_refs_p (tree expr)
2077 while (handled_component_p (expr))
2079 if (TREE_CODE (expr) == ARRAY_REF
2080 && !host_integerp (array_ref_low_bound (expr), 0))
2082 expr = TREE_OPERAND (expr, 0);
2087 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2088 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2089 sorts of access flags appropriately along the way, notably always set
2090 grp_read and grp_assign_read according to MARK_READ and grp_write when
2093 Creating a replacement for a scalar access is considered beneficial if its
2094 grp_hint is set (this means we are either attempting total scalarization or
2095 there is more than one direct read access) or according to the following
2098 Access written to through a scalar type (once or more times)
2100 | Written to in an assignment statement
2102 | | Access read as scalar _once_
2104 | | | Read in an assignment statement
2106 | | | | Scalarize Comment
2107 -----------------------------------------------------------------------------
2108 0 0 0 0 No access for the scalar
2109 0 0 0 1 No access for the scalar
2110 0 0 1 0 No Single read - won't help
2111 0 0 1 1 No The same case
2112 0 1 0 0 No access for the scalar
2113 0 1 0 1 No access for the scalar
2114 0 1 1 0 Yes s = *g; return s.i;
2115 0 1 1 1 Yes The same case as above
2116 1 0 0 0 No Won't help
2117 1 0 0 1 Yes s.i = 1; *g = s;
2118 1 0 1 0 Yes s.i = 5; g = s.i;
2119 1 0 1 1 Yes The same case as above
2120 1 1 0 0 No Won't help.
2121 1 1 0 1 Yes s.i = 1; *g = s;
2122 1 1 1 0 Yes s = *g; return s.i;
2123 1 1 1 1 Yes Any of the above yeses */
2126 analyze_access_subtree (struct access *root, struct access *parent,
2127 bool allow_replacements)
2129 struct access *child;
2130 HOST_WIDE_INT limit = root->offset + root->size;
2131 HOST_WIDE_INT covered_to = root->offset;
2132 bool scalar = is_gimple_reg_type (root->type);
2133 bool hole = false, sth_created = false;
2137 if (parent->grp_read)
2139 if (parent->grp_assignment_read)
2140 root->grp_assignment_read = 1;
2141 if (parent->grp_write)
2142 root->grp_write = 1;
2143 if (parent->grp_assignment_write)
2144 root->grp_assignment_write = 1;
2145 if (parent->grp_total_scalarization)
2146 root->grp_total_scalarization = 1;
2149 if (root->grp_unscalarizable_region)
2150 allow_replacements = false;
2152 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2153 allow_replacements = false;
2155 for (child = root->first_child; child; child = child->next_sibling)
2157 hole |= covered_to < child->offset;
2158 sth_created |= analyze_access_subtree (child, root,
2159 allow_replacements && !scalar);
2161 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2162 root->grp_total_scalarization &= child->grp_total_scalarization;
2163 if (child->grp_covered)
2164 covered_to += child->size;
2169 if (allow_replacements && scalar && !root->first_child
2171 || ((root->grp_scalar_read || root->grp_assignment_read)
2172 && (root->grp_scalar_write || root->grp_assignment_write))))
2174 bool new_integer_type;
2175 if (TREE_CODE (root->type) == ENUMERAL_TYPE)
2177 tree rt = root->type;
2178 root->type = build_nonstandard_integer_type (TYPE_PRECISION (rt),
2179 TYPE_UNSIGNED (rt));
2180 new_integer_type = true;
2183 new_integer_type = false;
2185 if (dump_file && (dump_flags & TDF_DETAILS))
2187 fprintf (dump_file, "Marking ");
2188 print_generic_expr (dump_file, root->base, 0);
2189 fprintf (dump_file, " offset: %u, size: %u ",
2190 (unsigned) root->offset, (unsigned) root->size);
2191 fprintf (dump_file, " to be replaced%s.\n",
2192 new_integer_type ? " with an integer": "");
2195 root->grp_to_be_replaced = 1;
2201 if (covered_to < limit)
2204 root->grp_total_scalarization = 0;
2208 && (!hole || root->grp_total_scalarization))
2210 root->grp_covered = 1;
2213 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2214 root->grp_unscalarized_data = 1; /* not covered and written to */
2220 /* Analyze all access trees linked by next_grp by the means of
2221 analyze_access_subtree. */
2223 analyze_access_trees (struct access *access)
2229 if (analyze_access_subtree (access, NULL, true))
2231 access = access->next_grp;
2237 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2238 SIZE would conflict with an already existing one. If exactly such a child
2239 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2242 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2243 HOST_WIDE_INT size, struct access **exact_match)
2245 struct access *child;
2247 for (child = lacc->first_child; child; child = child->next_sibling)
2249 if (child->offset == norm_offset && child->size == size)
2251 *exact_match = child;
2255 if (child->offset < norm_offset + size
2256 && child->offset + child->size > norm_offset)
2263 /* Create a new child access of PARENT, with all properties just like MODEL
2264 except for its offset and with its grp_write false and grp_read true.
2265 Return the new access or NULL if it cannot be created. Note that this access
2266 is created long after all splicing and sorting, it's not located in any
2267 access vector and is automatically a representative of its group. */
2269 static struct access *
2270 create_artificial_child_access (struct access *parent, struct access *model,
2271 HOST_WIDE_INT new_offset)
2273 struct access *access;
2274 struct access **child;
2275 tree expr = parent->base;
2277 gcc_assert (!model->grp_unscalarizable_region);
2279 access = (struct access *) pool_alloc (access_pool);
2280 memset (access, 0, sizeof (struct access));
2281 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2284 access->grp_no_warning = true;
2285 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2286 new_offset, model, NULL, false);
2289 access->base = parent->base;
2290 access->expr = expr;
2291 access->offset = new_offset;
2292 access->size = model->size;
2293 access->type = model->type;
2294 access->grp_write = true;
2295 access->grp_read = false;
2297 child = &parent->first_child;
2298 while (*child && (*child)->offset < new_offset)
2299 child = &(*child)->next_sibling;
2301 access->next_sibling = *child;
2308 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2309 true if any new subaccess was created. Additionally, if RACC is a scalar
2310 access but LACC is not, change the type of the latter, if possible. */
2313 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2315 struct access *rchild;
2316 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2319 if (is_gimple_reg_type (lacc->type)
2320 || lacc->grp_unscalarizable_region
2321 || racc->grp_unscalarizable_region)
2324 if (is_gimple_reg_type (racc->type))
2326 if (!lacc->first_child && !racc->first_child)
2328 tree t = lacc->base;
2330 lacc->type = racc->type;
2331 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t),
2332 lacc->offset, racc->type))
2336 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2337 lacc->base, lacc->offset,
2339 lacc->grp_no_warning = true;
2345 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2347 struct access *new_acc = NULL;
2348 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2350 if (rchild->grp_unscalarizable_region)
2353 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2358 rchild->grp_hint = 1;
2359 new_acc->grp_hint |= new_acc->grp_read;
2360 if (rchild->first_child)
2361 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2366 rchild->grp_hint = 1;
2367 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2371 if (racc->first_child)
2372 propagate_subaccesses_across_link (new_acc, rchild);
2379 /* Propagate all subaccesses across assignment links. */
2382 propagate_all_subaccesses (void)
2384 while (work_queue_head)
2386 struct access *racc = pop_access_from_work_queue ();
2387 struct assign_link *link;
2389 gcc_assert (racc->first_link);
2391 for (link = racc->first_link; link; link = link->next)
2393 struct access *lacc = link->lacc;
2395 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2397 lacc = lacc->group_representative;
2398 if (propagate_subaccesses_across_link (lacc, racc)
2399 && lacc->first_link)
2400 add_access_to_work_queue (lacc);
2405 /* Go through all accesses collected throughout the (intraprocedural) analysis
2406 stage, exclude overlapping ones, identify representatives and build trees
2407 out of them, making decisions about scalarization on the way. Return true
2408 iff there are any to-be-scalarized variables after this stage. */
2411 analyze_all_variable_accesses (void)
2414 bitmap tmp = BITMAP_ALLOC (NULL);
2416 unsigned i, max_total_scalarization_size;
2418 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2419 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2421 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2422 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2423 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2425 tree var = referenced_var (i);
2427 if (TREE_CODE (var) == VAR_DECL
2428 && type_consists_of_records_p (TREE_TYPE (var)))
2430 if ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2431 <= max_total_scalarization_size)
2433 completely_scalarize_var (var);
2434 if (dump_file && (dump_flags & TDF_DETAILS))
2436 fprintf (dump_file, "Will attempt to totally scalarize ");
2437 print_generic_expr (dump_file, var, 0);
2438 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2441 else if (dump_file && (dump_flags & TDF_DETAILS))
2443 fprintf (dump_file, "Too big to totally scalarize: ");
2444 print_generic_expr (dump_file, var, 0);
2445 fprintf (dump_file, " (UID: %u)\n", DECL_UID (var));
2450 bitmap_copy (tmp, candidate_bitmap);
2451 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2453 tree var = referenced_var (i);
2454 struct access *access;
2456 access = sort_and_splice_var_accesses (var);
2457 if (!access || !build_access_trees (access))
2458 disqualify_candidate (var,
2459 "No or inhibitingly overlapping accesses.");
2462 propagate_all_subaccesses ();
2464 bitmap_copy (tmp, candidate_bitmap);
2465 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2467 tree var = referenced_var (i);
2468 struct access *access = get_first_repr_for_decl (var);
2470 if (analyze_access_trees (access))
2473 if (dump_file && (dump_flags & TDF_DETAILS))
2475 fprintf (dump_file, "\nAccess trees for ");
2476 print_generic_expr (dump_file, var, 0);
2477 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2478 dump_access_tree (dump_file, access);
2479 fprintf (dump_file, "\n");
2483 disqualify_candidate (var, "No scalar replacements to be created.");
2490 statistics_counter_event (cfun, "Scalarized aggregates", res);
2497 /* Generate statements copying scalar replacements of accesses within a subtree
2498 into or out of AGG. ACCESS, all its children, siblings and their children
2499 are to be processed. AGG is an aggregate type expression (can be a
2500 declaration but does not have to be, it can for example also be a mem_ref or
2501 a series of handled components). TOP_OFFSET is the offset of the processed
2502 subtree which has to be subtracted from offsets of individual accesses to
2503 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2504 replacements in the interval <start_offset, start_offset + chunk_size>,
2505 otherwise copy all. GSI is a statement iterator used to place the new
2506 statements. WRITE should be true when the statements should write from AGG
2507 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2508 statements will be added after the current statement in GSI, they will be
2509 added before the statement otherwise. */
2512 generate_subtree_copies (struct access *access, tree agg,
2513 HOST_WIDE_INT top_offset,
2514 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2515 gimple_stmt_iterator *gsi, bool write,
2516 bool insert_after, location_t loc)
2520 if (chunk_size && access->offset >= start_offset + chunk_size)
2523 if (access->grp_to_be_replaced
2525 || access->offset + access->size > start_offset))
2527 tree expr, repl = get_access_replacement (access);
2530 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2531 access, gsi, insert_after);
2535 if (access->grp_partial_lhs)
2536 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2538 insert_after ? GSI_NEW_STMT
2540 stmt = gimple_build_assign (repl, expr);
2544 TREE_NO_WARNING (repl) = 1;
2545 if (access->grp_partial_lhs)
2546 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2548 insert_after ? GSI_NEW_STMT
2550 stmt = gimple_build_assign (expr, repl);
2552 gimple_set_location (stmt, loc);
2555 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2557 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2559 sra_stats.subtree_copies++;
2562 if (access->first_child)
2563 generate_subtree_copies (access->first_child, agg, top_offset,
2564 start_offset, chunk_size, gsi,
2565 write, insert_after, loc);
2567 access = access->next_sibling;
2572 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2573 the root of the subtree to be processed. GSI is the statement iterator used
2574 for inserting statements which are added after the current statement if
2575 INSERT_AFTER is true or before it otherwise. */
2578 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2579 bool insert_after, location_t loc)
2582 struct access *child;
2584 if (access->grp_to_be_replaced)
2588 stmt = gimple_build_assign (get_access_replacement (access),
2589 build_zero_cst (access->type));
2591 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2593 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2595 gimple_set_location (stmt, loc);
2598 for (child = access->first_child; child; child = child->next_sibling)
2599 init_subtree_with_zero (child, gsi, insert_after, loc);
2602 /* Search for an access representative for the given expression EXPR and
2603 return it or NULL if it cannot be found. */
2605 static struct access *
2606 get_access_for_expr (tree expr)
2608 HOST_WIDE_INT offset, size, max_size;
2611 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2612 a different size than the size of its argument and we need the latter
2614 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2615 expr = TREE_OPERAND (expr, 0);
2617 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2618 if (max_size == -1 || !DECL_P (base))
2621 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2624 return get_var_base_offset_size_access (base, offset, max_size);
2627 /* Replace the expression EXPR with a scalar replacement if there is one and
2628 generate other statements to do type conversion or subtree copying if
2629 necessary. GSI is used to place newly created statements, WRITE is true if
2630 the expression is being written to (it is on a LHS of a statement or output
2631 in an assembly statement). */
2634 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2637 struct access *access;
2640 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2643 expr = &TREE_OPERAND (*expr, 0);
2648 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2649 expr = &TREE_OPERAND (*expr, 0);
2650 access = get_access_for_expr (*expr);
2653 type = TREE_TYPE (*expr);
2655 loc = gimple_location (gsi_stmt (*gsi));
2656 if (access->grp_to_be_replaced)
2658 tree repl = get_access_replacement (access);
2659 /* If we replace a non-register typed access simply use the original
2660 access expression to extract the scalar component afterwards.
2661 This happens if scalarizing a function return value or parameter
2662 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2663 gcc.c-torture/compile/20011217-1.c.
2665 We also want to use this when accessing a complex or vector which can
2666 be accessed as a different type too, potentially creating a need for
2667 type conversion (see PR42196) and when scalarized unions are involved
2668 in assembler statements (see PR42398). */
2669 if (!useless_type_conversion_p (type, access->type))
2673 ref = build_ref_for_model (loc, access->base, access->offset, access,
2680 if (access->grp_partial_lhs)
2681 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2682 false, GSI_NEW_STMT);
2683 stmt = gimple_build_assign (repl, ref);
2684 gimple_set_location (stmt, loc);
2685 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2691 if (access->grp_partial_lhs)
2692 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2693 true, GSI_SAME_STMT);
2694 stmt = gimple_build_assign (ref, repl);
2695 gimple_set_location (stmt, loc);
2696 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2704 if (access->first_child)
2706 HOST_WIDE_INT start_offset, chunk_size;
2708 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2709 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2711 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2712 start_offset = access->offset
2713 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2716 start_offset = chunk_size = 0;
2718 generate_subtree_copies (access->first_child, access->base, 0,
2719 start_offset, chunk_size, gsi, write, write,
2725 /* Where scalar replacements of the RHS have been written to when a replacement
2726 of a LHS of an assigments cannot be direclty loaded from a replacement of
2728 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2729 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2730 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2732 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2733 base aggregate if there are unscalarized data or directly to LHS of the
2734 statement that is pointed to by GSI otherwise. */
2736 static enum unscalarized_data_handling
2737 handle_unscalarized_data_in_subtree (struct access *top_racc,
2738 gimple_stmt_iterator *gsi)
2740 if (top_racc->grp_unscalarized_data)
2742 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2744 gimple_location (gsi_stmt (*gsi)));
2745 return SRA_UDH_RIGHT;
2749 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2750 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2751 0, 0, gsi, false, false,
2752 gimple_location (gsi_stmt (*gsi)));
2753 return SRA_UDH_LEFT;
2758 /* Try to generate statements to load all sub-replacements in an access subtree
2759 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2760 If that is not possible, refresh the TOP_RACC base aggregate and load the
2761 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2762 copied. NEW_GSI is stmt iterator used for statement insertions after the
2763 original assignment, OLD_GSI is used to insert statements before the
2764 assignment. *REFRESHED keeps the information whether we have needed to
2765 refresh replacements of the LHS and from which side of the assignments this
2769 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2770 HOST_WIDE_INT left_offset,
2771 gimple_stmt_iterator *old_gsi,
2772 gimple_stmt_iterator *new_gsi,
2773 enum unscalarized_data_handling *refreshed)
2775 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2776 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2778 if (lacc->grp_to_be_replaced)
2780 struct access *racc;
2781 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2785 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2786 if (racc && racc->grp_to_be_replaced)
2788 rhs = get_access_replacement (racc);
2789 if (!useless_type_conversion_p (lacc->type, racc->type))
2790 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2792 if (racc->grp_partial_lhs && lacc->grp_partial_lhs)
2793 rhs = force_gimple_operand_gsi (old_gsi, rhs, true, NULL_TREE,
2794 true, GSI_SAME_STMT);
2798 /* No suitable access on the right hand side, need to load from
2799 the aggregate. See if we have to update it first... */
2800 if (*refreshed == SRA_UDH_NONE)
2801 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2804 if (*refreshed == SRA_UDH_LEFT)
2805 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2808 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2810 if (lacc->grp_partial_lhs)
2811 rhs = force_gimple_operand_gsi (new_gsi, rhs, true, NULL_TREE,
2812 false, GSI_NEW_STMT);
2815 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2816 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2817 gimple_set_location (stmt, loc);
2819 sra_stats.subreplacements++;
2821 else if (*refreshed == SRA_UDH_NONE
2822 && lacc->grp_read && !lacc->grp_covered)
2823 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2826 if (lacc->first_child)
2827 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2828 old_gsi, new_gsi, refreshed);
2832 /* Result code for SRA assignment modification. */
2833 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2834 SRA_AM_MODIFIED, /* stmt changed but not
2836 SRA_AM_REMOVED }; /* stmt eliminated */
2838 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2839 to the assignment and GSI is the statement iterator pointing at it. Returns
2840 the same values as sra_modify_assign. */
2842 static enum assignment_mod_result
2843 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2845 tree lhs = gimple_assign_lhs (*stmt);
2849 acc = get_access_for_expr (lhs);
2853 if (gimple_clobber_p (*stmt))
2855 /* Remove clobbers of fully scalarized variables, otherwise
2857 if (acc->grp_covered)
2859 unlink_stmt_vdef (*stmt);
2860 gsi_remove (gsi, true);
2861 return SRA_AM_REMOVED;
2867 loc = gimple_location (*stmt);
2868 if (VEC_length (constructor_elt,
2869 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2871 /* I have never seen this code path trigger but if it can happen the
2872 following should handle it gracefully. */
2873 if (access_has_children_p (acc))
2874 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2876 return SRA_AM_MODIFIED;
2879 if (acc->grp_covered)
2881 init_subtree_with_zero (acc, gsi, false, loc);
2882 unlink_stmt_vdef (*stmt);
2883 gsi_remove (gsi, true);
2884 return SRA_AM_REMOVED;
2888 init_subtree_with_zero (acc, gsi, true, loc);
2889 return SRA_AM_MODIFIED;
2893 /* Create and return a new suitable default definition SSA_NAME for RACC which
2894 is an access describing an uninitialized part of an aggregate that is being
2898 get_repl_default_def_ssa_name (struct access *racc)
2902 decl = get_unrenamed_access_replacement (racc);
2904 repl = gimple_default_def (cfun, decl);
2907 repl = make_ssa_name (decl, gimple_build_nop ());
2908 set_default_def (decl, repl);
2914 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2918 contains_bitfld_comp_ref_p (const_tree ref)
2920 while (handled_component_p (ref))
2922 if (TREE_CODE (ref) == COMPONENT_REF
2923 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2925 ref = TREE_OPERAND (ref, 0);
2931 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2932 bit-field field declaration somewhere in it. */
2935 contains_vce_or_bfcref_p (const_tree ref)
2937 while (handled_component_p (ref))
2939 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2940 || (TREE_CODE (ref) == COMPONENT_REF
2941 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2943 ref = TREE_OPERAND (ref, 0);
2949 /* Examine both sides of the assignment statement pointed to by STMT, replace
2950 them with a scalare replacement if there is one and generate copying of
2951 replacements if scalarized aggregates have been used in the assignment. GSI
2952 is used to hold generated statements for type conversions and subtree
2955 static enum assignment_mod_result
2956 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2958 struct access *lacc, *racc;
2960 bool modify_this_stmt = false;
2961 bool force_gimple_rhs = false;
2963 gimple_stmt_iterator orig_gsi = *gsi;
2965 if (!gimple_assign_single_p (*stmt))
2967 lhs = gimple_assign_lhs (*stmt);
2968 rhs = gimple_assign_rhs1 (*stmt);
2970 if (TREE_CODE (rhs) == CONSTRUCTOR)
2971 return sra_modify_constructor_assign (stmt, gsi);
2973 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2974 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2975 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2977 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2979 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2981 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2984 lacc = get_access_for_expr (lhs);
2985 racc = get_access_for_expr (rhs);
2989 loc = gimple_location (*stmt);
2990 if (lacc && lacc->grp_to_be_replaced)
2992 lhs = get_access_replacement (lacc);
2993 gimple_assign_set_lhs (*stmt, lhs);
2994 modify_this_stmt = true;
2995 if (lacc->grp_partial_lhs)
2996 force_gimple_rhs = true;
3000 if (racc && racc->grp_to_be_replaced)
3002 rhs = get_access_replacement (racc);
3003 modify_this_stmt = true;
3004 if (racc->grp_partial_lhs)
3005 force_gimple_rhs = true;
3009 && !racc->grp_unscalarized_data
3010 && TREE_CODE (lhs) == SSA_NAME
3011 && !access_has_replacements_p (racc))
3013 rhs = get_repl_default_def_ssa_name (racc);
3014 modify_this_stmt = true;
3018 if (modify_this_stmt)
3020 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3022 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
3023 ??? This should move to fold_stmt which we simply should
3024 call after building a VIEW_CONVERT_EXPR here. */
3025 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
3026 && !contains_bitfld_comp_ref_p (lhs)
3027 && !access_has_children_p (lacc))
3029 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
3030 gimple_assign_set_lhs (*stmt, lhs);
3032 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
3033 && !contains_vce_or_bfcref_p (rhs)
3034 && !access_has_children_p (racc))
3035 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
3037 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3039 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
3041 if (is_gimple_reg_type (TREE_TYPE (lhs))
3042 && TREE_CODE (lhs) != SSA_NAME)
3043 force_gimple_rhs = true;
3048 /* From this point on, the function deals with assignments in between
3049 aggregates when at least one has scalar reductions of some of its
3050 components. There are three possible scenarios: Both the LHS and RHS have
3051 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
3053 In the first case, we would like to load the LHS components from RHS
3054 components whenever possible. If that is not possible, we would like to
3055 read it directly from the RHS (after updating it by storing in it its own
3056 components). If there are some necessary unscalarized data in the LHS,
3057 those will be loaded by the original assignment too. If neither of these
3058 cases happen, the original statement can be removed. Most of this is done
3059 by load_assign_lhs_subreplacements.
3061 In the second case, we would like to store all RHS scalarized components
3062 directly into LHS and if they cover the aggregate completely, remove the
3063 statement too. In the third case, we want the LHS components to be loaded
3064 directly from the RHS (DSE will remove the original statement if it
3067 This is a bit complex but manageable when types match and when unions do
3068 not cause confusion in a way that we cannot really load a component of LHS
3069 from the RHS or vice versa (the access representing this level can have
3070 subaccesses that are accessible only through a different union field at a
3071 higher level - different from the one used in the examined expression).
3074 Therefore, I specially handle a fourth case, happening when there is a
3075 specific type cast or it is impossible to locate a scalarized subaccess on
3076 the other side of the expression. If that happens, I simply "refresh" the
3077 RHS by storing in it is scalarized components leave the original statement
3078 there to do the copying and then load the scalar replacements of the LHS.
3079 This is what the first branch does. */
3081 if (modify_this_stmt
3082 || gimple_has_volatile_ops (*stmt)
3083 || contains_vce_or_bfcref_p (rhs)
3084 || contains_vce_or_bfcref_p (lhs))
3086 if (access_has_children_p (racc))
3087 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3088 gsi, false, false, loc);
3089 if (access_has_children_p (lacc))
3090 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
3091 gsi, true, true, loc);
3092 sra_stats.separate_lhs_rhs_handling++;
3094 /* This gimplification must be done after generate_subtree_copies,
3095 lest we insert the subtree copies in the middle of the gimplified
3097 if (force_gimple_rhs)
3098 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
3099 true, GSI_SAME_STMT);
3100 if (gimple_assign_rhs1 (*stmt) != rhs)
3102 modify_this_stmt = true;
3103 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
3104 gcc_assert (*stmt == gsi_stmt (orig_gsi));
3107 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3111 if (access_has_children_p (lacc) && access_has_children_p (racc))
3113 gimple_stmt_iterator orig_gsi = *gsi;
3114 enum unscalarized_data_handling refreshed;
3116 if (lacc->grp_read && !lacc->grp_covered)
3117 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
3119 refreshed = SRA_UDH_NONE;
3121 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
3122 &orig_gsi, gsi, &refreshed);
3123 if (refreshed != SRA_UDH_RIGHT)
3126 unlink_stmt_vdef (*stmt);
3127 gsi_remove (&orig_gsi, true);
3128 sra_stats.deleted++;
3129 return SRA_AM_REMOVED;
3134 if (access_has_children_p (racc)
3135 && !racc->grp_unscalarized_data)
3139 fprintf (dump_file, "Removing load: ");
3140 print_gimple_stmt (dump_file, *stmt, 0, 0);
3142 generate_subtree_copies (racc->first_child, lhs,
3143 racc->offset, 0, 0, gsi,
3145 gcc_assert (*stmt == gsi_stmt (*gsi));
3146 unlink_stmt_vdef (*stmt);
3147 gsi_remove (gsi, true);
3148 sra_stats.deleted++;
3149 return SRA_AM_REMOVED;
3151 /* Restore the aggregate RHS from its components so the
3152 prevailing aggregate copy does the right thing. */
3153 if (access_has_children_p (racc))
3154 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3155 gsi, false, false, loc);
3156 /* Re-load the components of the aggregate copy destination.
3157 But use the RHS aggregate to load from to expose more
3158 optimization opportunities. */
3159 if (access_has_children_p (lacc))
3160 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
3161 0, 0, gsi, true, true, loc);
3168 /* Traverse the function body and all modifications as decided in
3169 analyze_all_variable_accesses. Return true iff the CFG has been
3173 sra_modify_function_body (void)
3175 bool cfg_changed = false;
3180 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3181 while (!gsi_end_p (gsi))
3183 gimple stmt = gsi_stmt (gsi);
3184 enum assignment_mod_result assign_result;
3185 bool modified = false, deleted = false;
3189 switch (gimple_code (stmt))
3192 t = gimple_return_retval_ptr (stmt);
3193 if (*t != NULL_TREE)
3194 modified |= sra_modify_expr (t, &gsi, false);
3198 assign_result = sra_modify_assign (&stmt, &gsi);
3199 modified |= assign_result == SRA_AM_MODIFIED;
3200 deleted = assign_result == SRA_AM_REMOVED;
3204 /* Operands must be processed before the lhs. */
3205 for (i = 0; i < gimple_call_num_args (stmt); i++)
3207 t = gimple_call_arg_ptr (stmt, i);
3208 modified |= sra_modify_expr (t, &gsi, false);
3211 if (gimple_call_lhs (stmt))
3213 t = gimple_call_lhs_ptr (stmt);
3214 modified |= sra_modify_expr (t, &gsi, true);
3219 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3221 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3222 modified |= sra_modify_expr (t, &gsi, false);
3224 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3226 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3227 modified |= sra_modify_expr (t, &gsi, true);
3238 if (maybe_clean_eh_stmt (stmt)
3239 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3250 /* Generate statements initializing scalar replacements of parts of function
3254 initialize_parameter_reductions (void)
3256 gimple_stmt_iterator gsi;
3257 gimple_seq seq = NULL;
3260 for (parm = DECL_ARGUMENTS (current_function_decl);
3262 parm = DECL_CHAIN (parm))
3264 VEC (access_p, heap) *access_vec;
3265 struct access *access;
3267 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3269 access_vec = get_base_access_vector (parm);
3275 seq = gimple_seq_alloc ();
3276 gsi = gsi_start (seq);
3279 for (access = VEC_index (access_p, access_vec, 0);
3281 access = access->next_grp)
3282 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3283 EXPR_LOCATION (parm));
3287 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3290 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3291 it reveals there are components of some aggregates to be scalarized, it runs
3292 the required transformations. */
3294 perform_intra_sra (void)
3299 if (!find_var_candidates ())
3302 if (!scan_function ())
3305 if (!analyze_all_variable_accesses ())
3308 if (sra_modify_function_body ())
3309 ret = TODO_update_ssa | TODO_cleanup_cfg;
3311 ret = TODO_update_ssa;
3312 initialize_parameter_reductions ();
3314 statistics_counter_event (cfun, "Scalar replacements created",
3315 sra_stats.replacements);
3316 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3317 statistics_counter_event (cfun, "Subtree copy stmts",
3318 sra_stats.subtree_copies);
3319 statistics_counter_event (cfun, "Subreplacement stmts",
3320 sra_stats.subreplacements);
3321 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3322 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3323 sra_stats.separate_lhs_rhs_handling);
3326 sra_deinitialize ();
3330 /* Perform early intraprocedural SRA. */
3332 early_intra_sra (void)
3334 sra_mode = SRA_MODE_EARLY_INTRA;
3335 return perform_intra_sra ();
3338 /* Perform "late" intraprocedural SRA. */
3340 late_intra_sra (void)
3342 sra_mode = SRA_MODE_INTRA;
3343 return perform_intra_sra ();
3348 gate_intra_sra (void)
3350 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3354 struct gimple_opt_pass pass_sra_early =
3359 gate_intra_sra, /* gate */
3360 early_intra_sra, /* execute */
3363 0, /* static_pass_number */
3364 TV_TREE_SRA, /* tv_id */
3365 PROP_cfg | PROP_ssa, /* properties_required */
3366 0, /* properties_provided */
3367 0, /* properties_destroyed */
3368 0, /* todo_flags_start */
3371 | TODO_verify_ssa /* todo_flags_finish */
3375 struct gimple_opt_pass pass_sra =
3380 gate_intra_sra, /* gate */
3381 late_intra_sra, /* execute */
3384 0, /* static_pass_number */
3385 TV_TREE_SRA, /* tv_id */
3386 PROP_cfg | PROP_ssa, /* properties_required */
3387 0, /* properties_provided */
3388 0, /* properties_destroyed */
3389 TODO_update_address_taken, /* todo_flags_start */
3392 | TODO_verify_ssa /* todo_flags_finish */
3397 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3401 is_unused_scalar_param (tree parm)
3404 return (is_gimple_reg (parm)
3405 && (!(name = gimple_default_def (cfun, parm))
3406 || has_zero_uses (name)));
3409 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3410 examine whether there are any direct or otherwise infeasible ones. If so,
3411 return true, otherwise return false. PARM must be a gimple register with a
3412 non-NULL default definition. */
3415 ptr_parm_has_direct_uses (tree parm)
3417 imm_use_iterator ui;
3419 tree name = gimple_default_def (cfun, parm);
3422 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3425 use_operand_p use_p;
3427 if (is_gimple_debug (stmt))
3430 /* Valid uses include dereferences on the lhs and the rhs. */
3431 if (gimple_has_lhs (stmt))
3433 tree lhs = gimple_get_lhs (stmt);
3434 while (handled_component_p (lhs))
3435 lhs = TREE_OPERAND (lhs, 0);
3436 if (TREE_CODE (lhs) == MEM_REF
3437 && TREE_OPERAND (lhs, 0) == name
3438 && integer_zerop (TREE_OPERAND (lhs, 1))
3439 && types_compatible_p (TREE_TYPE (lhs),
3440 TREE_TYPE (TREE_TYPE (name)))
3441 && !TREE_THIS_VOLATILE (lhs))
3444 if (gimple_assign_single_p (stmt))
3446 tree rhs = gimple_assign_rhs1 (stmt);
3447 while (handled_component_p (rhs))
3448 rhs = TREE_OPERAND (rhs, 0);
3449 if (TREE_CODE (rhs) == MEM_REF
3450 && TREE_OPERAND (rhs, 0) == name
3451 && integer_zerop (TREE_OPERAND (rhs, 1))
3452 && types_compatible_p (TREE_TYPE (rhs),
3453 TREE_TYPE (TREE_TYPE (name)))
3454 && !TREE_THIS_VOLATILE (rhs))
3457 else if (is_gimple_call (stmt))
3460 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3462 tree arg = gimple_call_arg (stmt, i);
3463 while (handled_component_p (arg))
3464 arg = TREE_OPERAND (arg, 0);
3465 if (TREE_CODE (arg) == MEM_REF
3466 && TREE_OPERAND (arg, 0) == name
3467 && integer_zerop (TREE_OPERAND (arg, 1))
3468 && types_compatible_p (TREE_TYPE (arg),
3469 TREE_TYPE (TREE_TYPE (name)))
3470 && !TREE_THIS_VOLATILE (arg))
3475 /* If the number of valid uses does not match the number of
3476 uses in this stmt there is an unhandled use. */
3477 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3484 BREAK_FROM_IMM_USE_STMT (ui);
3490 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3491 them in candidate_bitmap. Note that these do not necessarily include
3492 parameter which are unused and thus can be removed. Return true iff any
3493 such candidate has been found. */
3496 find_param_candidates (void)
3503 for (parm = DECL_ARGUMENTS (current_function_decl);
3505 parm = DECL_CHAIN (parm))
3507 tree type = TREE_TYPE (parm);
3511 if (TREE_THIS_VOLATILE (parm)
3512 || TREE_ADDRESSABLE (parm)
3513 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3516 if (is_unused_scalar_param (parm))
3522 if (POINTER_TYPE_P (type))
3524 type = TREE_TYPE (type);
3526 if (TREE_CODE (type) == FUNCTION_TYPE
3527 || TYPE_VOLATILE (type)
3528 || (TREE_CODE (type) == ARRAY_TYPE
3529 && TYPE_NONALIASED_COMPONENT (type))
3530 || !is_gimple_reg (parm)
3531 || is_va_list_type (type)
3532 || ptr_parm_has_direct_uses (parm))
3535 else if (!AGGREGATE_TYPE_P (type))
3538 if (!COMPLETE_TYPE_P (type)
3539 || !host_integerp (TYPE_SIZE (type), 1)
3540 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3541 || (AGGREGATE_TYPE_P (type)
3542 && type_internals_preclude_sra_p (type, &msg)))
3545 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3547 if (dump_file && (dump_flags & TDF_DETAILS))
3549 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3550 print_generic_expr (dump_file, parm, 0);
3551 fprintf (dump_file, "\n");
3555 func_param_count = count;
3559 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3563 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3566 struct access *repr = (struct access *) data;
3568 repr->grp_maybe_modified = 1;
3572 /* Analyze what representatives (in linked lists accessible from
3573 REPRESENTATIVES) can be modified by side effects of statements in the
3574 current function. */
3577 analyze_modified_params (VEC (access_p, heap) *representatives)
3581 for (i = 0; i < func_param_count; i++)
3583 struct access *repr;
3585 for (repr = VEC_index (access_p, representatives, i);
3587 repr = repr->next_grp)
3589 struct access *access;
3593 if (no_accesses_p (repr))
3595 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3596 || repr->grp_maybe_modified)
3599 ao_ref_init (&ar, repr->expr);
3600 visited = BITMAP_ALLOC (NULL);
3601 for (access = repr; access; access = access->next_sibling)
3603 /* All accesses are read ones, otherwise grp_maybe_modified would
3604 be trivially set. */
3605 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3606 mark_maybe_modified, repr, &visited);
3607 if (repr->grp_maybe_modified)
3610 BITMAP_FREE (visited);
3615 /* Propagate distances in bb_dereferences in the opposite direction than the
3616 control flow edges, in each step storing the maximum of the current value
3617 and the minimum of all successors. These steps are repeated until the table
3618 stabilizes. Note that BBs which might terminate the functions (according to
3619 final_bbs bitmap) never updated in this way. */
3622 propagate_dereference_distances (void)
3624 VEC (basic_block, heap) *queue;
3627 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3628 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3631 VEC_quick_push (basic_block, queue, bb);
3635 while (!VEC_empty (basic_block, queue))
3639 bool change = false;
3642 bb = VEC_pop (basic_block, queue);
3645 if (bitmap_bit_p (final_bbs, bb->index))
3648 for (i = 0; i < func_param_count; i++)
3650 int idx = bb->index * func_param_count + i;
3652 HOST_WIDE_INT inh = 0;
3654 FOR_EACH_EDGE (e, ei, bb->succs)
3656 int succ_idx = e->dest->index * func_param_count + i;
3658 if (e->src == EXIT_BLOCK_PTR)
3664 inh = bb_dereferences [succ_idx];
3666 else if (bb_dereferences [succ_idx] < inh)
3667 inh = bb_dereferences [succ_idx];
3670 if (!first && bb_dereferences[idx] < inh)
3672 bb_dereferences[idx] = inh;
3677 if (change && !bitmap_bit_p (final_bbs, bb->index))
3678 FOR_EACH_EDGE (e, ei, bb->preds)
3683 e->src->aux = e->src;
3684 VEC_quick_push (basic_block, queue, e->src);
3688 VEC_free (basic_block, heap, queue);
3691 /* Dump a dereferences TABLE with heading STR to file F. */
3694 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3698 fprintf (dump_file, str);
3699 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3701 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3702 if (bb != EXIT_BLOCK_PTR)
3705 for (i = 0; i < func_param_count; i++)
3707 int idx = bb->index * func_param_count + i;
3708 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3713 fprintf (dump_file, "\n");
3716 /* Determine what (parts of) parameters passed by reference that are not
3717 assigned to are not certainly dereferenced in this function and thus the
3718 dereferencing cannot be safely moved to the caller without potentially
3719 introducing a segfault. Mark such REPRESENTATIVES as
3720 grp_not_necessarilly_dereferenced.
3722 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3723 part is calculated rather than simple booleans are calculated for each
3724 pointer parameter to handle cases when only a fraction of the whole
3725 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3728 The maximum dereference distances for each pointer parameter and BB are
3729 already stored in bb_dereference. This routine simply propagates these
3730 values upwards by propagate_dereference_distances and then compares the
3731 distances of individual parameters in the ENTRY BB to the equivalent
3732 distances of each representative of a (fraction of a) parameter. */
3735 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3739 if (dump_file && (dump_flags & TDF_DETAILS))
3740 dump_dereferences_table (dump_file,
3741 "Dereference table before propagation:\n",
3744 propagate_dereference_distances ();
3746 if (dump_file && (dump_flags & TDF_DETAILS))
3747 dump_dereferences_table (dump_file,
3748 "Dereference table after propagation:\n",
3751 for (i = 0; i < func_param_count; i++)
3753 struct access *repr = VEC_index (access_p, representatives, i);
3754 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3756 if (!repr || no_accesses_p (repr))
3761 if ((repr->offset + repr->size) > bb_dereferences[idx])
3762 repr->grp_not_necessarilly_dereferenced = 1;
3763 repr = repr->next_grp;
3769 /* Return the representative access for the parameter declaration PARM if it is
3770 a scalar passed by reference which is not written to and the pointer value
3771 is not used directly. Thus, if it is legal to dereference it in the caller
3772 and we can rule out modifications through aliases, such parameter should be
3773 turned into one passed by value. Return NULL otherwise. */
3775 static struct access *
3776 unmodified_by_ref_scalar_representative (tree parm)
3778 int i, access_count;
3779 struct access *repr;
3780 VEC (access_p, heap) *access_vec;
3782 access_vec = get_base_access_vector (parm);
3783 gcc_assert (access_vec);
3784 repr = VEC_index (access_p, access_vec, 0);
3787 repr->group_representative = repr;
3789 access_count = VEC_length (access_p, access_vec);
3790 for (i = 1; i < access_count; i++)
3792 struct access *access = VEC_index (access_p, access_vec, i);
3795 access->group_representative = repr;
3796 access->next_sibling = repr->next_sibling;
3797 repr->next_sibling = access;
3801 repr->grp_scalar_ptr = 1;
3805 /* Return true iff this access precludes IPA-SRA of the parameter it is
3809 access_precludes_ipa_sra_p (struct access *access)
3811 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3812 is incompatible assign in a call statement (and possibly even in asm
3813 statements). This can be relaxed by using a new temporary but only for
3814 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3815 intraprocedural SRA we deal with this by keeping the old aggregate around,
3816 something we cannot do in IPA-SRA.) */
3818 && (is_gimple_call (access->stmt)
3819 || gimple_code (access->stmt) == GIMPLE_ASM))
3822 if (STRICT_ALIGNMENT
3823 && tree_non_aligned_mem_p (access->expr, TYPE_ALIGN (access->type)))
3830 /* Sort collected accesses for parameter PARM, identify representatives for
3831 each accessed region and link them together. Return NULL if there are
3832 different but overlapping accesses, return the special ptr value meaning
3833 there are no accesses for this parameter if that is the case and return the
3834 first representative otherwise. Set *RO_GRP if there is a group of accesses
3835 with only read (i.e. no write) accesses. */
3837 static struct access *
3838 splice_param_accesses (tree parm, bool *ro_grp)
3840 int i, j, access_count, group_count;
3841 int agg_size, total_size = 0;
3842 struct access *access, *res, **prev_acc_ptr = &res;
3843 VEC (access_p, heap) *access_vec;
3845 access_vec = get_base_access_vector (parm);
3847 return &no_accesses_representant;
3848 access_count = VEC_length (access_p, access_vec);
3850 VEC_qsort (access_p, access_vec, compare_access_positions);
3855 while (i < access_count)
3859 access = VEC_index (access_p, access_vec, i);
3860 modification = access->write;
3861 if (access_precludes_ipa_sra_p (access))
3863 a1_alias_type = reference_alias_ptr_type (access->expr);
3865 /* Access is about to become group representative unless we find some
3866 nasty overlap which would preclude us from breaking this parameter
3870 while (j < access_count)
3872 struct access *ac2 = VEC_index (access_p, access_vec, j);
3873 if (ac2->offset != access->offset)
3875 /* All or nothing law for parameters. */
3876 if (access->offset + access->size > ac2->offset)
3881 else if (ac2->size != access->size)
3884 if (access_precludes_ipa_sra_p (ac2)
3885 || (ac2->type != access->type
3886 && (TREE_ADDRESSABLE (ac2->type)
3887 || TREE_ADDRESSABLE (access->type)))
3888 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3891 modification |= ac2->write;
3892 ac2->group_representative = access;
3893 ac2->next_sibling = access->next_sibling;
3894 access->next_sibling = ac2;
3899 access->grp_maybe_modified = modification;
3902 *prev_acc_ptr = access;
3903 prev_acc_ptr = &access->next_grp;
3904 total_size += access->size;
3908 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3909 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3911 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3912 if (total_size >= agg_size)
3915 gcc_assert (group_count > 0);
3919 /* Decide whether parameters with representative accesses given by REPR should
3920 be reduced into components. */
3923 decide_one_param_reduction (struct access *repr)
3925 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3930 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3931 gcc_assert (cur_parm_size > 0);
3933 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3936 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3941 agg_size = cur_parm_size;
3947 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3948 print_generic_expr (dump_file, parm, 0);
3949 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3950 for (acc = repr; acc; acc = acc->next_grp)
3951 dump_access (dump_file, acc, true);
3955 new_param_count = 0;
3957 for (; repr; repr = repr->next_grp)
3959 gcc_assert (parm == repr->base);
3961 /* Taking the address of a non-addressable field is verboten. */
3962 if (by_ref && repr->non_addressable)
3965 /* Do not decompose a non-BLKmode param in a way that would
3966 create BLKmode params. Especially for by-reference passing
3967 (thus, pointer-type param) this is hardly worthwhile. */
3968 if (DECL_MODE (parm) != BLKmode
3969 && TYPE_MODE (repr->type) == BLKmode)
3972 if (!by_ref || (!repr->grp_maybe_modified
3973 && !repr->grp_not_necessarilly_dereferenced))
3974 total_size += repr->size;
3976 total_size += cur_parm_size;
3981 gcc_assert (new_param_count > 0);
3983 if (optimize_function_for_size_p (cfun))
3984 parm_size_limit = cur_parm_size;
3986 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3989 if (total_size < agg_size
3990 && total_size <= parm_size_limit)
3993 fprintf (dump_file, " ....will be split into %i components\n",
3995 return new_param_count;
4001 /* The order of the following enums is important, we need to do extra work for
4002 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
4003 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
4004 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
4006 /* Identify representatives of all accesses to all candidate parameters for
4007 IPA-SRA. Return result based on what representatives have been found. */
4009 static enum ipa_splicing_result
4010 splice_all_param_accesses (VEC (access_p, heap) **representatives)
4012 enum ipa_splicing_result result = NO_GOOD_ACCESS;
4014 struct access *repr;
4016 *representatives = VEC_alloc (access_p, heap, func_param_count);
4018 for (parm = DECL_ARGUMENTS (current_function_decl);
4020 parm = DECL_CHAIN (parm))
4022 if (is_unused_scalar_param (parm))
4024 VEC_quick_push (access_p, *representatives,
4025 &no_accesses_representant);
4026 if (result == NO_GOOD_ACCESS)
4027 result = UNUSED_PARAMS;
4029 else if (POINTER_TYPE_P (TREE_TYPE (parm))
4030 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
4031 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4033 repr = unmodified_by_ref_scalar_representative (parm);
4034 VEC_quick_push (access_p, *representatives, repr);
4036 result = UNMODIF_BY_REF_ACCESSES;
4038 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4040 bool ro_grp = false;
4041 repr = splice_param_accesses (parm, &ro_grp);
4042 VEC_quick_push (access_p, *representatives, repr);
4044 if (repr && !no_accesses_p (repr))
4046 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4049 result = UNMODIF_BY_REF_ACCESSES;
4050 else if (result < MODIF_BY_REF_ACCESSES)
4051 result = MODIF_BY_REF_ACCESSES;
4053 else if (result < BY_VAL_ACCESSES)
4054 result = BY_VAL_ACCESSES;
4056 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
4057 result = UNUSED_PARAMS;
4060 VEC_quick_push (access_p, *representatives, NULL);
4063 if (result == NO_GOOD_ACCESS)
4065 VEC_free (access_p, heap, *representatives);
4066 *representatives = NULL;
4067 return NO_GOOD_ACCESS;
4073 /* Return the index of BASE in PARMS. Abort if it is not found. */
4076 get_param_index (tree base, VEC(tree, heap) *parms)
4080 len = VEC_length (tree, parms);
4081 for (i = 0; i < len; i++)
4082 if (VEC_index (tree, parms, i) == base)
4087 /* Convert the decisions made at the representative level into compact
4088 parameter adjustments. REPRESENTATIVES are pointers to first
4089 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4090 final number of adjustments. */
4092 static ipa_parm_adjustment_vec
4093 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
4094 int adjustments_count)
4096 VEC (tree, heap) *parms;
4097 ipa_parm_adjustment_vec adjustments;
4101 gcc_assert (adjustments_count > 0);
4102 parms = ipa_get_vector_of_formal_parms (current_function_decl);
4103 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
4104 parm = DECL_ARGUMENTS (current_function_decl);
4105 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
4107 struct access *repr = VEC_index (access_p, representatives, i);
4109 if (!repr || no_accesses_p (repr))
4111 struct ipa_parm_adjustment *adj;
4113 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4114 memset (adj, 0, sizeof (*adj));
4115 adj->base_index = get_param_index (parm, parms);
4118 adj->copy_param = 1;
4120 adj->remove_param = 1;
4124 struct ipa_parm_adjustment *adj;
4125 int index = get_param_index (parm, parms);
4127 for (; repr; repr = repr->next_grp)
4129 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4130 memset (adj, 0, sizeof (*adj));
4131 gcc_assert (repr->base == parm);
4132 adj->base_index = index;
4133 adj->base = repr->base;
4134 adj->type = repr->type;
4135 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
4136 adj->offset = repr->offset;
4137 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
4138 && (repr->grp_maybe_modified
4139 || repr->grp_not_necessarilly_dereferenced));
4144 VEC_free (tree, heap, parms);
4148 /* Analyze the collected accesses and produce a plan what to do with the
4149 parameters in the form of adjustments, NULL meaning nothing. */
4151 static ipa_parm_adjustment_vec
4152 analyze_all_param_acesses (void)
4154 enum ipa_splicing_result repr_state;
4155 bool proceed = false;
4156 int i, adjustments_count = 0;
4157 VEC (access_p, heap) *representatives;
4158 ipa_parm_adjustment_vec adjustments;
4160 repr_state = splice_all_param_accesses (&representatives);
4161 if (repr_state == NO_GOOD_ACCESS)
4164 /* If there are any parameters passed by reference which are not modified
4165 directly, we need to check whether they can be modified indirectly. */
4166 if (repr_state == UNMODIF_BY_REF_ACCESSES)
4168 analyze_caller_dereference_legality (representatives);
4169 analyze_modified_params (representatives);
4172 for (i = 0; i < func_param_count; i++)
4174 struct access *repr = VEC_index (access_p, representatives, i);
4176 if (repr && !no_accesses_p (repr))
4178 if (repr->grp_scalar_ptr)
4180 adjustments_count++;
4181 if (repr->grp_not_necessarilly_dereferenced
4182 || repr->grp_maybe_modified)
4183 VEC_replace (access_p, representatives, i, NULL);
4187 sra_stats.scalar_by_ref_to_by_val++;
4192 int new_components = decide_one_param_reduction (repr);
4194 if (new_components == 0)
4196 VEC_replace (access_p, representatives, i, NULL);
4197 adjustments_count++;
4201 adjustments_count += new_components;
4202 sra_stats.aggregate_params_reduced++;
4203 sra_stats.param_reductions_created += new_components;
4210 if (no_accesses_p (repr))
4213 sra_stats.deleted_unused_parameters++;
4215 adjustments_count++;
4219 if (!proceed && dump_file)
4220 fprintf (dump_file, "NOT proceeding to change params.\n");
4223 adjustments = turn_representatives_into_adjustments (representatives,
4228 VEC_free (access_p, heap, representatives);
4232 /* If a parameter replacement identified by ADJ does not yet exist in the form
4233 of declaration, create it and record it, otherwise return the previously
4237 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4240 if (!adj->new_ssa_base)
4242 char *pretty_name = make_fancy_name (adj->base);
4244 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4245 DECL_NAME (repl) = get_identifier (pretty_name);
4246 obstack_free (&name_obstack, pretty_name);
4248 add_referenced_var (repl);
4249 adj->new_ssa_base = repl;
4252 repl = adj->new_ssa_base;
4256 /* Find the first adjustment for a particular parameter BASE in a vector of
4257 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4260 static struct ipa_parm_adjustment *
4261 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4265 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4266 for (i = 0; i < len; i++)
4268 struct ipa_parm_adjustment *adj;
4270 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4271 if (!adj->copy_param && adj->base == base)
4278 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4279 removed because its value is not used, replace the SSA_NAME with a one
4280 relating to a created VAR_DECL together all of its uses and return true.
4281 ADJUSTMENTS is a pointer to an adjustments vector. */
4284 replace_removed_params_ssa_names (gimple stmt,
4285 ipa_parm_adjustment_vec adjustments)
4287 struct ipa_parm_adjustment *adj;
4288 tree lhs, decl, repl, name;
4290 if (gimple_code (stmt) == GIMPLE_PHI)
4291 lhs = gimple_phi_result (stmt);
4292 else if (is_gimple_assign (stmt))
4293 lhs = gimple_assign_lhs (stmt);
4294 else if (is_gimple_call (stmt))
4295 lhs = gimple_call_lhs (stmt);
4299 if (TREE_CODE (lhs) != SSA_NAME)
4301 decl = SSA_NAME_VAR (lhs);
4302 if (TREE_CODE (decl) != PARM_DECL)
4305 adj = get_adjustment_for_base (adjustments, decl);
4309 repl = get_replaced_param_substitute (adj);
4310 name = make_ssa_name (repl, stmt);
4314 fprintf (dump_file, "replacing an SSA name of a removed param ");
4315 print_generic_expr (dump_file, lhs, 0);
4316 fprintf (dump_file, " with ");
4317 print_generic_expr (dump_file, name, 0);
4318 fprintf (dump_file, "\n");
4321 if (is_gimple_assign (stmt))
4322 gimple_assign_set_lhs (stmt, name);
4323 else if (is_gimple_call (stmt))
4324 gimple_call_set_lhs (stmt, name);
4326 gimple_phi_set_result (stmt, name);
4328 replace_uses_by (lhs, name);
4329 release_ssa_name (lhs);
4333 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4334 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4335 specifies whether the function should care about type incompatibility the
4336 current and new expressions. If it is false, the function will leave
4337 incompatibility issues to the caller. Return true iff the expression
4341 sra_ipa_modify_expr (tree *expr, bool convert,
4342 ipa_parm_adjustment_vec adjustments)
4345 struct ipa_parm_adjustment *adj, *cand = NULL;
4346 HOST_WIDE_INT offset, size, max_size;
4349 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4351 if (TREE_CODE (*expr) == BIT_FIELD_REF
4352 || TREE_CODE (*expr) == IMAGPART_EXPR
4353 || TREE_CODE (*expr) == REALPART_EXPR)
4355 expr = &TREE_OPERAND (*expr, 0);
4359 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4360 if (!base || size == -1 || max_size == -1)
4363 if (TREE_CODE (base) == MEM_REF)
4365 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4366 base = TREE_OPERAND (base, 0);
4369 base = get_ssa_base_param (base);
4370 if (!base || TREE_CODE (base) != PARM_DECL)
4373 for (i = 0; i < len; i++)
4375 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4377 if (adj->base == base &&
4378 (adj->offset == offset || adj->remove_param))
4384 if (!cand || cand->copy_param || cand->remove_param)
4388 src = build_simple_mem_ref (cand->reduction);
4390 src = cand->reduction;
4392 if (dump_file && (dump_flags & TDF_DETAILS))
4394 fprintf (dump_file, "About to replace expr ");
4395 print_generic_expr (dump_file, *expr, 0);
4396 fprintf (dump_file, " with ");
4397 print_generic_expr (dump_file, src, 0);
4398 fprintf (dump_file, "\n");
4401 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4403 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4411 /* If the statement pointed to by STMT_PTR contains any expressions that need
4412 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4413 potential type incompatibilities (GSI is used to accommodate conversion
4414 statements and must point to the statement). Return true iff the statement
4418 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4419 ipa_parm_adjustment_vec adjustments)
4421 gimple stmt = *stmt_ptr;
4422 tree *lhs_p, *rhs_p;
4425 if (!gimple_assign_single_p (stmt))
4428 rhs_p = gimple_assign_rhs1_ptr (stmt);
4429 lhs_p = gimple_assign_lhs_ptr (stmt);
4431 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4432 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4435 tree new_rhs = NULL_TREE;
4437 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4439 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4441 /* V_C_Es of constructors can cause trouble (PR 42714). */
4442 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4443 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4445 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4448 new_rhs = fold_build1_loc (gimple_location (stmt),
4449 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4452 else if (REFERENCE_CLASS_P (*rhs_p)
4453 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4454 && !is_gimple_reg (*lhs_p))
4455 /* This can happen when an assignment in between two single field
4456 structures is turned into an assignment in between two pointers to
4457 scalars (PR 42237). */
4462 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4463 true, GSI_SAME_STMT);
4465 gimple_assign_set_rhs_from_tree (gsi, tmp);
4474 /* Traverse the function body and all modifications as described in
4475 ADJUSTMENTS. Return true iff the CFG has been changed. */
4478 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4480 bool cfg_changed = false;
4485 gimple_stmt_iterator gsi;
4487 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4488 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4490 gsi = gsi_start_bb (bb);
4491 while (!gsi_end_p (gsi))
4493 gimple stmt = gsi_stmt (gsi);
4494 bool modified = false;
4498 switch (gimple_code (stmt))
4501 t = gimple_return_retval_ptr (stmt);
4502 if (*t != NULL_TREE)
4503 modified |= sra_ipa_modify_expr (t, true, adjustments);
4507 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4508 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4512 /* Operands must be processed before the lhs. */
4513 for (i = 0; i < gimple_call_num_args (stmt); i++)
4515 t = gimple_call_arg_ptr (stmt, i);
4516 modified |= sra_ipa_modify_expr (t, true, adjustments);
4519 if (gimple_call_lhs (stmt))
4521 t = gimple_call_lhs_ptr (stmt);
4522 modified |= sra_ipa_modify_expr (t, false, adjustments);
4523 modified |= replace_removed_params_ssa_names (stmt,
4529 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4531 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4532 modified |= sra_ipa_modify_expr (t, true, adjustments);
4534 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4536 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4537 modified |= sra_ipa_modify_expr (t, false, adjustments);
4548 if (maybe_clean_eh_stmt (stmt)
4549 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4559 /* Call gimple_debug_bind_reset_value on all debug statements describing
4560 gimple register parameters that are being removed or replaced. */
4563 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4566 gimple_stmt_iterator *gsip = NULL, gsi;
4568 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR))
4570 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
4573 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4574 for (i = 0; i < len; i++)
4576 struct ipa_parm_adjustment *adj;
4577 imm_use_iterator ui;
4578 gimple stmt, def_temp;
4579 tree name, vexpr, copy = NULL_TREE;
4580 use_operand_p use_p;
4582 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4583 if (adj->copy_param || !is_gimple_reg (adj->base))
4585 name = gimple_default_def (cfun, adj->base);
4588 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4590 /* All other users must have been removed by
4591 ipa_sra_modify_function_body. */
4592 gcc_assert (is_gimple_debug (stmt));
4593 if (vexpr == NULL && gsip != NULL)
4595 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4596 vexpr = make_node (DEBUG_EXPR_DECL);
4597 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4599 DECL_ARTIFICIAL (vexpr) = 1;
4600 TREE_TYPE (vexpr) = TREE_TYPE (name);
4601 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4602 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4606 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4607 SET_USE (use_p, vexpr);
4610 gimple_debug_bind_reset_value (stmt);
4613 /* Create a VAR_DECL for debug info purposes. */
4614 if (!DECL_IGNORED_P (adj->base))
4616 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4617 VAR_DECL, DECL_NAME (adj->base),
4618 TREE_TYPE (adj->base));
4619 if (DECL_PT_UID_SET_P (adj->base))
4620 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4621 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4622 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4623 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4624 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4625 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4626 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4627 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4628 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4629 SET_DECL_RTL (copy, 0);
4630 TREE_USED (copy) = 1;
4631 DECL_CONTEXT (copy) = current_function_decl;
4632 add_referenced_var (copy);
4633 add_local_decl (cfun, copy);
4635 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4636 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4638 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4640 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4642 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4644 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4646 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4651 /* Return false iff all callers have at least as many actual arguments as there
4652 are formal parameters in the current function. */
4655 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4656 void *data ATTRIBUTE_UNUSED)
4658 struct cgraph_edge *cs;
4659 for (cs = node->callers; cs; cs = cs->next_caller)
4660 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4666 /* Convert all callers of NODE. */
4669 convert_callers_for_node (struct cgraph_node *node,
4672 ipa_parm_adjustment_vec adjustments = (ipa_parm_adjustment_vec)data;
4673 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4674 struct cgraph_edge *cs;
4676 for (cs = node->callers; cs; cs = cs->next_caller)
4678 current_function_decl = cs->caller->decl;
4679 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4682 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4683 cs->caller->uid, cs->callee->uid,
4684 cgraph_node_name (cs->caller),
4685 cgraph_node_name (cs->callee));
4687 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4692 for (cs = node->callers; cs; cs = cs->next_caller)
4693 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4694 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4695 compute_inline_parameters (cs->caller, true);
4696 BITMAP_FREE (recomputed_callers);
4701 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4704 convert_callers (struct cgraph_node *node, tree old_decl,
4705 ipa_parm_adjustment_vec adjustments)
4707 tree old_cur_fndecl = current_function_decl;
4708 basic_block this_block;
4710 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4711 adjustments, false);
4713 current_function_decl = old_cur_fndecl;
4715 if (!encountered_recursive_call)
4718 FOR_EACH_BB (this_block)
4720 gimple_stmt_iterator gsi;
4722 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4724 gimple stmt = gsi_stmt (gsi);
4726 if (gimple_code (stmt) != GIMPLE_CALL)
4728 call_fndecl = gimple_call_fndecl (stmt);
4729 if (call_fndecl == old_decl)
4732 fprintf (dump_file, "Adjusting recursive call");
4733 gimple_call_set_fndecl (stmt, node->decl);
4734 ipa_modify_call_arguments (NULL, stmt, adjustments);
4742 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4743 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4746 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4748 struct cgraph_node *new_node;
4750 VEC (cgraph_edge_p, heap) * redirect_callers = collect_callers_of_node (node);
4752 rebuild_cgraph_edges ();
4753 free_dominance_info (CDI_DOMINATORS);
4755 current_function_decl = NULL_TREE;
4757 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4758 false, NULL, NULL, "isra");
4759 current_function_decl = new_node->decl;
4760 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4762 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4763 cfg_changed = ipa_sra_modify_function_body (adjustments);
4764 sra_ipa_reset_debug_stmts (adjustments);
4765 convert_callers (new_node, node->decl, adjustments);
4766 cgraph_make_node_local (new_node);
4770 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4771 attributes, return true otherwise. NODE is the cgraph node of the current
4775 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4777 if (!cgraph_node_can_be_local_p (node))
4780 fprintf (dump_file, "Function not local to this compilation unit.\n");
4784 if (!node->local.can_change_signature)
4787 fprintf (dump_file, "Function can not change signature.\n");
4791 if (!tree_versionable_function_p (node->decl))
4794 fprintf (dump_file, "Function is not versionable.\n");
4798 if (DECL_VIRTUAL_P (current_function_decl))
4801 fprintf (dump_file, "Function is a virtual method.\n");
4805 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4806 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4809 fprintf (dump_file, "Function too big to be made truly local.\n");
4817 "Function has no callers in this compilation unit.\n");
4824 fprintf (dump_file, "Function uses stdarg. \n");
4828 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4834 /* Perform early interprocedural SRA. */
4837 ipa_early_sra (void)
4839 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4840 ipa_parm_adjustment_vec adjustments;
4843 if (!ipa_sra_preliminary_function_checks (node))
4847 sra_mode = SRA_MODE_EARLY_IPA;
4849 if (!find_param_candidates ())
4852 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4856 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4860 fprintf (dump_file, "There are callers with insufficient number of "
4865 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4867 * last_basic_block_for_function (cfun));
4868 final_bbs = BITMAP_ALLOC (NULL);
4871 if (encountered_apply_args)
4874 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4878 if (encountered_unchangable_recursive_call)
4881 fprintf (dump_file, "Function calls itself with insufficient "
4882 "number of arguments.\n");
4886 adjustments = analyze_all_param_acesses ();
4890 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4892 if (modify_function (node, adjustments))
4893 ret = TODO_update_ssa | TODO_cleanup_cfg;
4895 ret = TODO_update_ssa;
4896 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4898 statistics_counter_event (cfun, "Unused parameters deleted",
4899 sra_stats.deleted_unused_parameters);
4900 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4901 sra_stats.scalar_by_ref_to_by_val);
4902 statistics_counter_event (cfun, "Aggregate parameters broken up",
4903 sra_stats.aggregate_params_reduced);
4904 statistics_counter_event (cfun, "Aggregate parameter components created",
4905 sra_stats.param_reductions_created);
4908 BITMAP_FREE (final_bbs);
4909 free (bb_dereferences);
4911 sra_deinitialize ();
4915 /* Return if early ipa sra shall be performed. */
4917 ipa_early_sra_gate (void)
4919 return flag_ipa_sra && dbg_cnt (eipa_sra);
4922 struct gimple_opt_pass pass_early_ipa_sra =
4926 "eipa_sra", /* name */
4927 ipa_early_sra_gate, /* gate */
4928 ipa_early_sra, /* execute */
4931 0, /* static_pass_number */
4932 TV_IPA_SRA, /* tv_id */
4933 0, /* properties_required */
4934 0, /* properties_provided */
4935 0, /* properties_destroyed */
4936 0, /* todo_flags_start */
4937 TODO_dump_cgraph /* todo_flags_finish */