1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
4 Copyright (C) 2008, 2009, 2010 Free Software Foundation, Inc.
5 Contributed by Martin Jambor <mjambor@suse.cz>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* This file implements Scalar Reduction of Aggregates (SRA). SRA is run
24 twice, once in the early stages of compilation (early SRA) and once in the
25 late stages (late SRA). The aim of both is to turn references to scalar
26 parts of aggregates into uses of independent scalar variables.
28 The two passes are nearly identical, the only difference is that early SRA
29 does not scalarize unions which are used as the result in a GIMPLE_RETURN
30 statement because together with inlining this can lead to weird type
33 Both passes operate in four stages:
35 1. The declarations that have properties which make them candidates for
36 scalarization are identified in function find_var_candidates(). The
37 candidates are stored in candidate_bitmap.
39 2. The function body is scanned. In the process, declarations which are
40 used in a manner that prevent their scalarization are removed from the
41 candidate bitmap. More importantly, for every access into an aggregate,
42 an access structure (struct access) is created by create_access() and
43 stored in a vector associated with the aggregate. Among other
44 information, the aggregate declaration, the offset and size of the access
45 and its type are stored in the structure.
47 On a related note, assign_link structures are created for every assign
48 statement between candidate aggregates and attached to the related
51 3. The vectors of accesses are analyzed. They are first sorted according to
52 their offset and size and then scanned for partially overlapping accesses
53 (i.e. those which overlap but one is not entirely within another). Such
54 an access disqualifies the whole aggregate from being scalarized.
56 If there is no such inhibiting overlap, a representative access structure
57 is chosen for every unique combination of offset and size. Afterwards,
58 the pass builds a set of trees from these structures, in which children
59 of an access are within their parent (in terms of offset and size).
61 Then accesses are propagated whenever possible (i.e. in cases when it
62 does not create a partially overlapping access) across assign_links from
63 the right hand side to the left hand side.
65 Then the set of trees for each declaration is traversed again and those
66 accesses which should be replaced by a scalar are identified.
68 4. The function is traversed again, and for every reference into an
69 aggregate that has some component which is about to be scalarized,
70 statements are amended and new statements are created as necessary.
71 Finally, if a parameter got scalarized, the scalar replacements are
72 initialized with values from respective parameter aggregates. */
76 #include "coretypes.h"
77 #include "alloc-pool.h"
82 #include "tree-flow.h"
84 #include "tree-pretty-print.h"
85 #include "statistics.h"
86 #include "tree-dump.h"
92 /* Enumeration of all aggregate reductions we can do. */
93 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
94 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
95 SRA_MODE_INTRA }; /* late intraprocedural SRA */
97 /* Global variable describing which aggregate reduction we are performing at
99 static enum sra_mode sra_mode;
103 /* ACCESS represents each access to an aggregate variable (as a whole or a
104 part). It can also represent a group of accesses that refer to exactly the
105 same fragment of an aggregate (i.e. those that have exactly the same offset
106 and size). Such representatives for a single aggregate, once determined,
107 are linked in a linked list and have the group fields set.
109 Moreover, when doing intraprocedural SRA, a tree is built from those
110 representatives (by the means of first_child and next_sibling pointers), in
111 which all items in a subtree are "within" the root, i.e. their offset is
112 greater or equal to offset of the root and offset+size is smaller or equal
113 to offset+size of the root. Children of an access are sorted by offset.
115 Note that accesses to parts of vector and complex number types always
116 represented by an access to the whole complex number or a vector. It is a
117 duty of the modifying functions to replace them appropriately. */
121 /* Values returned by `get_ref_base_and_extent' for each component reference
122 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
123 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
124 HOST_WIDE_INT offset;
128 /* Expression. It is context dependent so do not use it to create new
129 expressions to access the original aggregate. See PR 42154 for a
135 /* The statement this access belongs to. */
138 /* Next group representative for this aggregate. */
139 struct access *next_grp;
141 /* Pointer to the group representative. Pointer to itself if the struct is
142 the representative. */
143 struct access *group_representative;
145 /* If this access has any children (in terms of the definition above), this
146 points to the first one. */
147 struct access *first_child;
149 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
150 described above. In IPA-SRA this is a pointer to the next access
151 belonging to the same group (having the same representative). */
152 struct access *next_sibling;
154 /* Pointers to the first and last element in the linked list of assign
156 struct assign_link *first_link, *last_link;
158 /* Pointer to the next access in the work queue. */
159 struct access *next_queued;
161 /* Replacement variable for this access "region." Never to be accessed
162 directly, always only by the means of get_access_replacement() and only
163 when grp_to_be_replaced flag is set. */
164 tree replacement_decl;
166 /* Is this particular access write access? */
169 /* Is this access an artificial one created to scalarize some record
171 unsigned total_scalarization : 1;
173 /* Is this access currently in the work queue? */
174 unsigned grp_queued : 1;
176 /* Does this group contain a write access? This flag is propagated down the
178 unsigned grp_write : 1;
180 /* Does this group contain a read access? This flag is propagated down the
182 unsigned grp_read : 1;
184 /* Does this group contain a read access that comes from an assignment
185 statement? This flag is propagated down the access tree. */
186 unsigned grp_assignment_read : 1;
188 /* Other passes of the analysis use this bit to make function
189 analyze_access_subtree create scalar replacements for this group if
191 unsigned grp_hint : 1;
193 /* Is the subtree rooted in this access fully covered by scalar
195 unsigned grp_covered : 1;
197 /* If set to true, this access and all below it in an access tree must not be
199 unsigned grp_unscalarizable_region : 1;
201 /* Whether data have been written to parts of the aggregate covered by this
202 access which is not to be scalarized. This flag is propagated up in the
204 unsigned grp_unscalarized_data : 1;
206 /* Does this access and/or group contain a write access through a
208 unsigned grp_partial_lhs : 1;
210 /* Set when a scalar replacement should be created for this variable. We do
211 the decision and creation at different places because create_tmp_var
212 cannot be called from within FOR_EACH_REFERENCED_VAR. */
213 unsigned grp_to_be_replaced : 1;
215 /* Is it possible that the group refers to data which might be (directly or
216 otherwise) modified? */
217 unsigned grp_maybe_modified : 1;
219 /* Set when this is a representative of a pointer to scalar (i.e. by
220 reference) parameter which we consider for turning into a plain scalar
221 (i.e. a by value parameter). */
222 unsigned grp_scalar_ptr : 1;
224 /* Set when we discover that this pointer is not safe to dereference in the
226 unsigned grp_not_necessarilly_dereferenced : 1;
229 typedef struct access *access_p;
231 DEF_VEC_P (access_p);
232 DEF_VEC_ALLOC_P (access_p, heap);
234 /* Alloc pool for allocating access structures. */
235 static alloc_pool access_pool;
237 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
238 are used to propagate subaccesses from rhs to lhs as long as they don't
239 conflict with what is already there. */
242 struct access *lacc, *racc;
243 struct assign_link *next;
246 /* Alloc pool for allocating assign link structures. */
247 static alloc_pool link_pool;
249 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
250 static struct pointer_map_t *base_access_vec;
252 /* Bitmap of candidates. */
253 static bitmap candidate_bitmap;
255 /* Bitmap of candidates which we should try to entirely scalarize away and
256 those which cannot be (because they are and need be used as a whole). */
257 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
259 /* Obstack for creation of fancy names. */
260 static struct obstack name_obstack;
262 /* Head of a linked list of accesses that need to have its subaccesses
263 propagated to their assignment counterparts. */
264 static struct access *work_queue_head;
266 /* Number of parameters of the analyzed function when doing early ipa SRA. */
267 static int func_param_count;
269 /* scan_function sets the following to true if it encounters a call to
270 __builtin_apply_args. */
271 static bool encountered_apply_args;
273 /* Set by scan_function when it finds a recursive call. */
274 static bool encountered_recursive_call;
276 /* Set by scan_function when it finds a recursive call with less actual
277 arguments than formal parameters.. */
278 static bool encountered_unchangable_recursive_call;
280 /* This is a table in which for each basic block and parameter there is a
281 distance (offset + size) in that parameter which is dereferenced and
282 accessed in that BB. */
283 static HOST_WIDE_INT *bb_dereferences;
284 /* Bitmap of BBs that can cause the function to "stop" progressing by
285 returning, throwing externally, looping infinitely or calling a function
286 which might abort etc.. */
287 static bitmap final_bbs;
289 /* Representative of no accesses at all. */
290 static struct access no_accesses_representant;
292 /* Predicate to test the special value. */
295 no_accesses_p (struct access *access)
297 return access == &no_accesses_representant;
300 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
301 representative fields are dumped, otherwise those which only describe the
302 individual access are. */
306 /* Number of processed aggregates is readily available in
307 analyze_all_variable_accesses and so is not stored here. */
309 /* Number of created scalar replacements. */
312 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
316 /* Number of statements created by generate_subtree_copies. */
319 /* Number of statements created by load_assign_lhs_subreplacements. */
322 /* Number of times sra_modify_assign has deleted a statement. */
325 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
326 RHS reparately due to type conversions or nonexistent matching
328 int separate_lhs_rhs_handling;
330 /* Number of parameters that were removed because they were unused. */
331 int deleted_unused_parameters;
333 /* Number of scalars passed as parameters by reference that have been
334 converted to be passed by value. */
335 int scalar_by_ref_to_by_val;
337 /* Number of aggregate parameters that were replaced by one or more of their
339 int aggregate_params_reduced;
341 /* Numbber of components created when splitting aggregate parameters. */
342 int param_reductions_created;
346 dump_access (FILE *f, struct access *access, bool grp)
348 fprintf (f, "access { ");
349 fprintf (f, "base = (%d)'", DECL_UID (access->base));
350 print_generic_expr (f, access->base, 0);
351 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
352 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
353 fprintf (f, ", expr = ");
354 print_generic_expr (f, access->expr, 0);
355 fprintf (f, ", type = ");
356 print_generic_expr (f, access->type, 0);
358 fprintf (f, ", grp_write = %d, total_scalarization = %d, "
359 "grp_read = %d, grp_hint = %d, grp_assignment_read = %d,"
360 "grp_covered = %d, grp_unscalarizable_region = %d, "
361 "grp_unscalarized_data = %d, grp_partial_lhs = %d, "
362 "grp_to_be_replaced = %d, grp_maybe_modified = %d, "
363 "grp_not_necessarilly_dereferenced = %d\n",
364 access->grp_write, access->total_scalarization,
365 access->grp_read, access->grp_hint, access->grp_assignment_read,
366 access->grp_covered, access->grp_unscalarizable_region,
367 access->grp_unscalarized_data, access->grp_partial_lhs,
368 access->grp_to_be_replaced, access->grp_maybe_modified,
369 access->grp_not_necessarilly_dereferenced);
371 fprintf (f, ", write = %d, total_scalarization = %d, "
372 "grp_partial_lhs = %d\n",
373 access->write, access->total_scalarization,
374 access->grp_partial_lhs);
377 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
380 dump_access_tree_1 (FILE *f, struct access *access, int level)
386 for (i = 0; i < level; i++)
387 fputs ("* ", dump_file);
389 dump_access (f, access, true);
391 if (access->first_child)
392 dump_access_tree_1 (f, access->first_child, level + 1);
394 access = access->next_sibling;
399 /* Dump all access trees for a variable, given the pointer to the first root in
403 dump_access_tree (FILE *f, struct access *access)
405 for (; access; access = access->next_grp)
406 dump_access_tree_1 (f, access, 0);
409 /* Return true iff ACC is non-NULL and has subaccesses. */
412 access_has_children_p (struct access *acc)
414 return acc && acc->first_child;
417 /* Return a vector of pointers to accesses for the variable given in BASE or
418 NULL if there is none. */
420 static VEC (access_p, heap) *
421 get_base_access_vector (tree base)
425 slot = pointer_map_contains (base_access_vec, base);
429 return *(VEC (access_p, heap) **) slot;
432 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
433 in ACCESS. Return NULL if it cannot be found. */
435 static struct access *
436 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
439 while (access && (access->offset != offset || access->size != size))
441 struct access *child = access->first_child;
443 while (child && (child->offset + child->size <= offset))
444 child = child->next_sibling;
451 /* Return the first group representative for DECL or NULL if none exists. */
453 static struct access *
454 get_first_repr_for_decl (tree base)
456 VEC (access_p, heap) *access_vec;
458 access_vec = get_base_access_vector (base);
462 return VEC_index (access_p, access_vec, 0);
465 /* Find an access representative for the variable BASE and given OFFSET and
466 SIZE. Requires that access trees have already been built. Return NULL if
467 it cannot be found. */
469 static struct access *
470 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
473 struct access *access;
475 access = get_first_repr_for_decl (base);
476 while (access && (access->offset + access->size <= offset))
477 access = access->next_grp;
481 return find_access_in_subtree (access, offset, size);
484 /* Add LINK to the linked list of assign links of RACC. */
486 add_link_to_rhs (struct access *racc, struct assign_link *link)
488 gcc_assert (link->racc == racc);
490 if (!racc->first_link)
492 gcc_assert (!racc->last_link);
493 racc->first_link = link;
496 racc->last_link->next = link;
498 racc->last_link = link;
502 /* Move all link structures in their linked list in OLD_RACC to the linked list
505 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
507 if (!old_racc->first_link)
509 gcc_assert (!old_racc->last_link);
513 if (new_racc->first_link)
515 gcc_assert (!new_racc->last_link->next);
516 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
518 new_racc->last_link->next = old_racc->first_link;
519 new_racc->last_link = old_racc->last_link;
523 gcc_assert (!new_racc->last_link);
525 new_racc->first_link = old_racc->first_link;
526 new_racc->last_link = old_racc->last_link;
528 old_racc->first_link = old_racc->last_link = NULL;
531 /* Add ACCESS to the work queue (which is actually a stack). */
534 add_access_to_work_queue (struct access *access)
536 if (!access->grp_queued)
538 gcc_assert (!access->next_queued);
539 access->next_queued = work_queue_head;
540 access->grp_queued = 1;
541 work_queue_head = access;
545 /* Pop an access from the work queue, and return it, assuming there is one. */
547 static struct access *
548 pop_access_from_work_queue (void)
550 struct access *access = work_queue_head;
552 work_queue_head = access->next_queued;
553 access->next_queued = NULL;
554 access->grp_queued = 0;
559 /* Allocate necessary structures. */
562 sra_initialize (void)
564 candidate_bitmap = BITMAP_ALLOC (NULL);
565 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
566 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
567 gcc_obstack_init (&name_obstack);
568 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
569 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
570 base_access_vec = pointer_map_create ();
571 memset (&sra_stats, 0, sizeof (sra_stats));
572 encountered_apply_args = false;
573 encountered_recursive_call = false;
574 encountered_unchangable_recursive_call = false;
577 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
580 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
581 void *data ATTRIBUTE_UNUSED)
583 VEC (access_p, heap) *access_vec;
584 access_vec = (VEC (access_p, heap) *) *value;
585 VEC_free (access_p, heap, access_vec);
590 /* Deallocate all general structures. */
593 sra_deinitialize (void)
595 BITMAP_FREE (candidate_bitmap);
596 BITMAP_FREE (should_scalarize_away_bitmap);
597 BITMAP_FREE (cannot_scalarize_away_bitmap);
598 free_alloc_pool (access_pool);
599 free_alloc_pool (link_pool);
600 obstack_free (&name_obstack, NULL);
602 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
603 pointer_map_destroy (base_access_vec);
606 /* Remove DECL from candidates for SRA and write REASON to the dump file if
609 disqualify_candidate (tree decl, const char *reason)
611 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
613 if (dump_file && (dump_flags & TDF_DETAILS))
615 fprintf (dump_file, "! Disqualifying ");
616 print_generic_expr (dump_file, decl, 0);
617 fprintf (dump_file, " - %s\n", reason);
621 /* Return true iff the type contains a field or an element which does not allow
625 type_internals_preclude_sra_p (tree type)
630 switch (TREE_CODE (type))
634 case QUAL_UNION_TYPE:
635 for (fld = TYPE_FIELDS (type); fld; fld = TREE_CHAIN (fld))
636 if (TREE_CODE (fld) == FIELD_DECL)
638 tree ft = TREE_TYPE (fld);
640 if (TREE_THIS_VOLATILE (fld)
641 || !DECL_FIELD_OFFSET (fld) || !DECL_SIZE (fld)
642 || !host_integerp (DECL_FIELD_OFFSET (fld), 1)
643 || !host_integerp (DECL_SIZE (fld), 1))
646 if (AGGREGATE_TYPE_P (ft)
647 && type_internals_preclude_sra_p (ft))
654 et = TREE_TYPE (type);
656 if (AGGREGATE_TYPE_P (et))
657 return type_internals_preclude_sra_p (et);
666 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
667 base variable if it is. Return T if it is not an SSA_NAME. */
670 get_ssa_base_param (tree t)
672 if (TREE_CODE (t) == SSA_NAME)
674 if (SSA_NAME_IS_DEFAULT_DEF (t))
675 return SSA_NAME_VAR (t);
682 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
683 belongs to, unless the BB has already been marked as a potentially
687 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
689 basic_block bb = gimple_bb (stmt);
690 int idx, parm_index = 0;
693 if (bitmap_bit_p (final_bbs, bb->index))
696 for (parm = DECL_ARGUMENTS (current_function_decl);
697 parm && parm != base;
698 parm = TREE_CHAIN (parm))
701 gcc_assert (parm_index < func_param_count);
703 idx = bb->index * func_param_count + parm_index;
704 if (bb_dereferences[idx] < dist)
705 bb_dereferences[idx] = dist;
708 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
709 the three fields. Also add it to the vector of accesses corresponding to
710 the base. Finally, return the new access. */
712 static struct access *
713 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
715 VEC (access_p, heap) *vec;
716 struct access *access;
719 access = (struct access *) pool_alloc (access_pool);
720 memset (access, 0, sizeof (struct access));
722 access->offset = offset;
725 slot = pointer_map_contains (base_access_vec, base);
727 vec = (VEC (access_p, heap) *) *slot;
729 vec = VEC_alloc (access_p, heap, 32);
731 VEC_safe_push (access_p, heap, vec, access);
733 *((struct VEC (access_p,heap) **)
734 pointer_map_insert (base_access_vec, base)) = vec;
739 /* Create and insert access for EXPR. Return created access, or NULL if it is
742 static struct access *
743 create_access (tree expr, gimple stmt, bool write)
745 struct access *access;
746 HOST_WIDE_INT offset, size, max_size;
748 bool ptr, unscalarizable_region = false;
750 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
752 if (sra_mode == SRA_MODE_EARLY_IPA && INDIRECT_REF_P (base))
754 base = get_ssa_base_param (TREE_OPERAND (base, 0));
762 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
765 if (sra_mode == SRA_MODE_EARLY_IPA)
767 if (size < 0 || size != max_size)
769 disqualify_candidate (base, "Encountered a variable sized access.");
772 if ((offset % BITS_PER_UNIT) != 0 || (size % BITS_PER_UNIT) != 0)
774 disqualify_candidate (base,
775 "Encountered an acces not aligned to a byte.");
780 mark_parm_dereference (base, offset + size, stmt);
784 if (size != max_size)
787 unscalarizable_region = true;
791 disqualify_candidate (base, "Encountered an unconstrained access.");
796 access = create_access_1 (base, offset, size);
798 access->type = TREE_TYPE (expr);
799 access->write = write;
800 access->grp_unscalarizable_region = unscalarizable_region;
807 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
808 register types or (recursively) records with only these two kinds of fields.
809 It also returns false if any of these records has a zero-size field as its
813 type_consists_of_records_p (tree type)
816 bool last_fld_has_zero_size = false;
818 if (TREE_CODE (type) != RECORD_TYPE)
821 for (fld = TYPE_FIELDS (type); fld; fld = TREE_CHAIN (fld))
822 if (TREE_CODE (fld) == FIELD_DECL)
824 tree ft = TREE_TYPE (fld);
826 if (!is_gimple_reg_type (ft)
827 && !type_consists_of_records_p (ft))
830 last_fld_has_zero_size = tree_low_cst (DECL_SIZE (fld), 1) == 0;
833 if (last_fld_has_zero_size)
839 /* Create total_scalarization accesses for all scalar type fields in DECL that
840 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
841 must be the top-most VAR_DECL representing the variable, OFFSET must be the
842 offset of DECL within BASE. */
845 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset)
847 tree fld, decl_type = TREE_TYPE (decl);
849 for (fld = TYPE_FIELDS (decl_type); fld; fld = TREE_CHAIN (fld))
850 if (TREE_CODE (fld) == FIELD_DECL)
852 HOST_WIDE_INT pos = offset + int_bit_position (fld);
853 tree ft = TREE_TYPE (fld);
855 if (is_gimple_reg_type (ft))
857 struct access *access;
862 size = tree_low_cst (DECL_SIZE (fld), 1);
864 ok = build_ref_for_offset (&expr, TREE_TYPE (base), pos,
868 access = create_access_1 (base, pos, size);
871 access->total_scalarization = 1;
872 /* Accesses for intraprocedural SRA can have their stmt NULL. */
875 completely_scalarize_record (base, fld, pos);
880 /* Search the given tree for a declaration by skipping handled components and
881 exclude it from the candidates. */
884 disqualify_base_of_expr (tree t, const char *reason)
886 while (handled_component_p (t))
887 t = TREE_OPERAND (t, 0);
889 if (sra_mode == SRA_MODE_EARLY_IPA)
891 if (INDIRECT_REF_P (t))
892 t = TREE_OPERAND (t, 0);
893 t = get_ssa_base_param (t);
897 disqualify_candidate (t, reason);
900 /* Scan expression EXPR and create access structures for all accesses to
901 candidates for scalarization. Return the created access or NULL if none is
904 static struct access *
905 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
907 struct access *ret = NULL;
910 if (TREE_CODE (expr) == BIT_FIELD_REF
911 || TREE_CODE (expr) == IMAGPART_EXPR
912 || TREE_CODE (expr) == REALPART_EXPR)
914 expr = TREE_OPERAND (expr, 0);
920 /* We need to dive through V_C_Es in order to get the size of its parameter
921 and not the result type. Ada produces such statements. We are also
922 capable of handling the topmost V_C_E but not any of those buried in other
923 handled components. */
924 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
925 expr = TREE_OPERAND (expr, 0);
927 if (contains_view_convert_expr_p (expr))
929 disqualify_base_of_expr (expr, "V_C_E under a different handled "
934 switch (TREE_CODE (expr))
937 if (sra_mode != SRA_MODE_EARLY_IPA)
945 case ARRAY_RANGE_REF:
946 ret = create_access (expr, stmt, write);
953 if (write && partial_ref && ret)
954 ret->grp_partial_lhs = 1;
959 /* Scan expression EXPR and create access structures for all accesses to
960 candidates for scalarization. Return true if any access has been inserted.
961 STMT must be the statement from which the expression is taken, WRITE must be
962 true if the expression is a store and false otherwise. */
965 build_access_from_expr (tree expr, gimple stmt, bool write)
967 struct access *access;
969 access = build_access_from_expr_1 (expr, stmt, write);
972 /* This means the aggregate is accesses as a whole in a way other than an
973 assign statement and thus cannot be removed even if we had a scalar
974 replacement for everything. */
975 if (cannot_scalarize_away_bitmap)
976 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
982 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
983 modes in which it matters, return true iff they have been disqualified. RHS
984 may be NULL, in that case ignore it. If we scalarize an aggregate in
985 intra-SRA we may need to add statements after each statement. This is not
986 possible if a statement unconditionally has to end the basic block. */
988 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
990 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
991 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
993 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
995 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1001 /* Scan expressions occuring in STMT, create access structures for all accesses
1002 to candidates for scalarization and remove those candidates which occur in
1003 statements or expressions that prevent them from being split apart. Return
1004 true if any access has been inserted. */
1007 build_accesses_from_assign (gimple stmt)
1010 struct access *lacc, *racc;
1012 if (!gimple_assign_single_p (stmt))
1015 lhs = gimple_assign_lhs (stmt);
1016 rhs = gimple_assign_rhs1 (stmt);
1018 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1021 racc = build_access_from_expr_1 (rhs, stmt, false);
1022 lacc = build_access_from_expr_1 (lhs, stmt, true);
1026 racc->grp_assignment_read = 1;
1027 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1028 && !is_gimple_reg_type (racc->type))
1029 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1033 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1034 && !lacc->grp_unscalarizable_region
1035 && !racc->grp_unscalarizable_region
1036 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1037 /* FIXME: Turn the following line into an assert after PR 40058 is
1039 && lacc->size == racc->size
1040 && useless_type_conversion_p (lacc->type, racc->type))
1042 struct assign_link *link;
1044 link = (struct assign_link *) pool_alloc (link_pool);
1045 memset (link, 0, sizeof (struct assign_link));
1050 add_link_to_rhs (racc, link);
1053 return lacc || racc;
1056 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1057 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1060 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1061 void *data ATTRIBUTE_UNUSED)
1063 op = get_base_address (op);
1066 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1071 /* Return true iff callsite CALL has at least as many actual arguments as there
1072 are formal parameters of the function currently processed by IPA-SRA. */
1075 callsite_has_enough_arguments_p (gimple call)
1077 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1080 /* Scan function and look for interesting expressions and create access
1081 structures for them. Return true iff any access is created. */
1084 scan_function (void)
1091 gimple_stmt_iterator gsi;
1092 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1094 gimple stmt = gsi_stmt (gsi);
1098 if (final_bbs && stmt_can_throw_external (stmt))
1099 bitmap_set_bit (final_bbs, bb->index);
1100 switch (gimple_code (stmt))
1103 t = gimple_return_retval (stmt);
1105 ret |= build_access_from_expr (t, stmt, false);
1107 bitmap_set_bit (final_bbs, bb->index);
1111 ret |= build_accesses_from_assign (stmt);
1115 for (i = 0; i < gimple_call_num_args (stmt); i++)
1116 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1119 if (sra_mode == SRA_MODE_EARLY_IPA)
1121 tree dest = gimple_call_fndecl (stmt);
1122 int flags = gimple_call_flags (stmt);
1126 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1127 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1128 encountered_apply_args = true;
1129 if (cgraph_get_node (dest)
1130 == cgraph_get_node (current_function_decl))
1132 encountered_recursive_call = true;
1133 if (!callsite_has_enough_arguments_p (stmt))
1134 encountered_unchangable_recursive_call = true;
1139 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1140 bitmap_set_bit (final_bbs, bb->index);
1143 t = gimple_call_lhs (stmt);
1144 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1145 ret |= build_access_from_expr (t, stmt, true);
1149 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1152 bitmap_set_bit (final_bbs, bb->index);
1154 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1156 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1157 ret |= build_access_from_expr (t, stmt, false);
1159 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1161 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1162 ret |= build_access_from_expr (t, stmt, true);
1175 /* Helper of QSORT function. There are pointers to accesses in the array. An
1176 access is considered smaller than another if it has smaller offset or if the
1177 offsets are the same but is size is bigger. */
1180 compare_access_positions (const void *a, const void *b)
1182 const access_p *fp1 = (const access_p *) a;
1183 const access_p *fp2 = (const access_p *) b;
1184 const access_p f1 = *fp1;
1185 const access_p f2 = *fp2;
1187 if (f1->offset != f2->offset)
1188 return f1->offset < f2->offset ? -1 : 1;
1190 if (f1->size == f2->size)
1192 if (f1->type == f2->type)
1194 /* Put any non-aggregate type before any aggregate type. */
1195 else if (!is_gimple_reg_type (f1->type)
1196 && is_gimple_reg_type (f2->type))
1198 else if (is_gimple_reg_type (f1->type)
1199 && !is_gimple_reg_type (f2->type))
1201 /* Put any complex or vector type before any other scalar type. */
1202 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1203 && TREE_CODE (f1->type) != VECTOR_TYPE
1204 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1205 || TREE_CODE (f2->type) == VECTOR_TYPE))
1207 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1208 || TREE_CODE (f1->type) == VECTOR_TYPE)
1209 && TREE_CODE (f2->type) != COMPLEX_TYPE
1210 && TREE_CODE (f2->type) != VECTOR_TYPE)
1212 /* Put the integral type with the bigger precision first. */
1213 else if (INTEGRAL_TYPE_P (f1->type)
1214 && INTEGRAL_TYPE_P (f2->type))
1215 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1216 /* Put any integral type with non-full precision last. */
1217 else if (INTEGRAL_TYPE_P (f1->type)
1218 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1219 != TYPE_PRECISION (f1->type)))
1221 else if (INTEGRAL_TYPE_P (f2->type)
1222 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1223 != TYPE_PRECISION (f2->type)))
1225 /* Stabilize the sort. */
1226 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1229 /* We want the bigger accesses first, thus the opposite operator in the next
1231 return f1->size > f2->size ? -1 : 1;
1235 /* Append a name of the declaration to the name obstack. A helper function for
1239 make_fancy_decl_name (tree decl)
1243 tree name = DECL_NAME (decl);
1245 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1246 IDENTIFIER_LENGTH (name));
1249 sprintf (buffer, "D%u", DECL_UID (decl));
1250 obstack_grow (&name_obstack, buffer, strlen (buffer));
1254 /* Helper for make_fancy_name. */
1257 make_fancy_name_1 (tree expr)
1264 make_fancy_decl_name (expr);
1268 switch (TREE_CODE (expr))
1271 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1272 obstack_1grow (&name_obstack, '$');
1273 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1277 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1278 obstack_1grow (&name_obstack, '$');
1279 /* Arrays with only one element may not have a constant as their
1281 index = TREE_OPERAND (expr, 1);
1282 if (TREE_CODE (index) != INTEGER_CST)
1284 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1285 obstack_grow (&name_obstack, buffer, strlen (buffer));
1292 gcc_unreachable (); /* we treat these as scalars. */
1299 /* Create a human readable name for replacement variable of ACCESS. */
1302 make_fancy_name (tree expr)
1304 make_fancy_name_1 (expr);
1305 obstack_1grow (&name_obstack, '\0');
1306 return XOBFINISH (&name_obstack, char *);
1309 /* Helper function for build_ref_for_offset. */
1312 build_ref_for_offset_1 (tree *res, tree type, HOST_WIDE_INT offset,
1318 tree tr_size, index, minidx;
1319 HOST_WIDE_INT el_size;
1321 if (offset == 0 && exp_type
1322 && types_compatible_p (exp_type, type))
1325 switch (TREE_CODE (type))
1328 case QUAL_UNION_TYPE:
1330 for (fld = TYPE_FIELDS (type); fld; fld = TREE_CHAIN (fld))
1332 HOST_WIDE_INT pos, size;
1333 tree expr, *expr_ptr;
1335 if (TREE_CODE (fld) != FIELD_DECL)
1338 pos = int_bit_position (fld);
1339 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1340 tr_size = DECL_SIZE (fld);
1341 if (!tr_size || !host_integerp (tr_size, 1))
1343 size = tree_low_cst (tr_size, 1);
1349 else if (pos > offset || (pos + size) <= offset)
1354 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1360 if (build_ref_for_offset_1 (expr_ptr, TREE_TYPE (fld),
1361 offset - pos, exp_type))
1371 tr_size = TYPE_SIZE (TREE_TYPE (type));
1372 if (!tr_size || !host_integerp (tr_size, 1))
1374 el_size = tree_low_cst (tr_size, 1);
1376 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1377 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1381 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1382 if (!integer_zerop (minidx))
1383 index = int_const_binop (PLUS_EXPR, index, minidx, 0);
1384 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1385 NULL_TREE, NULL_TREE);
1387 offset = offset % el_size;
1388 type = TREE_TYPE (type);
1403 /* Construct an expression that would reference a part of aggregate *EXPR of
1404 type TYPE at the given OFFSET of the type EXP_TYPE. If EXPR is NULL, the
1405 function only determines whether it can build such a reference without
1406 actually doing it, otherwise, the tree it points to is unshared first and
1407 then used as a base for furhter sub-references.
1409 FIXME: Eventually this should be replaced with
1410 maybe_fold_offset_to_reference() from tree-ssa-ccp.c but that requires a
1411 minor rewrite of fold_stmt.
1415 build_ref_for_offset (tree *expr, tree type, HOST_WIDE_INT offset,
1416 tree exp_type, bool allow_ptr)
1418 location_t loc = expr ? EXPR_LOCATION (*expr) : UNKNOWN_LOCATION;
1421 *expr = unshare_expr (*expr);
1423 if (allow_ptr && POINTER_TYPE_P (type))
1425 type = TREE_TYPE (type);
1427 *expr = fold_build1_loc (loc, INDIRECT_REF, type, *expr);
1430 return build_ref_for_offset_1 (expr, type, offset, exp_type);
1433 /* Return true iff TYPE is stdarg va_list type. */
1436 is_va_list_type (tree type)
1438 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1441 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1442 those with type which is suitable for scalarization. */
1445 find_var_candidates (void)
1448 referenced_var_iterator rvi;
1451 FOR_EACH_REFERENCED_VAR (var, rvi)
1453 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1455 type = TREE_TYPE (var);
1457 if (!AGGREGATE_TYPE_P (type)
1458 || needs_to_live_in_memory (var)
1459 || TREE_THIS_VOLATILE (var)
1460 || !COMPLETE_TYPE_P (type)
1461 || !host_integerp (TYPE_SIZE (type), 1)
1462 || tree_low_cst (TYPE_SIZE (type), 1) == 0
1463 || type_internals_preclude_sra_p (type)
1464 /* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1465 we also want to schedule it rather late. Thus we ignore it in
1467 || (sra_mode == SRA_MODE_EARLY_INTRA
1468 && is_va_list_type (type)))
1471 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1473 if (dump_file && (dump_flags & TDF_DETAILS))
1475 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1476 print_generic_expr (dump_file, var, 0);
1477 fprintf (dump_file, "\n");
1485 /* Sort all accesses for the given variable, check for partial overlaps and
1486 return NULL if there are any. If there are none, pick a representative for
1487 each combination of offset and size and create a linked list out of them.
1488 Return the pointer to the first representative and make sure it is the first
1489 one in the vector of accesses. */
1491 static struct access *
1492 sort_and_splice_var_accesses (tree var)
1494 int i, j, access_count;
1495 struct access *res, **prev_acc_ptr = &res;
1496 VEC (access_p, heap) *access_vec;
1498 HOST_WIDE_INT low = -1, high = 0;
1500 access_vec = get_base_access_vector (var);
1503 access_count = VEC_length (access_p, access_vec);
1505 /* Sort by <OFFSET, SIZE>. */
1506 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
1507 compare_access_positions);
1510 while (i < access_count)
1512 struct access *access = VEC_index (access_p, access_vec, i);
1513 bool grp_write = access->write;
1514 bool grp_read = !access->write;
1515 bool grp_assignment_read = access->grp_assignment_read;
1516 bool multiple_reads = false;
1517 bool total_scalarization = access->total_scalarization;
1518 bool grp_partial_lhs = access->grp_partial_lhs;
1519 bool first_scalar = is_gimple_reg_type (access->type);
1520 bool unscalarizable_region = access->grp_unscalarizable_region;
1522 if (first || access->offset >= high)
1525 low = access->offset;
1526 high = access->offset + access->size;
1528 else if (access->offset > low && access->offset + access->size > high)
1531 gcc_assert (access->offset >= low
1532 && access->offset + access->size <= high);
1535 while (j < access_count)
1537 struct access *ac2 = VEC_index (access_p, access_vec, j);
1538 if (ac2->offset != access->offset || ac2->size != access->size)
1545 multiple_reads = true;
1549 grp_assignment_read |= ac2->grp_assignment_read;
1550 grp_partial_lhs |= ac2->grp_partial_lhs;
1551 unscalarizable_region |= ac2->grp_unscalarizable_region;
1552 total_scalarization |= ac2->total_scalarization;
1553 relink_to_new_repr (access, ac2);
1555 /* If there are both aggregate-type and scalar-type accesses with
1556 this combination of size and offset, the comparison function
1557 should have put the scalars first. */
1558 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1559 ac2->group_representative = access;
1565 access->group_representative = access;
1566 access->grp_write = grp_write;
1567 access->grp_read = grp_read;
1568 access->grp_assignment_read = grp_assignment_read;
1569 access->grp_hint = multiple_reads || total_scalarization;
1570 access->grp_partial_lhs = grp_partial_lhs;
1571 access->grp_unscalarizable_region = unscalarizable_region;
1572 if (access->first_link)
1573 add_access_to_work_queue (access);
1575 *prev_acc_ptr = access;
1576 prev_acc_ptr = &access->next_grp;
1579 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1583 /* Create a variable for the given ACCESS which determines the type, name and a
1584 few other properties. Return the variable declaration and store it also to
1585 ACCESS->replacement. */
1588 create_access_replacement (struct access *access, bool rename)
1592 repl = create_tmp_var (access->type, "SR");
1594 add_referenced_var (repl);
1596 mark_sym_for_renaming (repl);
1598 if (!access->grp_partial_lhs
1599 && (TREE_CODE (access->type) == COMPLEX_TYPE
1600 || TREE_CODE (access->type) == VECTOR_TYPE))
1601 DECL_GIMPLE_REG_P (repl) = 1;
1603 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1604 DECL_ARTIFICIAL (repl) = 1;
1605 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1607 if (DECL_NAME (access->base)
1608 && !DECL_IGNORED_P (access->base)
1609 && !DECL_ARTIFICIAL (access->base))
1611 char *pretty_name = make_fancy_name (access->expr);
1612 tree debug_expr = unshare_expr (access->expr), d;
1614 DECL_NAME (repl) = get_identifier (pretty_name);
1615 obstack_free (&name_obstack, pretty_name);
1617 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1618 as DECL_DEBUG_EXPR isn't considered when looking for still
1619 used SSA_NAMEs and thus they could be freed. All debug info
1620 generation cares is whether something is constant or variable
1621 and that get_ref_base_and_extent works properly on the
1623 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1624 switch (TREE_CODE (d))
1627 case ARRAY_RANGE_REF:
1628 if (TREE_OPERAND (d, 1)
1629 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1630 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1631 if (TREE_OPERAND (d, 3)
1632 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1633 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1636 if (TREE_OPERAND (d, 2)
1637 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1638 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1643 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1644 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1645 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1648 TREE_NO_WARNING (repl) = 1;
1652 fprintf (dump_file, "Created a replacement for ");
1653 print_generic_expr (dump_file, access->base, 0);
1654 fprintf (dump_file, " offset: %u, size: %u: ",
1655 (unsigned) access->offset, (unsigned) access->size);
1656 print_generic_expr (dump_file, repl, 0);
1657 fprintf (dump_file, "\n");
1659 sra_stats.replacements++;
1664 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1667 get_access_replacement (struct access *access)
1669 gcc_assert (access->grp_to_be_replaced);
1671 if (!access->replacement_decl)
1672 access->replacement_decl = create_access_replacement (access, true);
1673 return access->replacement_decl;
1676 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1677 not mark it for renaming. */
1680 get_unrenamed_access_replacement (struct access *access)
1682 gcc_assert (!access->grp_to_be_replaced);
1684 if (!access->replacement_decl)
1685 access->replacement_decl = create_access_replacement (access, false);
1686 return access->replacement_decl;
1690 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1691 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1692 to it is not "within" the root. */
1695 build_access_subtree (struct access **access)
1697 struct access *root = *access, *last_child = NULL;
1698 HOST_WIDE_INT limit = root->offset + root->size;
1700 *access = (*access)->next_grp;
1701 while (*access && (*access)->offset + (*access)->size <= limit)
1704 root->first_child = *access;
1706 last_child->next_sibling = *access;
1707 last_child = *access;
1709 build_access_subtree (access);
1713 /* Build a tree of access representatives, ACCESS is the pointer to the first
1714 one, others are linked in a list by the next_grp field. Decide about scalar
1715 replacements on the way, return true iff any are to be created. */
1718 build_access_trees (struct access *access)
1722 struct access *root = access;
1724 build_access_subtree (&access);
1725 root->next_grp = access;
1729 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1733 expr_with_var_bounded_array_refs_p (tree expr)
1735 while (handled_component_p (expr))
1737 if (TREE_CODE (expr) == ARRAY_REF
1738 && !host_integerp (array_ref_low_bound (expr), 0))
1740 expr = TREE_OPERAND (expr, 0);
1745 enum mark_read_status { SRA_MR_NOT_READ, SRA_MR_READ, SRA_MR_ASSIGN_READ};
1747 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1748 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1749 sorts of access flags appropriately along the way, notably always set
1750 grp_read and grp_assign_read according to MARK_READ and grp_write when
1751 MARK_WRITE is true. */
1754 analyze_access_subtree (struct access *root, bool allow_replacements,
1755 enum mark_read_status mark_read, bool mark_write)
1757 struct access *child;
1758 HOST_WIDE_INT limit = root->offset + root->size;
1759 HOST_WIDE_INT covered_to = root->offset;
1760 bool scalar = is_gimple_reg_type (root->type);
1761 bool hole = false, sth_created = false;
1762 bool direct_read = root->grp_read;
1764 if (mark_read == SRA_MR_ASSIGN_READ)
1767 root->grp_assignment_read = 1;
1769 if (mark_read == SRA_MR_READ)
1771 else if (root->grp_assignment_read)
1772 mark_read = SRA_MR_ASSIGN_READ;
1773 else if (root->grp_read)
1774 mark_read = SRA_MR_READ;
1777 root->grp_write = true;
1778 else if (root->grp_write)
1781 if (root->grp_unscalarizable_region)
1782 allow_replacements = false;
1784 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
1785 allow_replacements = false;
1787 for (child = root->first_child; child; child = child->next_sibling)
1789 if (!hole && child->offset < covered_to)
1792 covered_to += child->size;
1794 sth_created |= analyze_access_subtree (child,
1795 allow_replacements && !scalar,
1796 mark_read, mark_write);
1798 root->grp_unscalarized_data |= child->grp_unscalarized_data;
1799 hole |= !child->grp_covered;
1802 if (allow_replacements && scalar && !root->first_child
1804 || (root->grp_write && (direct_read || root->grp_assignment_read)))
1805 /* We must not ICE later on when trying to build an access to the
1806 original data within the aggregate even when it is impossible to do in
1807 a defined way like in the PR 42703 testcase. Therefore we check
1808 pre-emptively here that we will be able to do that. */
1809 && build_ref_for_offset (NULL, TREE_TYPE (root->base), root->offset,
1812 if (dump_file && (dump_flags & TDF_DETAILS))
1814 fprintf (dump_file, "Marking ");
1815 print_generic_expr (dump_file, root->base, 0);
1816 fprintf (dump_file, " offset: %u, size: %u: ",
1817 (unsigned) root->offset, (unsigned) root->size);
1818 fprintf (dump_file, " to be replaced.\n");
1821 root->grp_to_be_replaced = 1;
1825 else if (covered_to < limit)
1828 if (sth_created && !hole)
1830 root->grp_covered = 1;
1833 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
1834 root->grp_unscalarized_data = 1; /* not covered and written to */
1840 /* Analyze all access trees linked by next_grp by the means of
1841 analyze_access_subtree. */
1843 analyze_access_trees (struct access *access)
1849 if (analyze_access_subtree (access, true, SRA_MR_NOT_READ, false))
1851 access = access->next_grp;
1857 /* Return true iff a potential new child of LACC at offset OFFSET and with size
1858 SIZE would conflict with an already existing one. If exactly such a child
1859 already exists in LACC, store a pointer to it in EXACT_MATCH. */
1862 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
1863 HOST_WIDE_INT size, struct access **exact_match)
1865 struct access *child;
1867 for (child = lacc->first_child; child; child = child->next_sibling)
1869 if (child->offset == norm_offset && child->size == size)
1871 *exact_match = child;
1875 if (child->offset < norm_offset + size
1876 && child->offset + child->size > norm_offset)
1883 /* Create a new child access of PARENT, with all properties just like MODEL
1884 except for its offset and with its grp_write false and grp_read true.
1885 Return the new access or NULL if it cannot be created. Note that this access
1886 is created long after all splicing and sorting, it's not located in any
1887 access vector and is automatically a representative of its group. */
1889 static struct access *
1890 create_artificial_child_access (struct access *parent, struct access *model,
1891 HOST_WIDE_INT new_offset)
1893 struct access *access;
1894 struct access **child;
1895 tree expr = parent->base;;
1897 gcc_assert (!model->grp_unscalarizable_region);
1899 if (!build_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
1900 model->type, false))
1903 access = (struct access *) pool_alloc (access_pool);
1904 memset (access, 0, sizeof (struct access));
1905 access->base = parent->base;
1906 access->expr = expr;
1907 access->offset = new_offset;
1908 access->size = model->size;
1909 access->type = model->type;
1910 access->grp_write = true;
1911 access->grp_read = false;
1913 child = &parent->first_child;
1914 while (*child && (*child)->offset < new_offset)
1915 child = &(*child)->next_sibling;
1917 access->next_sibling = *child;
1924 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
1925 true if any new subaccess was created. Additionally, if RACC is a scalar
1926 access but LACC is not, change the type of the latter, if possible. */
1929 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
1931 struct access *rchild;
1932 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
1935 if (is_gimple_reg_type (lacc->type)
1936 || lacc->grp_unscalarizable_region
1937 || racc->grp_unscalarizable_region)
1940 if (!lacc->first_child && !racc->first_child
1941 && is_gimple_reg_type (racc->type))
1943 tree t = lacc->base;
1945 if (build_ref_for_offset (&t, TREE_TYPE (t), lacc->offset, racc->type,
1949 lacc->type = racc->type;
1954 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
1956 struct access *new_acc = NULL;
1957 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
1959 if (rchild->grp_unscalarizable_region)
1962 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
1967 rchild->grp_hint = 1;
1968 new_acc->grp_hint |= new_acc->grp_read;
1969 if (rchild->first_child)
1970 ret |= propagate_subaccesses_across_link (new_acc, rchild);
1975 /* If a (part of) a union field is on the RHS of an assignment, it can
1976 have sub-accesses which do not make sense on the LHS (PR 40351).
1977 Check that this is not the case. */
1978 if (!build_ref_for_offset (NULL, TREE_TYPE (lacc->base), norm_offset,
1979 rchild->type, false))
1982 rchild->grp_hint = 1;
1983 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
1987 if (racc->first_child)
1988 propagate_subaccesses_across_link (new_acc, rchild);
1995 /* Propagate all subaccesses across assignment links. */
1998 propagate_all_subaccesses (void)
2000 while (work_queue_head)
2002 struct access *racc = pop_access_from_work_queue ();
2003 struct assign_link *link;
2005 gcc_assert (racc->first_link);
2007 for (link = racc->first_link; link; link = link->next)
2009 struct access *lacc = link->lacc;
2011 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2013 lacc = lacc->group_representative;
2014 if (propagate_subaccesses_across_link (lacc, racc)
2015 && lacc->first_link)
2016 add_access_to_work_queue (lacc);
2021 /* Go through all accesses collected throughout the (intraprocedural) analysis
2022 stage, exclude overlapping ones, identify representatives and build trees
2023 out of them, making decisions about scalarization on the way. Return true
2024 iff there are any to-be-scalarized variables after this stage. */
2027 analyze_all_variable_accesses (void)
2030 bitmap tmp = BITMAP_ALLOC (NULL);
2032 unsigned i, max_total_scalarization_size;
2034 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2035 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2037 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2038 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2039 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2041 tree var = referenced_var (i);
2043 if (TREE_CODE (var) == VAR_DECL
2044 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2045 <= max_total_scalarization_size)
2046 && type_consists_of_records_p (TREE_TYPE (var)))
2048 completely_scalarize_record (var, var, 0);
2049 if (dump_file && (dump_flags & TDF_DETAILS))
2051 fprintf (dump_file, "Will attempt to totally scalarize ");
2052 print_generic_expr (dump_file, var, 0);
2053 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2058 bitmap_copy (tmp, candidate_bitmap);
2059 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2061 tree var = referenced_var (i);
2062 struct access *access;
2064 access = sort_and_splice_var_accesses (var);
2066 build_access_trees (access);
2068 disqualify_candidate (var,
2069 "No or inhibitingly overlapping accesses.");
2072 propagate_all_subaccesses ();
2074 bitmap_copy (tmp, candidate_bitmap);
2075 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2077 tree var = referenced_var (i);
2078 struct access *access = get_first_repr_for_decl (var);
2080 if (analyze_access_trees (access))
2083 if (dump_file && (dump_flags & TDF_DETAILS))
2085 fprintf (dump_file, "\nAccess trees for ");
2086 print_generic_expr (dump_file, var, 0);
2087 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2088 dump_access_tree (dump_file, access);
2089 fprintf (dump_file, "\n");
2093 disqualify_candidate (var, "No scalar replacements to be created.");
2100 statistics_counter_event (cfun, "Scalarized aggregates", res);
2107 /* Return true iff a reference statement into aggregate AGG can be built for
2108 every single to-be-replaced accesses that is a child of ACCESS, its sibling
2109 or a child of its sibling. TOP_OFFSET is the offset from the processed
2110 access subtree that has to be subtracted from offset of each access. */
2113 ref_expr_for_all_replacements_p (struct access *access, tree agg,
2114 HOST_WIDE_INT top_offset)
2118 if (access->grp_to_be_replaced
2119 && !build_ref_for_offset (NULL, TREE_TYPE (agg),
2120 access->offset - top_offset,
2121 access->type, false))
2124 if (access->first_child
2125 && !ref_expr_for_all_replacements_p (access->first_child, agg,
2129 access = access->next_sibling;
2136 /* Generate statements copying scalar replacements of accesses within a subtree
2137 into or out of AGG. ACCESS is the first child of the root of the subtree to
2138 be processed. AGG is an aggregate type expression (can be a declaration but
2139 does not have to be, it can for example also be an indirect_ref).
2140 TOP_OFFSET is the offset of the processed subtree which has to be subtracted
2141 from offsets of individual accesses to get corresponding offsets for AGG.
2142 If CHUNK_SIZE is non-null, copy only replacements in the interval
2143 <start_offset, start_offset + chunk_size>, otherwise copy all. GSI is a
2144 statement iterator used to place the new statements. WRITE should be true
2145 when the statements should write from AGG to the replacement and false if
2146 vice versa. if INSERT_AFTER is true, new statements will be added after the
2147 current statement in GSI, they will be added before the statement
2151 generate_subtree_copies (struct access *access, tree agg,
2152 HOST_WIDE_INT top_offset,
2153 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2154 gimple_stmt_iterator *gsi, bool write,
2161 if (chunk_size && access->offset >= start_offset + chunk_size)
2164 if (access->grp_to_be_replaced
2166 || access->offset + access->size > start_offset))
2168 tree repl = get_access_replacement (access);
2172 ref_found = build_ref_for_offset (&expr, TREE_TYPE (agg),
2173 access->offset - top_offset,
2174 access->type, false);
2175 gcc_assert (ref_found);
2179 if (access->grp_partial_lhs)
2180 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2182 insert_after ? GSI_NEW_STMT
2184 stmt = gimple_build_assign (repl, expr);
2188 TREE_NO_WARNING (repl) = 1;
2189 if (access->grp_partial_lhs)
2190 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2192 insert_after ? GSI_NEW_STMT
2194 stmt = gimple_build_assign (expr, repl);
2198 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2200 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2202 sra_stats.subtree_copies++;
2205 if (access->first_child)
2206 generate_subtree_copies (access->first_child, agg, top_offset,
2207 start_offset, chunk_size, gsi,
2208 write, insert_after);
2210 access = access->next_sibling;
2215 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2216 the root of the subtree to be processed. GSI is the statement iterator used
2217 for inserting statements which are added after the current statement if
2218 INSERT_AFTER is true or before it otherwise. */
2221 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2225 struct access *child;
2227 if (access->grp_to_be_replaced)
2231 stmt = gimple_build_assign (get_access_replacement (access),
2232 fold_convert (access->type,
2233 integer_zero_node));
2235 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2237 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2241 for (child = access->first_child; child; child = child->next_sibling)
2242 init_subtree_with_zero (child, gsi, insert_after);
2245 /* Search for an access representative for the given expression EXPR and
2246 return it or NULL if it cannot be found. */
2248 static struct access *
2249 get_access_for_expr (tree expr)
2251 HOST_WIDE_INT offset, size, max_size;
2254 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2255 a different size than the size of its argument and we need the latter
2257 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2258 expr = TREE_OPERAND (expr, 0);
2260 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2261 if (max_size == -1 || !DECL_P (base))
2264 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2267 return get_var_base_offset_size_access (base, offset, max_size);
2270 /* Replace the expression EXPR with a scalar replacement if there is one and
2271 generate other statements to do type conversion or subtree copying if
2272 necessary. GSI is used to place newly created statements, WRITE is true if
2273 the expression is being written to (it is on a LHS of a statement or output
2274 in an assembly statement). */
2277 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2279 struct access *access;
2282 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2285 expr = &TREE_OPERAND (*expr, 0);
2290 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2291 expr = &TREE_OPERAND (*expr, 0);
2292 access = get_access_for_expr (*expr);
2295 type = TREE_TYPE (*expr);
2297 if (access->grp_to_be_replaced)
2299 tree repl = get_access_replacement (access);
2300 /* If we replace a non-register typed access simply use the original
2301 access expression to extract the scalar component afterwards.
2302 This happens if scalarizing a function return value or parameter
2303 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2304 gcc.c-torture/compile/20011217-1.c.
2306 We also want to use this when accessing a complex or vector which can
2307 be accessed as a different type too, potentially creating a need for
2308 type conversion (see PR42196) and when scalarized unions are involved
2309 in assembler statements (see PR42398). */
2310 if (!useless_type_conversion_p (type, access->type))
2312 tree ref = access->base;
2315 ok = build_ref_for_offset (&ref, TREE_TYPE (ref),
2316 access->offset, access->type, false);
2323 if (access->grp_partial_lhs)
2324 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2325 false, GSI_NEW_STMT);
2326 stmt = gimple_build_assign (repl, ref);
2327 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2333 if (access->grp_partial_lhs)
2334 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2335 true, GSI_SAME_STMT);
2336 stmt = gimple_build_assign (ref, repl);
2337 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2345 if (access->first_child)
2347 HOST_WIDE_INT start_offset, chunk_size;
2349 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2350 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2352 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2353 start_offset = access->offset
2354 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2357 start_offset = chunk_size = 0;
2359 generate_subtree_copies (access->first_child, access->base, 0,
2360 start_offset, chunk_size, gsi, write, write);
2365 /* Where scalar replacements of the RHS have been written to when a replacement
2366 of a LHS of an assigments cannot be direclty loaded from a replacement of
2368 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2369 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2370 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2372 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2373 base aggregate if there are unscalarized data or directly to LHS
2376 static enum unscalarized_data_handling
2377 handle_unscalarized_data_in_subtree (struct access *top_racc, tree lhs,
2378 gimple_stmt_iterator *gsi)
2380 if (top_racc->grp_unscalarized_data)
2382 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2384 return SRA_UDH_RIGHT;
2388 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2389 0, 0, gsi, false, false);
2390 return SRA_UDH_LEFT;
2395 /* Try to generate statements to load all sub-replacements in an access
2396 (sub)tree (LACC is the first child) from scalar replacements in the TOP_RACC
2397 (sub)tree. If that is not possible, refresh the TOP_RACC base aggregate and
2398 load the accesses from it. LEFT_OFFSET is the offset of the left whole
2399 subtree being copied, RIGHT_OFFSET is the same thing for the right subtree.
2400 GSI is stmt iterator used for statement insertions. *REFRESHED is true iff
2401 the rhs top aggregate has already been refreshed by contents of its scalar
2402 reductions and is set to true if this function has to do it. */
2405 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2406 HOST_WIDE_INT left_offset,
2407 HOST_WIDE_INT right_offset,
2408 gimple_stmt_iterator *old_gsi,
2409 gimple_stmt_iterator *new_gsi,
2410 enum unscalarized_data_handling *refreshed,
2413 location_t loc = EXPR_LOCATION (lacc->expr);
2416 if (lacc->grp_to_be_replaced)
2418 struct access *racc;
2419 HOST_WIDE_INT offset = lacc->offset - left_offset + right_offset;
2423 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2424 if (racc && racc->grp_to_be_replaced)
2426 rhs = get_access_replacement (racc);
2427 if (!useless_type_conversion_p (lacc->type, racc->type))
2428 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2432 /* No suitable access on the right hand side, need to load from
2433 the aggregate. See if we have to update it first... */
2434 if (*refreshed == SRA_UDH_NONE)
2435 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2438 if (*refreshed == SRA_UDH_LEFT)
2443 repl_found = build_ref_for_offset (&rhs, TREE_TYPE (rhs),
2444 lacc->offset, lacc->type,
2446 gcc_assert (repl_found);
2452 rhs = top_racc->base;
2453 repl_found = build_ref_for_offset (&rhs,
2454 TREE_TYPE (top_racc->base),
2455 offset, lacc->type, false);
2456 gcc_assert (repl_found);
2460 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2461 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2463 sra_stats.subreplacements++;
2465 else if (*refreshed == SRA_UDH_NONE
2466 && lacc->grp_read && !lacc->grp_covered)
2467 *refreshed = handle_unscalarized_data_in_subtree (top_racc, lhs,
2470 if (lacc->first_child)
2471 load_assign_lhs_subreplacements (lacc->first_child, top_racc,
2472 left_offset, right_offset,
2473 old_gsi, new_gsi, refreshed, lhs);
2474 lacc = lacc->next_sibling;
2479 /* Result code for SRA assignment modification. */
2480 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2481 SRA_AM_MODIFIED, /* stmt changed but not
2483 SRA_AM_REMOVED }; /* stmt eliminated */
2485 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2486 to the assignment and GSI is the statement iterator pointing at it. Returns
2487 the same values as sra_modify_assign. */
2489 static enum assignment_mod_result
2490 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2492 tree lhs = gimple_assign_lhs (*stmt);
2495 acc = get_access_for_expr (lhs);
2499 if (VEC_length (constructor_elt,
2500 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2502 /* I have never seen this code path trigger but if it can happen the
2503 following should handle it gracefully. */
2504 if (access_has_children_p (acc))
2505 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2507 return SRA_AM_MODIFIED;
2510 if (acc->grp_covered)
2512 init_subtree_with_zero (acc, gsi, false);
2513 unlink_stmt_vdef (*stmt);
2514 gsi_remove (gsi, true);
2515 return SRA_AM_REMOVED;
2519 init_subtree_with_zero (acc, gsi, true);
2520 return SRA_AM_MODIFIED;
2524 /* Create a new suitable default definition SSA_NAME and replace all uses of
2525 SSA with it, RACC is access describing the uninitialized part of an
2526 aggregate that is being loaded. */
2529 replace_uses_with_default_def_ssa_name (tree ssa, struct access *racc)
2533 decl = get_unrenamed_access_replacement (racc);
2535 repl = gimple_default_def (cfun, decl);
2538 repl = make_ssa_name (decl, gimple_build_nop ());
2539 set_default_def (decl, repl);
2542 replace_uses_by (ssa, repl);
2545 /* Examine both sides of the assignment statement pointed to by STMT, replace
2546 them with a scalare replacement if there is one and generate copying of
2547 replacements if scalarized aggregates have been used in the assignment. GSI
2548 is used to hold generated statements for type conversions and subtree
2551 static enum assignment_mod_result
2552 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2554 struct access *lacc, *racc;
2556 bool modify_this_stmt = false;
2557 bool force_gimple_rhs = false;
2558 location_t loc = gimple_location (*stmt);
2559 gimple_stmt_iterator orig_gsi = *gsi;
2561 if (!gimple_assign_single_p (*stmt))
2563 lhs = gimple_assign_lhs (*stmt);
2564 rhs = gimple_assign_rhs1 (*stmt);
2566 if (TREE_CODE (rhs) == CONSTRUCTOR)
2567 return sra_modify_constructor_assign (stmt, gsi);
2569 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2570 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2571 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2573 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2575 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2577 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2580 lacc = get_access_for_expr (lhs);
2581 racc = get_access_for_expr (rhs);
2585 if (lacc && lacc->grp_to_be_replaced)
2587 lhs = get_access_replacement (lacc);
2588 gimple_assign_set_lhs (*stmt, lhs);
2589 modify_this_stmt = true;
2590 if (lacc->grp_partial_lhs)
2591 force_gimple_rhs = true;
2595 if (racc && racc->grp_to_be_replaced)
2597 rhs = get_access_replacement (racc);
2598 modify_this_stmt = true;
2599 if (racc->grp_partial_lhs)
2600 force_gimple_rhs = true;
2604 if (modify_this_stmt)
2606 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2608 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2609 ??? This should move to fold_stmt which we simply should
2610 call after building a VIEW_CONVERT_EXPR here. */
2611 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2612 && !access_has_children_p (lacc))
2615 if (build_ref_for_offset (&expr, TREE_TYPE (lhs), 0,
2616 TREE_TYPE (rhs), false))
2619 gimple_assign_set_lhs (*stmt, expr);
2622 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2623 && !access_has_children_p (racc))
2626 if (build_ref_for_offset (&expr, TREE_TYPE (rhs), 0,
2627 TREE_TYPE (lhs), false))
2630 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2632 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs), rhs);
2633 if (is_gimple_reg_type (TREE_TYPE (lhs))
2634 && TREE_CODE (lhs) != SSA_NAME)
2635 force_gimple_rhs = true;
2640 /* From this point on, the function deals with assignments in between
2641 aggregates when at least one has scalar reductions of some of its
2642 components. There are three possible scenarios: Both the LHS and RHS have
2643 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2645 In the first case, we would like to load the LHS components from RHS
2646 components whenever possible. If that is not possible, we would like to
2647 read it directly from the RHS (after updating it by storing in it its own
2648 components). If there are some necessary unscalarized data in the LHS,
2649 those will be loaded by the original assignment too. If neither of these
2650 cases happen, the original statement can be removed. Most of this is done
2651 by load_assign_lhs_subreplacements.
2653 In the second case, we would like to store all RHS scalarized components
2654 directly into LHS and if they cover the aggregate completely, remove the
2655 statement too. In the third case, we want the LHS components to be loaded
2656 directly from the RHS (DSE will remove the original statement if it
2659 This is a bit complex but manageable when types match and when unions do
2660 not cause confusion in a way that we cannot really load a component of LHS
2661 from the RHS or vice versa (the access representing this level can have
2662 subaccesses that are accessible only through a different union field at a
2663 higher level - different from the one used in the examined expression).
2666 Therefore, I specially handle a fourth case, happening when there is a
2667 specific type cast or it is impossible to locate a scalarized subaccess on
2668 the other side of the expression. If that happens, I simply "refresh" the
2669 RHS by storing in it is scalarized components leave the original statement
2670 there to do the copying and then load the scalar replacements of the LHS.
2671 This is what the first branch does. */
2673 if (gimple_has_volatile_ops (*stmt)
2674 || contains_view_convert_expr_p (rhs)
2675 || contains_view_convert_expr_p (lhs)
2676 || (access_has_children_p (racc)
2677 && !ref_expr_for_all_replacements_p (racc, lhs, racc->offset))
2678 || (access_has_children_p (lacc)
2679 && !ref_expr_for_all_replacements_p (lacc, rhs, lacc->offset)))
2681 if (access_has_children_p (racc))
2682 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2684 if (access_has_children_p (lacc))
2685 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2687 sra_stats.separate_lhs_rhs_handling++;
2691 if (access_has_children_p (lacc) && access_has_children_p (racc))
2693 gimple_stmt_iterator orig_gsi = *gsi;
2694 enum unscalarized_data_handling refreshed;
2696 if (lacc->grp_read && !lacc->grp_covered)
2697 refreshed = handle_unscalarized_data_in_subtree (racc, lhs, gsi);
2699 refreshed = SRA_UDH_NONE;
2701 load_assign_lhs_subreplacements (lacc->first_child, racc,
2702 lacc->offset, racc->offset,
2703 &orig_gsi, gsi, &refreshed, lhs);
2704 if (refreshed != SRA_UDH_RIGHT)
2706 if (*stmt == gsi_stmt (*gsi))
2709 unlink_stmt_vdef (*stmt);
2710 gsi_remove (&orig_gsi, true);
2711 sra_stats.deleted++;
2712 return SRA_AM_REMOVED;
2719 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2721 if (racc->first_child)
2722 generate_subtree_copies (racc->first_child, lhs,
2723 racc->offset, 0, 0, gsi,
2725 gcc_assert (*stmt == gsi_stmt (*gsi));
2726 if (TREE_CODE (lhs) == SSA_NAME)
2727 replace_uses_with_default_def_ssa_name (lhs, racc);
2729 unlink_stmt_vdef (*stmt);
2730 gsi_remove (gsi, true);
2731 sra_stats.deleted++;
2732 return SRA_AM_REMOVED;
2734 else if (racc->first_child)
2735 generate_subtree_copies (racc->first_child, lhs,
2736 racc->offset, 0, 0, gsi, false, true);
2738 if (access_has_children_p (lacc))
2739 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2740 0, 0, gsi, true, true);
2744 /* This gimplification must be done after generate_subtree_copies, lest we
2745 insert the subtree copies in the middle of the gimplified sequence. */
2746 if (force_gimple_rhs)
2747 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2748 true, GSI_SAME_STMT);
2749 if (gimple_assign_rhs1 (*stmt) != rhs)
2751 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2752 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2755 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2758 /* Traverse the function body and all modifications as decided in
2759 analyze_all_variable_accesses. */
2762 sra_modify_function_body (void)
2768 gimple_stmt_iterator gsi = gsi_start_bb (bb);
2769 while (!gsi_end_p (gsi))
2771 gimple stmt = gsi_stmt (gsi);
2772 enum assignment_mod_result assign_result;
2773 bool modified = false, deleted = false;
2777 switch (gimple_code (stmt))
2780 t = gimple_return_retval_ptr (stmt);
2781 if (*t != NULL_TREE)
2782 modified |= sra_modify_expr (t, &gsi, false);
2786 assign_result = sra_modify_assign (&stmt, &gsi);
2787 modified |= assign_result == SRA_AM_MODIFIED;
2788 deleted = assign_result == SRA_AM_REMOVED;
2792 /* Operands must be processed before the lhs. */
2793 for (i = 0; i < gimple_call_num_args (stmt); i++)
2795 t = gimple_call_arg_ptr (stmt, i);
2796 modified |= sra_modify_expr (t, &gsi, false);
2799 if (gimple_call_lhs (stmt))
2801 t = gimple_call_lhs_ptr (stmt);
2802 modified |= sra_modify_expr (t, &gsi, true);
2807 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
2809 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
2810 modified |= sra_modify_expr (t, &gsi, false);
2812 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
2814 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
2815 modified |= sra_modify_expr (t, &gsi, true);
2826 maybe_clean_eh_stmt (stmt);
2834 /* Generate statements initializing scalar replacements of parts of function
2838 initialize_parameter_reductions (void)
2840 gimple_stmt_iterator gsi;
2841 gimple_seq seq = NULL;
2844 for (parm = DECL_ARGUMENTS (current_function_decl);
2846 parm = TREE_CHAIN (parm))
2848 VEC (access_p, heap) *access_vec;
2849 struct access *access;
2851 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
2853 access_vec = get_base_access_vector (parm);
2859 seq = gimple_seq_alloc ();
2860 gsi = gsi_start (seq);
2863 for (access = VEC_index (access_p, access_vec, 0);
2865 access = access->next_grp)
2866 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true);
2870 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
2873 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
2874 it reveals there are components of some aggregates to be scalarized, it runs
2875 the required transformations. */
2877 perform_intra_sra (void)
2882 if (!find_var_candidates ())
2885 if (!scan_function ())
2888 if (!analyze_all_variable_accesses ())
2891 sra_modify_function_body ();
2892 initialize_parameter_reductions ();
2894 statistics_counter_event (cfun, "Scalar replacements created",
2895 sra_stats.replacements);
2896 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
2897 statistics_counter_event (cfun, "Subtree copy stmts",
2898 sra_stats.subtree_copies);
2899 statistics_counter_event (cfun, "Subreplacement stmts",
2900 sra_stats.subreplacements);
2901 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
2902 statistics_counter_event (cfun, "Separate LHS and RHS handling",
2903 sra_stats.separate_lhs_rhs_handling);
2905 ret = TODO_update_ssa;
2908 sra_deinitialize ();
2912 /* Perform early intraprocedural SRA. */
2914 early_intra_sra (void)
2916 sra_mode = SRA_MODE_EARLY_INTRA;
2917 return perform_intra_sra ();
2920 /* Perform "late" intraprocedural SRA. */
2922 late_intra_sra (void)
2924 sra_mode = SRA_MODE_INTRA;
2925 return perform_intra_sra ();
2930 gate_intra_sra (void)
2932 return flag_tree_sra != 0;
2936 struct gimple_opt_pass pass_sra_early =
2941 gate_intra_sra, /* gate */
2942 early_intra_sra, /* execute */
2945 0, /* static_pass_number */
2946 TV_TREE_SRA, /* tv_id */
2947 PROP_cfg | PROP_ssa, /* properties_required */
2948 0, /* properties_provided */
2949 0, /* properties_destroyed */
2950 0, /* todo_flags_start */
2954 | TODO_verify_ssa /* todo_flags_finish */
2958 struct gimple_opt_pass pass_sra =
2963 gate_intra_sra, /* gate */
2964 late_intra_sra, /* execute */
2967 0, /* static_pass_number */
2968 TV_TREE_SRA, /* tv_id */
2969 PROP_cfg | PROP_ssa, /* properties_required */
2970 0, /* properties_provided */
2971 0, /* properties_destroyed */
2972 TODO_update_address_taken, /* todo_flags_start */
2976 | TODO_verify_ssa /* todo_flags_finish */
2981 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
2985 is_unused_scalar_param (tree parm)
2988 return (is_gimple_reg (parm)
2989 && (!(name = gimple_default_def (cfun, parm))
2990 || has_zero_uses (name)));
2993 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
2994 examine whether there are any direct or otherwise infeasible ones. If so,
2995 return true, otherwise return false. PARM must be a gimple register with a
2996 non-NULL default definition. */
2999 ptr_parm_has_direct_uses (tree parm)
3001 imm_use_iterator ui;
3003 tree name = gimple_default_def (cfun, parm);
3006 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3009 use_operand_p use_p;
3011 if (is_gimple_debug (stmt))
3014 /* Valid uses include dereferences on the lhs and the rhs. */
3015 if (gimple_has_lhs (stmt))
3017 tree lhs = gimple_get_lhs (stmt);
3018 while (handled_component_p (lhs))
3019 lhs = TREE_OPERAND (lhs, 0);
3020 if (INDIRECT_REF_P (lhs)
3021 && TREE_OPERAND (lhs, 0) == name)
3024 if (gimple_assign_single_p (stmt))
3026 tree rhs = gimple_assign_rhs1 (stmt);
3027 while (handled_component_p (rhs))
3028 rhs = TREE_OPERAND (rhs, 0);
3029 if (INDIRECT_REF_P (rhs)
3030 && TREE_OPERAND (rhs, 0) == name)
3033 else if (is_gimple_call (stmt))
3036 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3038 tree arg = gimple_call_arg (stmt, i);
3039 while (handled_component_p (arg))
3040 arg = TREE_OPERAND (arg, 0);
3041 if (INDIRECT_REF_P (arg)
3042 && TREE_OPERAND (arg, 0) == name)
3047 /* If the number of valid uses does not match the number of
3048 uses in this stmt there is an unhandled use. */
3049 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3056 BREAK_FROM_IMM_USE_STMT (ui);
3062 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3063 them in candidate_bitmap. Note that these do not necessarily include
3064 parameter which are unused and thus can be removed. Return true iff any
3065 such candidate has been found. */
3068 find_param_candidates (void)
3074 for (parm = DECL_ARGUMENTS (current_function_decl);
3076 parm = TREE_CHAIN (parm))
3078 tree type = TREE_TYPE (parm);
3082 if (TREE_THIS_VOLATILE (parm)
3083 || TREE_ADDRESSABLE (parm)
3084 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3087 if (is_unused_scalar_param (parm))
3093 if (POINTER_TYPE_P (type))
3095 type = TREE_TYPE (type);
3097 if (TREE_CODE (type) == FUNCTION_TYPE
3098 || TYPE_VOLATILE (type)
3099 || !is_gimple_reg (parm)
3100 || is_va_list_type (type)
3101 || ptr_parm_has_direct_uses (parm))
3104 else if (!AGGREGATE_TYPE_P (type))
3107 if (!COMPLETE_TYPE_P (type)
3108 || !host_integerp (TYPE_SIZE (type), 1)
3109 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3110 || (AGGREGATE_TYPE_P (type)
3111 && type_internals_preclude_sra_p (type)))
3114 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3116 if (dump_file && (dump_flags & TDF_DETAILS))
3118 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3119 print_generic_expr (dump_file, parm, 0);
3120 fprintf (dump_file, "\n");
3124 func_param_count = count;
3128 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3132 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3135 struct access *repr = (struct access *) data;
3137 repr->grp_maybe_modified = 1;
3141 /* Analyze what representatives (in linked lists accessible from
3142 REPRESENTATIVES) can be modified by side effects of statements in the
3143 current function. */
3146 analyze_modified_params (VEC (access_p, heap) *representatives)
3150 for (i = 0; i < func_param_count; i++)
3152 struct access *repr;
3154 for (repr = VEC_index (access_p, representatives, i);
3156 repr = repr->next_grp)
3158 struct access *access;
3162 if (no_accesses_p (repr))
3164 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3165 || repr->grp_maybe_modified)
3168 ao_ref_init (&ar, repr->expr);
3169 visited = BITMAP_ALLOC (NULL);
3170 for (access = repr; access; access = access->next_sibling)
3172 /* All accesses are read ones, otherwise grp_maybe_modified would
3173 be trivially set. */
3174 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3175 mark_maybe_modified, repr, &visited);
3176 if (repr->grp_maybe_modified)
3179 BITMAP_FREE (visited);
3184 /* Propagate distances in bb_dereferences in the opposite direction than the
3185 control flow edges, in each step storing the maximum of the current value
3186 and the minimum of all successors. These steps are repeated until the table
3187 stabilizes. Note that BBs which might terminate the functions (according to
3188 final_bbs bitmap) never updated in this way. */
3191 propagate_dereference_distances (void)
3193 VEC (basic_block, heap) *queue;
3196 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3197 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3200 VEC_quick_push (basic_block, queue, bb);
3204 while (!VEC_empty (basic_block, queue))
3208 bool change = false;
3211 bb = VEC_pop (basic_block, queue);
3214 if (bitmap_bit_p (final_bbs, bb->index))
3217 for (i = 0; i < func_param_count; i++)
3219 int idx = bb->index * func_param_count + i;
3221 HOST_WIDE_INT inh = 0;
3223 FOR_EACH_EDGE (e, ei, bb->succs)
3225 int succ_idx = e->dest->index * func_param_count + i;
3227 if (e->src == EXIT_BLOCK_PTR)
3233 inh = bb_dereferences [succ_idx];
3235 else if (bb_dereferences [succ_idx] < inh)
3236 inh = bb_dereferences [succ_idx];
3239 if (!first && bb_dereferences[idx] < inh)
3241 bb_dereferences[idx] = inh;
3246 if (change && !bitmap_bit_p (final_bbs, bb->index))
3247 FOR_EACH_EDGE (e, ei, bb->preds)
3252 e->src->aux = e->src;
3253 VEC_quick_push (basic_block, queue, e->src);
3257 VEC_free (basic_block, heap, queue);
3260 /* Dump a dereferences TABLE with heading STR to file F. */
3263 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3267 fprintf (dump_file, str);
3268 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3270 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3271 if (bb != EXIT_BLOCK_PTR)
3274 for (i = 0; i < func_param_count; i++)
3276 int idx = bb->index * func_param_count + i;
3277 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3282 fprintf (dump_file, "\n");
3285 /* Determine what (parts of) parameters passed by reference that are not
3286 assigned to are not certainly dereferenced in this function and thus the
3287 dereferencing cannot be safely moved to the caller without potentially
3288 introducing a segfault. Mark such REPRESENTATIVES as
3289 grp_not_necessarilly_dereferenced.
3291 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3292 part is calculated rather than simple booleans are calculated for each
3293 pointer parameter to handle cases when only a fraction of the whole
3294 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3297 The maximum dereference distances for each pointer parameter and BB are
3298 already stored in bb_dereference. This routine simply propagates these
3299 values upwards by propagate_dereference_distances and then compares the
3300 distances of individual parameters in the ENTRY BB to the equivalent
3301 distances of each representative of a (fraction of a) parameter. */
3304 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3308 if (dump_file && (dump_flags & TDF_DETAILS))
3309 dump_dereferences_table (dump_file,
3310 "Dereference table before propagation:\n",
3313 propagate_dereference_distances ();
3315 if (dump_file && (dump_flags & TDF_DETAILS))
3316 dump_dereferences_table (dump_file,
3317 "Dereference table after propagation:\n",
3320 for (i = 0; i < func_param_count; i++)
3322 struct access *repr = VEC_index (access_p, representatives, i);
3323 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3325 if (!repr || no_accesses_p (repr))
3330 if ((repr->offset + repr->size) > bb_dereferences[idx])
3331 repr->grp_not_necessarilly_dereferenced = 1;
3332 repr = repr->next_grp;
3338 /* Return the representative access for the parameter declaration PARM if it is
3339 a scalar passed by reference which is not written to and the pointer value
3340 is not used directly. Thus, if it is legal to dereference it in the caller
3341 and we can rule out modifications through aliases, such parameter should be
3342 turned into one passed by value. Return NULL otherwise. */
3344 static struct access *
3345 unmodified_by_ref_scalar_representative (tree parm)
3347 int i, access_count;
3348 struct access *repr;
3349 VEC (access_p, heap) *access_vec;
3351 access_vec = get_base_access_vector (parm);
3352 gcc_assert (access_vec);
3353 repr = VEC_index (access_p, access_vec, 0);
3356 repr->group_representative = repr;
3358 access_count = VEC_length (access_p, access_vec);
3359 for (i = 1; i < access_count; i++)
3361 struct access *access = VEC_index (access_p, access_vec, i);
3364 access->group_representative = repr;
3365 access->next_sibling = repr->next_sibling;
3366 repr->next_sibling = access;
3370 repr->grp_scalar_ptr = 1;
3374 /* Return true iff this access precludes IPA-SRA of the parameter it is
3378 access_precludes_ipa_sra_p (struct access *access)
3380 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3381 is incompatible assign in a call statement (and possibly even in asm
3382 statements). This can be relaxed by using a new temporary but only for
3383 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3384 intraprocedural SRA we deal with this by keeping the old aggregate around,
3385 something we cannot do in IPA-SRA.) */
3387 && (is_gimple_call (access->stmt)
3388 || gimple_code (access->stmt) == GIMPLE_ASM))
3395 /* Sort collected accesses for parameter PARM, identify representatives for
3396 each accessed region and link them together. Return NULL if there are
3397 different but overlapping accesses, return the special ptr value meaning
3398 there are no accesses for this parameter if that is the case and return the
3399 first representative otherwise. Set *RO_GRP if there is a group of accesses
3400 with only read (i.e. no write) accesses. */
3402 static struct access *
3403 splice_param_accesses (tree parm, bool *ro_grp)
3405 int i, j, access_count, group_count;
3406 int agg_size, total_size = 0;
3407 struct access *access, *res, **prev_acc_ptr = &res;
3408 VEC (access_p, heap) *access_vec;
3410 access_vec = get_base_access_vector (parm);
3412 return &no_accesses_representant;
3413 access_count = VEC_length (access_p, access_vec);
3415 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
3416 compare_access_positions);
3421 while (i < access_count)
3424 access = VEC_index (access_p, access_vec, i);
3425 modification = access->write;
3426 if (access_precludes_ipa_sra_p (access))
3429 /* Access is about to become group representative unless we find some
3430 nasty overlap which would preclude us from breaking this parameter
3434 while (j < access_count)
3436 struct access *ac2 = VEC_index (access_p, access_vec, j);
3437 if (ac2->offset != access->offset)
3439 /* All or nothing law for parameters. */
3440 if (access->offset + access->size > ac2->offset)
3445 else if (ac2->size != access->size)
3448 if (access_precludes_ipa_sra_p (ac2))
3451 modification |= ac2->write;
3452 ac2->group_representative = access;
3453 ac2->next_sibling = access->next_sibling;
3454 access->next_sibling = ac2;
3459 access->grp_maybe_modified = modification;
3462 *prev_acc_ptr = access;
3463 prev_acc_ptr = &access->next_grp;
3464 total_size += access->size;
3468 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3469 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3471 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3472 if (total_size >= agg_size)
3475 gcc_assert (group_count > 0);
3479 /* Decide whether parameters with representative accesses given by REPR should
3480 be reduced into components. */
3483 decide_one_param_reduction (struct access *repr)
3485 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3490 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3491 gcc_assert (cur_parm_size > 0);
3493 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3496 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3501 agg_size = cur_parm_size;
3507 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3508 print_generic_expr (dump_file, parm, 0);
3509 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3510 for (acc = repr; acc; acc = acc->next_grp)
3511 dump_access (dump_file, acc, true);
3515 new_param_count = 0;
3517 for (; repr; repr = repr->next_grp)
3519 gcc_assert (parm == repr->base);
3522 if (!by_ref || (!repr->grp_maybe_modified
3523 && !repr->grp_not_necessarilly_dereferenced))
3524 total_size += repr->size;
3526 total_size += cur_parm_size;
3529 gcc_assert (new_param_count > 0);
3531 if (optimize_function_for_size_p (cfun))
3532 parm_size_limit = cur_parm_size;
3534 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3537 if (total_size < agg_size
3538 && total_size <= parm_size_limit)
3541 fprintf (dump_file, " ....will be split into %i components\n",
3543 return new_param_count;
3549 /* The order of the following enums is important, we need to do extra work for
3550 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3551 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3552 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3554 /* Identify representatives of all accesses to all candidate parameters for
3555 IPA-SRA. Return result based on what representatives have been found. */
3557 static enum ipa_splicing_result
3558 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3560 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3562 struct access *repr;
3564 *representatives = VEC_alloc (access_p, heap, func_param_count);
3566 for (parm = DECL_ARGUMENTS (current_function_decl);
3568 parm = TREE_CHAIN (parm))
3570 if (is_unused_scalar_param (parm))
3572 VEC_quick_push (access_p, *representatives,
3573 &no_accesses_representant);
3574 if (result == NO_GOOD_ACCESS)
3575 result = UNUSED_PARAMS;
3577 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3578 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3579 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3581 repr = unmodified_by_ref_scalar_representative (parm);
3582 VEC_quick_push (access_p, *representatives, repr);
3584 result = UNMODIF_BY_REF_ACCESSES;
3586 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3588 bool ro_grp = false;
3589 repr = splice_param_accesses (parm, &ro_grp);
3590 VEC_quick_push (access_p, *representatives, repr);
3592 if (repr && !no_accesses_p (repr))
3594 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3597 result = UNMODIF_BY_REF_ACCESSES;
3598 else if (result < MODIF_BY_REF_ACCESSES)
3599 result = MODIF_BY_REF_ACCESSES;
3601 else if (result < BY_VAL_ACCESSES)
3602 result = BY_VAL_ACCESSES;
3604 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3605 result = UNUSED_PARAMS;
3608 VEC_quick_push (access_p, *representatives, NULL);
3611 if (result == NO_GOOD_ACCESS)
3613 VEC_free (access_p, heap, *representatives);
3614 *representatives = NULL;
3615 return NO_GOOD_ACCESS;
3621 /* Return the index of BASE in PARMS. Abort if it is not found. */
3624 get_param_index (tree base, VEC(tree, heap) *parms)
3628 len = VEC_length (tree, parms);
3629 for (i = 0; i < len; i++)
3630 if (VEC_index (tree, parms, i) == base)
3635 /* Convert the decisions made at the representative level into compact
3636 parameter adjustments. REPRESENTATIVES are pointers to first
3637 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3638 final number of adjustments. */
3640 static ipa_parm_adjustment_vec
3641 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3642 int adjustments_count)
3644 VEC (tree, heap) *parms;
3645 ipa_parm_adjustment_vec adjustments;
3649 gcc_assert (adjustments_count > 0);
3650 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3651 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3652 parm = DECL_ARGUMENTS (current_function_decl);
3653 for (i = 0; i < func_param_count; i++, parm = TREE_CHAIN (parm))
3655 struct access *repr = VEC_index (access_p, representatives, i);
3657 if (!repr || no_accesses_p (repr))
3659 struct ipa_parm_adjustment *adj;
3661 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3662 memset (adj, 0, sizeof (*adj));
3663 adj->base_index = get_param_index (parm, parms);
3666 adj->copy_param = 1;
3668 adj->remove_param = 1;
3672 struct ipa_parm_adjustment *adj;
3673 int index = get_param_index (parm, parms);
3675 for (; repr; repr = repr->next_grp)
3677 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3678 memset (adj, 0, sizeof (*adj));
3679 gcc_assert (repr->base == parm);
3680 adj->base_index = index;
3681 adj->base = repr->base;
3682 adj->type = repr->type;
3683 adj->offset = repr->offset;
3684 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3685 && (repr->grp_maybe_modified
3686 || repr->grp_not_necessarilly_dereferenced));
3691 VEC_free (tree, heap, parms);
3695 /* Analyze the collected accesses and produce a plan what to do with the
3696 parameters in the form of adjustments, NULL meaning nothing. */
3698 static ipa_parm_adjustment_vec
3699 analyze_all_param_acesses (void)
3701 enum ipa_splicing_result repr_state;
3702 bool proceed = false;
3703 int i, adjustments_count = 0;
3704 VEC (access_p, heap) *representatives;
3705 ipa_parm_adjustment_vec adjustments;
3707 repr_state = splice_all_param_accesses (&representatives);
3708 if (repr_state == NO_GOOD_ACCESS)
3711 /* If there are any parameters passed by reference which are not modified
3712 directly, we need to check whether they can be modified indirectly. */
3713 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3715 analyze_caller_dereference_legality (representatives);
3716 analyze_modified_params (representatives);
3719 for (i = 0; i < func_param_count; i++)
3721 struct access *repr = VEC_index (access_p, representatives, i);
3723 if (repr && !no_accesses_p (repr))
3725 if (repr->grp_scalar_ptr)
3727 adjustments_count++;
3728 if (repr->grp_not_necessarilly_dereferenced
3729 || repr->grp_maybe_modified)
3730 VEC_replace (access_p, representatives, i, NULL);
3734 sra_stats.scalar_by_ref_to_by_val++;
3739 int new_components = decide_one_param_reduction (repr);
3741 if (new_components == 0)
3743 VEC_replace (access_p, representatives, i, NULL);
3744 adjustments_count++;
3748 adjustments_count += new_components;
3749 sra_stats.aggregate_params_reduced++;
3750 sra_stats.param_reductions_created += new_components;
3757 if (no_accesses_p (repr))
3760 sra_stats.deleted_unused_parameters++;
3762 adjustments_count++;
3766 if (!proceed && dump_file)
3767 fprintf (dump_file, "NOT proceeding to change params.\n");
3770 adjustments = turn_representatives_into_adjustments (representatives,
3775 VEC_free (access_p, heap, representatives);
3779 /* If a parameter replacement identified by ADJ does not yet exist in the form
3780 of declaration, create it and record it, otherwise return the previously
3784 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
3787 if (!adj->new_ssa_base)
3789 char *pretty_name = make_fancy_name (adj->base);
3791 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
3792 DECL_NAME (repl) = get_identifier (pretty_name);
3793 obstack_free (&name_obstack, pretty_name);
3796 add_referenced_var (repl);
3797 adj->new_ssa_base = repl;
3800 repl = adj->new_ssa_base;
3804 /* Find the first adjustment for a particular parameter BASE in a vector of
3805 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3808 static struct ipa_parm_adjustment *
3809 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
3813 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3814 for (i = 0; i < len; i++)
3816 struct ipa_parm_adjustment *adj;
3818 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3819 if (!adj->copy_param && adj->base == base)
3826 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3827 removed because its value is not used, replace the SSA_NAME with a one
3828 relating to a created VAR_DECL together all of its uses and return true.
3829 ADJUSTMENTS is a pointer to an adjustments vector. */
3832 replace_removed_params_ssa_names (gimple stmt,
3833 ipa_parm_adjustment_vec adjustments)
3835 struct ipa_parm_adjustment *adj;
3836 tree lhs, decl, repl, name;
3838 if (gimple_code (stmt) == GIMPLE_PHI)
3839 lhs = gimple_phi_result (stmt);
3840 else if (is_gimple_assign (stmt))
3841 lhs = gimple_assign_lhs (stmt);
3842 else if (is_gimple_call (stmt))
3843 lhs = gimple_call_lhs (stmt);
3847 if (TREE_CODE (lhs) != SSA_NAME)
3849 decl = SSA_NAME_VAR (lhs);
3850 if (TREE_CODE (decl) != PARM_DECL)
3853 adj = get_adjustment_for_base (adjustments, decl);
3857 repl = get_replaced_param_substitute (adj);
3858 name = make_ssa_name (repl, stmt);
3862 fprintf (dump_file, "replacing an SSA name of a removed param ");
3863 print_generic_expr (dump_file, lhs, 0);
3864 fprintf (dump_file, " with ");
3865 print_generic_expr (dump_file, name, 0);
3866 fprintf (dump_file, "\n");
3869 if (is_gimple_assign (stmt))
3870 gimple_assign_set_lhs (stmt, name);
3871 else if (is_gimple_call (stmt))
3872 gimple_call_set_lhs (stmt, name);
3874 gimple_phi_set_result (stmt, name);
3876 replace_uses_by (lhs, name);
3880 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
3881 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
3882 specifies whether the function should care about type incompatibility the
3883 current and new expressions. If it is false, the function will leave
3884 incompatibility issues to the caller. Return true iff the expression
3888 sra_ipa_modify_expr (tree *expr, bool convert,
3889 ipa_parm_adjustment_vec adjustments)
3892 struct ipa_parm_adjustment *adj, *cand = NULL;
3893 HOST_WIDE_INT offset, size, max_size;
3896 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3898 if (TREE_CODE (*expr) == BIT_FIELD_REF
3899 || TREE_CODE (*expr) == IMAGPART_EXPR
3900 || TREE_CODE (*expr) == REALPART_EXPR)
3902 expr = &TREE_OPERAND (*expr, 0);
3906 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
3907 if (!base || size == -1 || max_size == -1)
3910 if (INDIRECT_REF_P (base))
3911 base = TREE_OPERAND (base, 0);
3913 base = get_ssa_base_param (base);
3914 if (!base || TREE_CODE (base) != PARM_DECL)
3917 for (i = 0; i < len; i++)
3919 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3921 if (adj->base == base &&
3922 (adj->offset == offset || adj->remove_param))
3928 if (!cand || cand->copy_param || cand->remove_param)
3934 src = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (cand->reduction)),
3936 folded = gimple_fold_indirect_ref (src);
3941 src = cand->reduction;
3943 if (dump_file && (dump_flags & TDF_DETAILS))
3945 fprintf (dump_file, "About to replace expr ");
3946 print_generic_expr (dump_file, *expr, 0);
3947 fprintf (dump_file, " with ");
3948 print_generic_expr (dump_file, src, 0);
3949 fprintf (dump_file, "\n");
3952 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
3954 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
3962 /* If the statement pointed to by STMT_PTR contains any expressions that need
3963 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
3964 potential type incompatibilities (GSI is used to accommodate conversion
3965 statements and must point to the statement). Return true iff the statement
3969 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
3970 ipa_parm_adjustment_vec adjustments)
3972 gimple stmt = *stmt_ptr;
3973 tree *lhs_p, *rhs_p;
3976 if (!gimple_assign_single_p (stmt))
3979 rhs_p = gimple_assign_rhs1_ptr (stmt);
3980 lhs_p = gimple_assign_lhs_ptr (stmt);
3982 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
3983 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
3986 tree new_rhs = NULL_TREE;
3988 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
3990 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
3992 /* V_C_Es of constructors can cause trouble (PR 42714). */
3993 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
3994 *rhs_p = fold_convert (TREE_TYPE (*lhs_p), integer_zero_node);
3996 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
3999 new_rhs = fold_build1_loc (gimple_location (stmt),
4000 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4003 else if (REFERENCE_CLASS_P (*rhs_p)
4004 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4005 && !is_gimple_reg (*lhs_p))
4006 /* This can happen when an assignment in between two single field
4007 structures is turned into an assignment in between two pointers to
4008 scalars (PR 42237). */
4013 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4014 true, GSI_SAME_STMT);
4016 gimple_assign_set_rhs_from_tree (gsi, tmp);
4025 /* Traverse the function body and all modifications as described in
4029 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4035 gimple_stmt_iterator gsi;
4036 bool bb_changed = false;
4038 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4039 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4041 gsi = gsi_start_bb (bb);
4042 while (!gsi_end_p (gsi))
4044 gimple stmt = gsi_stmt (gsi);
4045 bool modified = false;
4049 switch (gimple_code (stmt))
4052 t = gimple_return_retval_ptr (stmt);
4053 if (*t != NULL_TREE)
4054 modified |= sra_ipa_modify_expr (t, true, adjustments);
4058 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4059 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4063 /* Operands must be processed before the lhs. */
4064 for (i = 0; i < gimple_call_num_args (stmt); i++)
4066 t = gimple_call_arg_ptr (stmt, i);
4067 modified |= sra_ipa_modify_expr (t, true, adjustments);
4070 if (gimple_call_lhs (stmt))
4072 t = gimple_call_lhs_ptr (stmt);
4073 modified |= sra_ipa_modify_expr (t, false, adjustments);
4074 modified |= replace_removed_params_ssa_names (stmt,
4080 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4082 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4083 modified |= sra_ipa_modify_expr (t, true, adjustments);
4085 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4087 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4088 modified |= sra_ipa_modify_expr (t, false, adjustments);
4100 maybe_clean_eh_stmt (stmt);
4105 gimple_purge_dead_eh_edges (bb);
4109 /* Call gimple_debug_bind_reset_value on all debug statements describing
4110 gimple register parameters that are being removed or replaced. */
4113 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4117 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4118 for (i = 0; i < len; i++)
4120 struct ipa_parm_adjustment *adj;
4121 imm_use_iterator ui;
4125 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4126 if (adj->copy_param || !is_gimple_reg (adj->base))
4128 name = gimple_default_def (cfun, adj->base);
4131 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4133 /* All other users must have been removed by
4134 ipa_sra_modify_function_body. */
4135 gcc_assert (is_gimple_debug (stmt));
4136 gimple_debug_bind_reset_value (stmt);
4142 /* Return true iff all callers have at least as many actual arguments as there
4143 are formal parameters in the current function. */
4146 all_callers_have_enough_arguments_p (struct cgraph_node *node)
4148 struct cgraph_edge *cs;
4149 for (cs = node->callers; cs; cs = cs->next_caller)
4150 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4157 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4160 convert_callers (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4162 tree old_cur_fndecl = current_function_decl;
4163 struct cgraph_edge *cs;
4164 basic_block this_block;
4165 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4167 for (cs = node->callers; cs; cs = cs->next_caller)
4169 current_function_decl = cs->caller->decl;
4170 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4173 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4174 cs->caller->uid, cs->callee->uid,
4175 cgraph_node_name (cs->caller),
4176 cgraph_node_name (cs->callee));
4178 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4183 for (cs = node->callers; cs; cs = cs->next_caller)
4184 if (!bitmap_bit_p (recomputed_callers, cs->caller->uid))
4186 compute_inline_parameters (cs->caller);
4187 bitmap_set_bit (recomputed_callers, cs->caller->uid);
4189 BITMAP_FREE (recomputed_callers);
4191 current_function_decl = old_cur_fndecl;
4193 if (!encountered_recursive_call)
4196 FOR_EACH_BB (this_block)
4198 gimple_stmt_iterator gsi;
4200 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4202 gimple stmt = gsi_stmt (gsi);
4204 if (gimple_code (stmt) != GIMPLE_CALL)
4206 call_fndecl = gimple_call_fndecl (stmt);
4207 if (call_fndecl && cgraph_get_node (call_fndecl) == node)
4210 fprintf (dump_file, "Adjusting recursive call");
4211 ipa_modify_call_arguments (NULL, stmt, adjustments);
4219 /* Create an abstract origin declaration for OLD_DECL and make it an abstract
4220 origin of the provided decl so that there are preserved parameters for debug
4224 create_abstract_origin (tree old_decl)
4226 if (!DECL_ABSTRACT_ORIGIN (old_decl))
4228 tree new_decl = copy_node (old_decl);
4230 DECL_ABSTRACT (new_decl) = 1;
4231 SET_DECL_ASSEMBLER_NAME (new_decl, NULL_TREE);
4232 SET_DECL_RTL (new_decl, NULL);
4233 DECL_STRUCT_FUNCTION (new_decl) = NULL;
4234 DECL_ARTIFICIAL (old_decl) = 1;
4235 DECL_ABSTRACT_ORIGIN (old_decl) = new_decl;
4239 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4240 as given in ADJUSTMENTS. */
4243 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4245 struct cgraph_node *alias;
4246 for (alias = node->same_body; alias; alias = alias->next)
4247 ipa_modify_formal_parameters (alias->decl, adjustments, "ISRA");
4248 /* current_function_decl must be handled last, after same_body aliases,
4249 as following functions will use what it computed. */
4250 create_abstract_origin (current_function_decl);
4251 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4252 ipa_sra_modify_function_body (adjustments);
4253 sra_ipa_reset_debug_stmts (adjustments);
4254 convert_callers (node, adjustments);
4255 cgraph_make_node_local (node);
4259 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4260 attributes, return true otherwise. NODE is the cgraph node of the current
4264 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4266 if (!cgraph_node_can_be_local_p (node))
4269 fprintf (dump_file, "Function not local to this compilation unit.\n");
4273 if (DECL_VIRTUAL_P (current_function_decl))
4276 fprintf (dump_file, "Function is a virtual method.\n");
4280 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4281 && node->global.size >= MAX_INLINE_INSNS_AUTO)
4284 fprintf (dump_file, "Function too big to be made truly local.\n");
4292 "Function has no callers in this compilation unit.\n");
4299 fprintf (dump_file, "Function uses stdarg. \n");
4303 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4309 /* Perform early interprocedural SRA. */
4312 ipa_early_sra (void)
4314 struct cgraph_node *node = cgraph_node (current_function_decl);
4315 ipa_parm_adjustment_vec adjustments;
4318 if (!ipa_sra_preliminary_function_checks (node))
4322 sra_mode = SRA_MODE_EARLY_IPA;
4324 if (!find_param_candidates ())
4327 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4331 if (!all_callers_have_enough_arguments_p (node))
4334 fprintf (dump_file, "There are callers with insufficient number of "
4339 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4341 * last_basic_block_for_function (cfun));
4342 final_bbs = BITMAP_ALLOC (NULL);
4345 if (encountered_apply_args)
4348 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4352 if (encountered_unchangable_recursive_call)
4355 fprintf (dump_file, "Function calls itself with insufficient "
4356 "number of arguments.\n");
4360 adjustments = analyze_all_param_acesses ();
4364 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4366 modify_function (node, adjustments);
4367 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4368 ret = TODO_update_ssa;
4370 statistics_counter_event (cfun, "Unused parameters deleted",
4371 sra_stats.deleted_unused_parameters);
4372 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4373 sra_stats.scalar_by_ref_to_by_val);
4374 statistics_counter_event (cfun, "Aggregate parameters broken up",
4375 sra_stats.aggregate_params_reduced);
4376 statistics_counter_event (cfun, "Aggregate parameter components created",
4377 sra_stats.param_reductions_created);
4380 BITMAP_FREE (final_bbs);
4381 free (bb_dereferences);
4383 sra_deinitialize ();
4387 /* Return if early ipa sra shall be performed. */
4389 ipa_early_sra_gate (void)
4391 return flag_ipa_sra;
4394 struct gimple_opt_pass pass_early_ipa_sra =
4398 "eipa_sra", /* name */
4399 ipa_early_sra_gate, /* gate */
4400 ipa_early_sra, /* execute */
4403 0, /* static_pass_number */
4404 TV_IPA_SRA, /* tv_id */
4405 0, /* properties_required */
4406 0, /* properties_provided */
4407 0, /* properties_destroyed */
4408 0, /* todo_flags_start */
4409 TODO_dump_func | TODO_dump_cgraph /* todo_flags_finish */