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, "
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,
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, allow_replacements,
1795 mark_read, mark_write);
1797 root->grp_unscalarized_data |= child->grp_unscalarized_data;
1798 hole |= !child->grp_covered;
1801 if (allow_replacements && scalar && !root->first_child
1803 || (root->grp_write && (direct_read || root->grp_assignment_read)))
1804 /* We must not ICE later on when trying to build an access to the
1805 original data within the aggregate even when it is impossible to do in
1806 a defined way like in the PR 42703 testcase. Therefore we check
1807 pre-emptively here that we will be able to do that. */
1808 && build_ref_for_offset (NULL, TREE_TYPE (root->base), root->offset,
1811 if (dump_file && (dump_flags & TDF_DETAILS))
1813 fprintf (dump_file, "Marking ");
1814 print_generic_expr (dump_file, root->base, 0);
1815 fprintf (dump_file, " offset: %u, size: %u: ",
1816 (unsigned) root->offset, (unsigned) root->size);
1817 fprintf (dump_file, " to be replaced.\n");
1820 root->grp_to_be_replaced = 1;
1824 else if (covered_to < limit)
1827 if (sth_created && !hole)
1829 root->grp_covered = 1;
1832 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
1833 root->grp_unscalarized_data = 1; /* not covered and written to */
1839 /* Analyze all access trees linked by next_grp by the means of
1840 analyze_access_subtree. */
1842 analyze_access_trees (struct access *access)
1848 if (analyze_access_subtree (access, true, SRA_MR_NOT_READ, false))
1850 access = access->next_grp;
1856 /* Return true iff a potential new child of LACC at offset OFFSET and with size
1857 SIZE would conflict with an already existing one. If exactly such a child
1858 already exists in LACC, store a pointer to it in EXACT_MATCH. */
1861 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
1862 HOST_WIDE_INT size, struct access **exact_match)
1864 struct access *child;
1866 for (child = lacc->first_child; child; child = child->next_sibling)
1868 if (child->offset == norm_offset && child->size == size)
1870 *exact_match = child;
1874 if (child->offset < norm_offset + size
1875 && child->offset + child->size > norm_offset)
1882 /* Create a new child access of PARENT, with all properties just like MODEL
1883 except for its offset and with its grp_write false and grp_read true.
1884 Return the new access or NULL if it cannot be created. Note that this access
1885 is created long after all splicing and sorting, it's not located in any
1886 access vector and is automatically a representative of its group. */
1888 static struct access *
1889 create_artificial_child_access (struct access *parent, struct access *model,
1890 HOST_WIDE_INT new_offset)
1892 struct access *access;
1893 struct access **child;
1894 tree expr = parent->base;;
1896 gcc_assert (!model->grp_unscalarizable_region);
1898 if (!build_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
1899 model->type, false))
1902 access = (struct access *) pool_alloc (access_pool);
1903 memset (access, 0, sizeof (struct access));
1904 access->base = parent->base;
1905 access->expr = expr;
1906 access->offset = new_offset;
1907 access->size = model->size;
1908 access->type = model->type;
1909 access->grp_write = true;
1910 access->grp_read = false;
1912 child = &parent->first_child;
1913 while (*child && (*child)->offset < new_offset)
1914 child = &(*child)->next_sibling;
1916 access->next_sibling = *child;
1923 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
1924 true if any new subaccess was created. Additionally, if RACC is a scalar
1925 access but LACC is not, change the type of the latter, if possible. */
1928 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
1930 struct access *rchild;
1931 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
1934 if (is_gimple_reg_type (lacc->type)
1935 || lacc->grp_unscalarizable_region
1936 || racc->grp_unscalarizable_region)
1939 if (!lacc->first_child && !racc->first_child
1940 && is_gimple_reg_type (racc->type))
1942 tree t = lacc->base;
1944 if (build_ref_for_offset (&t, TREE_TYPE (t), lacc->offset, racc->type,
1948 lacc->type = racc->type;
1953 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
1955 struct access *new_acc = NULL;
1956 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
1958 if (rchild->grp_unscalarizable_region)
1961 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
1966 rchild->grp_hint = 1;
1967 new_acc->grp_hint |= new_acc->grp_read;
1968 if (rchild->first_child)
1969 ret |= propagate_subaccesses_across_link (new_acc, rchild);
1974 /* If a (part of) a union field is on the RHS of an assignment, it can
1975 have sub-accesses which do not make sense on the LHS (PR 40351).
1976 Check that this is not the case. */
1977 if (!build_ref_for_offset (NULL, TREE_TYPE (lacc->base), norm_offset,
1978 rchild->type, false))
1981 rchild->grp_hint = 1;
1982 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
1986 if (racc->first_child)
1987 propagate_subaccesses_across_link (new_acc, rchild);
1994 /* Propagate all subaccesses across assignment links. */
1997 propagate_all_subaccesses (void)
1999 while (work_queue_head)
2001 struct access *racc = pop_access_from_work_queue ();
2002 struct assign_link *link;
2004 gcc_assert (racc->first_link);
2006 for (link = racc->first_link; link; link = link->next)
2008 struct access *lacc = link->lacc;
2010 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2012 lacc = lacc->group_representative;
2013 if (propagate_subaccesses_across_link (lacc, racc)
2014 && lacc->first_link)
2015 add_access_to_work_queue (lacc);
2020 /* Go through all accesses collected throughout the (intraprocedural) analysis
2021 stage, exclude overlapping ones, identify representatives and build trees
2022 out of them, making decisions about scalarization on the way. Return true
2023 iff there are any to-be-scalarized variables after this stage. */
2026 analyze_all_variable_accesses (void)
2029 bitmap tmp = BITMAP_ALLOC (NULL);
2031 unsigned i, max_total_scalarization_size;
2033 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2034 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2036 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2037 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2038 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2040 tree var = referenced_var (i);
2042 if (TREE_CODE (var) == VAR_DECL
2043 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2044 <= max_total_scalarization_size)
2045 && type_consists_of_records_p (TREE_TYPE (var)))
2047 completely_scalarize_record (var, var, 0);
2048 if (dump_file && (dump_flags & TDF_DETAILS))
2050 fprintf (dump_file, "Will attempt to totally scalarize ");
2051 print_generic_expr (dump_file, var, 0);
2052 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2057 bitmap_copy (tmp, candidate_bitmap);
2058 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2060 tree var = referenced_var (i);
2061 struct access *access;
2063 access = sort_and_splice_var_accesses (var);
2065 build_access_trees (access);
2067 disqualify_candidate (var,
2068 "No or inhibitingly overlapping accesses.");
2071 propagate_all_subaccesses ();
2073 bitmap_copy (tmp, candidate_bitmap);
2074 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2076 tree var = referenced_var (i);
2077 struct access *access = get_first_repr_for_decl (var);
2079 if (analyze_access_trees (access))
2082 if (dump_file && (dump_flags & TDF_DETAILS))
2084 fprintf (dump_file, "\nAccess trees for ");
2085 print_generic_expr (dump_file, var, 0);
2086 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2087 dump_access_tree (dump_file, access);
2088 fprintf (dump_file, "\n");
2092 disqualify_candidate (var, "No scalar replacements to be created.");
2099 statistics_counter_event (cfun, "Scalarized aggregates", res);
2106 /* Return true iff a reference statement into aggregate AGG can be built for
2107 every single to-be-replaced accesses that is a child of ACCESS, its sibling
2108 or a child of its sibling. TOP_OFFSET is the offset from the processed
2109 access subtree that has to be subtracted from offset of each access. */
2112 ref_expr_for_all_replacements_p (struct access *access, tree agg,
2113 HOST_WIDE_INT top_offset)
2117 if (access->grp_to_be_replaced
2118 && !build_ref_for_offset (NULL, TREE_TYPE (agg),
2119 access->offset - top_offset,
2120 access->type, false))
2123 if (access->first_child
2124 && !ref_expr_for_all_replacements_p (access->first_child, agg,
2128 access = access->next_sibling;
2135 /* Generate statements copying scalar replacements of accesses within a subtree
2136 into or out of AGG. ACCESS is the first child of the root of the subtree to
2137 be processed. AGG is an aggregate type expression (can be a declaration but
2138 does not have to be, it can for example also be an indirect_ref).
2139 TOP_OFFSET is the offset of the processed subtree which has to be subtracted
2140 from offsets of individual accesses to get corresponding offsets for AGG.
2141 If CHUNK_SIZE is non-null, copy only replacements in the interval
2142 <start_offset, start_offset + chunk_size>, otherwise copy all. GSI is a
2143 statement iterator used to place the new statements. WRITE should be true
2144 when the statements should write from AGG to the replacement and false if
2145 vice versa. if INSERT_AFTER is true, new statements will be added after the
2146 current statement in GSI, they will be added before the statement
2150 generate_subtree_copies (struct access *access, tree agg,
2151 HOST_WIDE_INT top_offset,
2152 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2153 gimple_stmt_iterator *gsi, bool write,
2160 if (chunk_size && access->offset >= start_offset + chunk_size)
2163 if (access->grp_to_be_replaced
2165 || access->offset + access->size > start_offset))
2167 tree repl = get_access_replacement (access);
2171 ref_found = build_ref_for_offset (&expr, TREE_TYPE (agg),
2172 access->offset - top_offset,
2173 access->type, false);
2174 gcc_assert (ref_found);
2178 if (access->grp_partial_lhs)
2179 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2181 insert_after ? GSI_NEW_STMT
2183 stmt = gimple_build_assign (repl, expr);
2187 TREE_NO_WARNING (repl) = 1;
2188 if (access->grp_partial_lhs)
2189 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2191 insert_after ? GSI_NEW_STMT
2193 stmt = gimple_build_assign (expr, repl);
2197 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2199 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2201 sra_stats.subtree_copies++;
2204 if (access->first_child)
2205 generate_subtree_copies (access->first_child, agg, top_offset,
2206 start_offset, chunk_size, gsi,
2207 write, insert_after);
2209 access = access->next_sibling;
2214 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2215 the root of the subtree to be processed. GSI is the statement iterator used
2216 for inserting statements which are added after the current statement if
2217 INSERT_AFTER is true or before it otherwise. */
2220 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2224 struct access *child;
2226 if (access->grp_to_be_replaced)
2230 stmt = gimple_build_assign (get_access_replacement (access),
2231 fold_convert (access->type,
2232 integer_zero_node));
2234 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2236 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2240 for (child = access->first_child; child; child = child->next_sibling)
2241 init_subtree_with_zero (child, gsi, insert_after);
2244 /* Search for an access representative for the given expression EXPR and
2245 return it or NULL if it cannot be found. */
2247 static struct access *
2248 get_access_for_expr (tree expr)
2250 HOST_WIDE_INT offset, size, max_size;
2253 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2254 a different size than the size of its argument and we need the latter
2256 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2257 expr = TREE_OPERAND (expr, 0);
2259 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2260 if (max_size == -1 || !DECL_P (base))
2263 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2266 return get_var_base_offset_size_access (base, offset, max_size);
2269 /* Replace the expression EXPR with a scalar replacement if there is one and
2270 generate other statements to do type conversion or subtree copying if
2271 necessary. GSI is used to place newly created statements, WRITE is true if
2272 the expression is being written to (it is on a LHS of a statement or output
2273 in an assembly statement). */
2276 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2278 struct access *access;
2281 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2284 expr = &TREE_OPERAND (*expr, 0);
2289 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2290 expr = &TREE_OPERAND (*expr, 0);
2291 access = get_access_for_expr (*expr);
2294 type = TREE_TYPE (*expr);
2296 if (access->grp_to_be_replaced)
2298 tree repl = get_access_replacement (access);
2299 /* If we replace a non-register typed access simply use the original
2300 access expression to extract the scalar component afterwards.
2301 This happens if scalarizing a function return value or parameter
2302 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2303 gcc.c-torture/compile/20011217-1.c.
2305 We also want to use this when accessing a complex or vector which can
2306 be accessed as a different type too, potentially creating a need for
2307 type conversion (see PR42196) and when scalarized unions are involved
2308 in assembler statements (see PR42398). */
2309 if (!useless_type_conversion_p (type, access->type))
2311 tree ref = access->base;
2314 ok = build_ref_for_offset (&ref, TREE_TYPE (ref),
2315 access->offset, access->type, false);
2322 if (access->grp_partial_lhs)
2323 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2324 false, GSI_NEW_STMT);
2325 stmt = gimple_build_assign (repl, ref);
2326 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2332 if (access->grp_partial_lhs)
2333 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2334 true, GSI_SAME_STMT);
2335 stmt = gimple_build_assign (ref, repl);
2336 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2344 if (access->first_child)
2346 HOST_WIDE_INT start_offset, chunk_size;
2348 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2349 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2351 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2352 start_offset = access->offset
2353 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2356 start_offset = chunk_size = 0;
2358 generate_subtree_copies (access->first_child, access->base, 0,
2359 start_offset, chunk_size, gsi, write, write);
2364 /* Where scalar replacements of the RHS have been written to when a replacement
2365 of a LHS of an assigments cannot be direclty loaded from a replacement of
2367 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2368 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2369 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2371 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2372 base aggregate if there are unscalarized data or directly to LHS
2375 static enum unscalarized_data_handling
2376 handle_unscalarized_data_in_subtree (struct access *top_racc, tree lhs,
2377 gimple_stmt_iterator *gsi)
2379 if (top_racc->grp_unscalarized_data)
2381 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2383 return SRA_UDH_RIGHT;
2387 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2388 0, 0, gsi, false, false);
2389 return SRA_UDH_LEFT;
2394 /* Try to generate statements to load all sub-replacements in an access
2395 (sub)tree (LACC is the first child) from scalar replacements in the TOP_RACC
2396 (sub)tree. If that is not possible, refresh the TOP_RACC base aggregate and
2397 load the accesses from it. LEFT_OFFSET is the offset of the left whole
2398 subtree being copied, RIGHT_OFFSET is the same thing for the right subtree.
2399 GSI is stmt iterator used for statement insertions. *REFRESHED is true iff
2400 the rhs top aggregate has already been refreshed by contents of its scalar
2401 reductions and is set to true if this function has to do it. */
2404 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2405 HOST_WIDE_INT left_offset,
2406 HOST_WIDE_INT right_offset,
2407 gimple_stmt_iterator *old_gsi,
2408 gimple_stmt_iterator *new_gsi,
2409 enum unscalarized_data_handling *refreshed,
2412 location_t loc = EXPR_LOCATION (lacc->expr);
2415 if (lacc->grp_to_be_replaced)
2417 struct access *racc;
2418 HOST_WIDE_INT offset = lacc->offset - left_offset + right_offset;
2422 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2423 if (racc && racc->grp_to_be_replaced)
2425 rhs = get_access_replacement (racc);
2426 if (!useless_type_conversion_p (lacc->type, racc->type))
2427 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2431 /* No suitable access on the right hand side, need to load from
2432 the aggregate. See if we have to update it first... */
2433 if (*refreshed == SRA_UDH_NONE)
2434 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2437 if (*refreshed == SRA_UDH_LEFT)
2442 repl_found = build_ref_for_offset (&rhs, TREE_TYPE (rhs),
2443 lacc->offset, lacc->type,
2445 gcc_assert (repl_found);
2451 rhs = top_racc->base;
2452 repl_found = build_ref_for_offset (&rhs,
2453 TREE_TYPE (top_racc->base),
2454 offset, lacc->type, false);
2455 gcc_assert (repl_found);
2459 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2460 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2462 sra_stats.subreplacements++;
2464 else if (*refreshed == SRA_UDH_NONE
2465 && lacc->grp_read && !lacc->grp_covered)
2466 *refreshed = handle_unscalarized_data_in_subtree (top_racc, lhs,
2469 if (lacc->first_child)
2470 load_assign_lhs_subreplacements (lacc->first_child, top_racc,
2471 left_offset, right_offset,
2472 old_gsi, new_gsi, refreshed, lhs);
2473 lacc = lacc->next_sibling;
2478 /* Result code for SRA assignment modification. */
2479 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2480 SRA_AM_MODIFIED, /* stmt changed but not
2482 SRA_AM_REMOVED }; /* stmt eliminated */
2484 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2485 to the assignment and GSI is the statement iterator pointing at it. Returns
2486 the same values as sra_modify_assign. */
2488 static enum assignment_mod_result
2489 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2491 tree lhs = gimple_assign_lhs (*stmt);
2494 acc = get_access_for_expr (lhs);
2498 if (VEC_length (constructor_elt,
2499 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2501 /* I have never seen this code path trigger but if it can happen the
2502 following should handle it gracefully. */
2503 if (access_has_children_p (acc))
2504 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2506 return SRA_AM_MODIFIED;
2509 if (acc->grp_covered)
2511 init_subtree_with_zero (acc, gsi, false);
2512 unlink_stmt_vdef (*stmt);
2513 gsi_remove (gsi, true);
2514 return SRA_AM_REMOVED;
2518 init_subtree_with_zero (acc, gsi, true);
2519 return SRA_AM_MODIFIED;
2523 /* Create a new suitable default definition SSA_NAME and replace all uses of
2524 SSA with it, RACC is access describing the uninitialized part of an
2525 aggregate that is being loaded. */
2528 replace_uses_with_default_def_ssa_name (tree ssa, struct access *racc)
2532 decl = get_unrenamed_access_replacement (racc);
2534 repl = gimple_default_def (cfun, decl);
2537 repl = make_ssa_name (decl, gimple_build_nop ());
2538 set_default_def (decl, repl);
2541 replace_uses_by (ssa, repl);
2544 /* Examine both sides of the assignment statement pointed to by STMT, replace
2545 them with a scalare replacement if there is one and generate copying of
2546 replacements if scalarized aggregates have been used in the assignment. GSI
2547 is used to hold generated statements for type conversions and subtree
2550 static enum assignment_mod_result
2551 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2553 struct access *lacc, *racc;
2555 bool modify_this_stmt = false;
2556 bool force_gimple_rhs = false;
2557 location_t loc = gimple_location (*stmt);
2558 gimple_stmt_iterator orig_gsi = *gsi;
2560 if (!gimple_assign_single_p (*stmt))
2562 lhs = gimple_assign_lhs (*stmt);
2563 rhs = gimple_assign_rhs1 (*stmt);
2565 if (TREE_CODE (rhs) == CONSTRUCTOR)
2566 return sra_modify_constructor_assign (stmt, gsi);
2568 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2569 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2570 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2572 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2574 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2576 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2579 lacc = get_access_for_expr (lhs);
2580 racc = get_access_for_expr (rhs);
2584 if (lacc && lacc->grp_to_be_replaced)
2586 lhs = get_access_replacement (lacc);
2587 gimple_assign_set_lhs (*stmt, lhs);
2588 modify_this_stmt = true;
2589 if (lacc->grp_partial_lhs)
2590 force_gimple_rhs = true;
2594 if (racc && racc->grp_to_be_replaced)
2596 rhs = get_access_replacement (racc);
2597 modify_this_stmt = true;
2598 if (racc->grp_partial_lhs)
2599 force_gimple_rhs = true;
2603 if (modify_this_stmt)
2605 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2607 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2608 ??? This should move to fold_stmt which we simply should
2609 call after building a VIEW_CONVERT_EXPR here. */
2610 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2611 && !access_has_children_p (lacc))
2614 if (build_ref_for_offset (&expr, TREE_TYPE (lhs), 0,
2615 TREE_TYPE (rhs), false))
2618 gimple_assign_set_lhs (*stmt, expr);
2621 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2622 && !access_has_children_p (racc))
2625 if (build_ref_for_offset (&expr, TREE_TYPE (rhs), 0,
2626 TREE_TYPE (lhs), false))
2629 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2631 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs), rhs);
2632 if (is_gimple_reg_type (TREE_TYPE (lhs))
2633 && TREE_CODE (lhs) != SSA_NAME)
2634 force_gimple_rhs = true;
2639 /* From this point on, the function deals with assignments in between
2640 aggregates when at least one has scalar reductions of some of its
2641 components. There are three possible scenarios: Both the LHS and RHS have
2642 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2644 In the first case, we would like to load the LHS components from RHS
2645 components whenever possible. If that is not possible, we would like to
2646 read it directly from the RHS (after updating it by storing in it its own
2647 components). If there are some necessary unscalarized data in the LHS,
2648 those will be loaded by the original assignment too. If neither of these
2649 cases happen, the original statement can be removed. Most of this is done
2650 by load_assign_lhs_subreplacements.
2652 In the second case, we would like to store all RHS scalarized components
2653 directly into LHS and if they cover the aggregate completely, remove the
2654 statement too. In the third case, we want the LHS components to be loaded
2655 directly from the RHS (DSE will remove the original statement if it
2658 This is a bit complex but manageable when types match and when unions do
2659 not cause confusion in a way that we cannot really load a component of LHS
2660 from the RHS or vice versa (the access representing this level can have
2661 subaccesses that are accessible only through a different union field at a
2662 higher level - different from the one used in the examined expression).
2665 Therefore, I specially handle a fourth case, happening when there is a
2666 specific type cast or it is impossible to locate a scalarized subaccess on
2667 the other side of the expression. If that happens, I simply "refresh" the
2668 RHS by storing in it is scalarized components leave the original statement
2669 there to do the copying and then load the scalar replacements of the LHS.
2670 This is what the first branch does. */
2672 if (gimple_has_volatile_ops (*stmt)
2673 || contains_view_convert_expr_p (rhs)
2674 || contains_view_convert_expr_p (lhs)
2675 || (access_has_children_p (racc)
2676 && !ref_expr_for_all_replacements_p (racc, lhs, racc->offset))
2677 || (access_has_children_p (lacc)
2678 && !ref_expr_for_all_replacements_p (lacc, rhs, lacc->offset)))
2680 if (access_has_children_p (racc))
2681 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2683 if (access_has_children_p (lacc))
2684 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2686 sra_stats.separate_lhs_rhs_handling++;
2690 if (access_has_children_p (lacc) && access_has_children_p (racc))
2692 gimple_stmt_iterator orig_gsi = *gsi;
2693 enum unscalarized_data_handling refreshed;
2695 if (lacc->grp_read && !lacc->grp_covered)
2696 refreshed = handle_unscalarized_data_in_subtree (racc, lhs, gsi);
2698 refreshed = SRA_UDH_NONE;
2700 load_assign_lhs_subreplacements (lacc->first_child, racc,
2701 lacc->offset, racc->offset,
2702 &orig_gsi, gsi, &refreshed, lhs);
2703 if (refreshed != SRA_UDH_RIGHT)
2705 if (*stmt == gsi_stmt (*gsi))
2708 unlink_stmt_vdef (*stmt);
2709 gsi_remove (&orig_gsi, true);
2710 sra_stats.deleted++;
2711 return SRA_AM_REMOVED;
2718 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2720 if (racc->first_child)
2721 generate_subtree_copies (racc->first_child, lhs,
2722 racc->offset, 0, 0, gsi,
2724 gcc_assert (*stmt == gsi_stmt (*gsi));
2725 if (TREE_CODE (lhs) == SSA_NAME)
2726 replace_uses_with_default_def_ssa_name (lhs, racc);
2728 unlink_stmt_vdef (*stmt);
2729 gsi_remove (gsi, true);
2730 sra_stats.deleted++;
2731 return SRA_AM_REMOVED;
2733 else if (racc->first_child)
2734 generate_subtree_copies (racc->first_child, lhs,
2735 racc->offset, 0, 0, gsi, false, true);
2737 if (access_has_children_p (lacc))
2738 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2739 0, 0, gsi, true, true);
2743 /* This gimplification must be done after generate_subtree_copies, lest we
2744 insert the subtree copies in the middle of the gimplified sequence. */
2745 if (force_gimple_rhs)
2746 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2747 true, GSI_SAME_STMT);
2748 if (gimple_assign_rhs1 (*stmt) != rhs)
2750 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2751 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2754 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2757 /* Traverse the function body and all modifications as decided in
2758 analyze_all_variable_accesses. */
2761 sra_modify_function_body (void)
2767 gimple_stmt_iterator gsi = gsi_start_bb (bb);
2768 while (!gsi_end_p (gsi))
2770 gimple stmt = gsi_stmt (gsi);
2771 enum assignment_mod_result assign_result;
2772 bool modified = false, deleted = false;
2776 switch (gimple_code (stmt))
2779 t = gimple_return_retval_ptr (stmt);
2780 if (*t != NULL_TREE)
2781 modified |= sra_modify_expr (t, &gsi, false);
2785 assign_result = sra_modify_assign (&stmt, &gsi);
2786 modified |= assign_result == SRA_AM_MODIFIED;
2787 deleted = assign_result == SRA_AM_REMOVED;
2791 /* Operands must be processed before the lhs. */
2792 for (i = 0; i < gimple_call_num_args (stmt); i++)
2794 t = gimple_call_arg_ptr (stmt, i);
2795 modified |= sra_modify_expr (t, &gsi, false);
2798 if (gimple_call_lhs (stmt))
2800 t = gimple_call_lhs_ptr (stmt);
2801 modified |= sra_modify_expr (t, &gsi, true);
2806 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
2808 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
2809 modified |= sra_modify_expr (t, &gsi, false);
2811 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
2813 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
2814 modified |= sra_modify_expr (t, &gsi, true);
2825 maybe_clean_eh_stmt (stmt);
2833 /* Generate statements initializing scalar replacements of parts of function
2837 initialize_parameter_reductions (void)
2839 gimple_stmt_iterator gsi;
2840 gimple_seq seq = NULL;
2843 for (parm = DECL_ARGUMENTS (current_function_decl);
2845 parm = TREE_CHAIN (parm))
2847 VEC (access_p, heap) *access_vec;
2848 struct access *access;
2850 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
2852 access_vec = get_base_access_vector (parm);
2858 seq = gimple_seq_alloc ();
2859 gsi = gsi_start (seq);
2862 for (access = VEC_index (access_p, access_vec, 0);
2864 access = access->next_grp)
2865 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true);
2869 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
2872 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
2873 it reveals there are components of some aggregates to be scalarized, it runs
2874 the required transformations. */
2876 perform_intra_sra (void)
2881 if (!find_var_candidates ())
2884 if (!scan_function ())
2887 if (!analyze_all_variable_accesses ())
2890 sra_modify_function_body ();
2891 initialize_parameter_reductions ();
2893 statistics_counter_event (cfun, "Scalar replacements created",
2894 sra_stats.replacements);
2895 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
2896 statistics_counter_event (cfun, "Subtree copy stmts",
2897 sra_stats.subtree_copies);
2898 statistics_counter_event (cfun, "Subreplacement stmts",
2899 sra_stats.subreplacements);
2900 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
2901 statistics_counter_event (cfun, "Separate LHS and RHS handling",
2902 sra_stats.separate_lhs_rhs_handling);
2904 ret = TODO_update_ssa;
2907 sra_deinitialize ();
2911 /* Perform early intraprocedural SRA. */
2913 early_intra_sra (void)
2915 sra_mode = SRA_MODE_EARLY_INTRA;
2916 return perform_intra_sra ();
2919 /* Perform "late" intraprocedural SRA. */
2921 late_intra_sra (void)
2923 sra_mode = SRA_MODE_INTRA;
2924 return perform_intra_sra ();
2929 gate_intra_sra (void)
2931 return flag_tree_sra != 0;
2935 struct gimple_opt_pass pass_sra_early =
2940 gate_intra_sra, /* gate */
2941 early_intra_sra, /* execute */
2944 0, /* static_pass_number */
2945 TV_TREE_SRA, /* tv_id */
2946 PROP_cfg | PROP_ssa, /* properties_required */
2947 0, /* properties_provided */
2948 0, /* properties_destroyed */
2949 0, /* todo_flags_start */
2953 | TODO_verify_ssa /* todo_flags_finish */
2957 struct gimple_opt_pass pass_sra =
2962 gate_intra_sra, /* gate */
2963 late_intra_sra, /* execute */
2966 0, /* static_pass_number */
2967 TV_TREE_SRA, /* tv_id */
2968 PROP_cfg | PROP_ssa, /* properties_required */
2969 0, /* properties_provided */
2970 0, /* properties_destroyed */
2971 TODO_update_address_taken, /* todo_flags_start */
2975 | TODO_verify_ssa /* todo_flags_finish */
2980 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
2984 is_unused_scalar_param (tree parm)
2987 return (is_gimple_reg (parm)
2988 && (!(name = gimple_default_def (cfun, parm))
2989 || has_zero_uses (name)));
2992 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
2993 examine whether there are any direct or otherwise infeasible ones. If so,
2994 return true, otherwise return false. PARM must be a gimple register with a
2995 non-NULL default definition. */
2998 ptr_parm_has_direct_uses (tree parm)
3000 imm_use_iterator ui;
3002 tree name = gimple_default_def (cfun, parm);
3005 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3008 use_operand_p use_p;
3010 if (is_gimple_debug (stmt))
3013 /* Valid uses include dereferences on the lhs and the rhs. */
3014 if (gimple_has_lhs (stmt))
3016 tree lhs = gimple_get_lhs (stmt);
3017 while (handled_component_p (lhs))
3018 lhs = TREE_OPERAND (lhs, 0);
3019 if (INDIRECT_REF_P (lhs)
3020 && TREE_OPERAND (lhs, 0) == name)
3023 if (gimple_assign_single_p (stmt))
3025 tree rhs = gimple_assign_rhs1 (stmt);
3026 while (handled_component_p (rhs))
3027 rhs = TREE_OPERAND (rhs, 0);
3028 if (INDIRECT_REF_P (rhs)
3029 && TREE_OPERAND (rhs, 0) == name)
3032 else if (is_gimple_call (stmt))
3035 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3037 tree arg = gimple_call_arg (stmt, i);
3038 while (handled_component_p (arg))
3039 arg = TREE_OPERAND (arg, 0);
3040 if (INDIRECT_REF_P (arg)
3041 && TREE_OPERAND (arg, 0) == name)
3046 /* If the number of valid uses does not match the number of
3047 uses in this stmt there is an unhandled use. */
3048 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3055 BREAK_FROM_IMM_USE_STMT (ui);
3061 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3062 them in candidate_bitmap. Note that these do not necessarily include
3063 parameter which are unused and thus can be removed. Return true iff any
3064 such candidate has been found. */
3067 find_param_candidates (void)
3073 for (parm = DECL_ARGUMENTS (current_function_decl);
3075 parm = TREE_CHAIN (parm))
3077 tree type = TREE_TYPE (parm);
3081 if (TREE_THIS_VOLATILE (parm)
3082 || TREE_ADDRESSABLE (parm)
3083 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3086 if (is_unused_scalar_param (parm))
3092 if (POINTER_TYPE_P (type))
3094 type = TREE_TYPE (type);
3096 if (TREE_CODE (type) == FUNCTION_TYPE
3097 || TYPE_VOLATILE (type)
3098 || !is_gimple_reg (parm)
3099 || is_va_list_type (type)
3100 || ptr_parm_has_direct_uses (parm))
3103 else if (!AGGREGATE_TYPE_P (type))
3106 if (!COMPLETE_TYPE_P (type)
3107 || !host_integerp (TYPE_SIZE (type), 1)
3108 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3109 || (AGGREGATE_TYPE_P (type)
3110 && type_internals_preclude_sra_p (type)))
3113 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3115 if (dump_file && (dump_flags & TDF_DETAILS))
3117 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3118 print_generic_expr (dump_file, parm, 0);
3119 fprintf (dump_file, "\n");
3123 func_param_count = count;
3127 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3131 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3134 struct access *repr = (struct access *) data;
3136 repr->grp_maybe_modified = 1;
3140 /* Analyze what representatives (in linked lists accessible from
3141 REPRESENTATIVES) can be modified by side effects of statements in the
3142 current function. */
3145 analyze_modified_params (VEC (access_p, heap) *representatives)
3149 for (i = 0; i < func_param_count; i++)
3151 struct access *repr;
3153 for (repr = VEC_index (access_p, representatives, i);
3155 repr = repr->next_grp)
3157 struct access *access;
3161 if (no_accesses_p (repr))
3163 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3164 || repr->grp_maybe_modified)
3167 ao_ref_init (&ar, repr->expr);
3168 visited = BITMAP_ALLOC (NULL);
3169 for (access = repr; access; access = access->next_sibling)
3171 /* All accesses are read ones, otherwise grp_maybe_modified would
3172 be trivially set. */
3173 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3174 mark_maybe_modified, repr, &visited);
3175 if (repr->grp_maybe_modified)
3178 BITMAP_FREE (visited);
3183 /* Propagate distances in bb_dereferences in the opposite direction than the
3184 control flow edges, in each step storing the maximum of the current value
3185 and the minimum of all successors. These steps are repeated until the table
3186 stabilizes. Note that BBs which might terminate the functions (according to
3187 final_bbs bitmap) never updated in this way. */
3190 propagate_dereference_distances (void)
3192 VEC (basic_block, heap) *queue;
3195 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3196 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3199 VEC_quick_push (basic_block, queue, bb);
3203 while (!VEC_empty (basic_block, queue))
3207 bool change = false;
3210 bb = VEC_pop (basic_block, queue);
3213 if (bitmap_bit_p (final_bbs, bb->index))
3216 for (i = 0; i < func_param_count; i++)
3218 int idx = bb->index * func_param_count + i;
3220 HOST_WIDE_INT inh = 0;
3222 FOR_EACH_EDGE (e, ei, bb->succs)
3224 int succ_idx = e->dest->index * func_param_count + i;
3226 if (e->src == EXIT_BLOCK_PTR)
3232 inh = bb_dereferences [succ_idx];
3234 else if (bb_dereferences [succ_idx] < inh)
3235 inh = bb_dereferences [succ_idx];
3238 if (!first && bb_dereferences[idx] < inh)
3240 bb_dereferences[idx] = inh;
3245 if (change && !bitmap_bit_p (final_bbs, bb->index))
3246 FOR_EACH_EDGE (e, ei, bb->preds)
3251 e->src->aux = e->src;
3252 VEC_quick_push (basic_block, queue, e->src);
3256 VEC_free (basic_block, heap, queue);
3259 /* Dump a dereferences TABLE with heading STR to file F. */
3262 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3266 fprintf (dump_file, str);
3267 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3269 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3270 if (bb != EXIT_BLOCK_PTR)
3273 for (i = 0; i < func_param_count; i++)
3275 int idx = bb->index * func_param_count + i;
3276 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3281 fprintf (dump_file, "\n");
3284 /* Determine what (parts of) parameters passed by reference that are not
3285 assigned to are not certainly dereferenced in this function and thus the
3286 dereferencing cannot be safely moved to the caller without potentially
3287 introducing a segfault. Mark such REPRESENTATIVES as
3288 grp_not_necessarilly_dereferenced.
3290 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3291 part is calculated rather than simple booleans are calculated for each
3292 pointer parameter to handle cases when only a fraction of the whole
3293 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3296 The maximum dereference distances for each pointer parameter and BB are
3297 already stored in bb_dereference. This routine simply propagates these
3298 values upwards by propagate_dereference_distances and then compares the
3299 distances of individual parameters in the ENTRY BB to the equivalent
3300 distances of each representative of a (fraction of a) parameter. */
3303 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3307 if (dump_file && (dump_flags & TDF_DETAILS))
3308 dump_dereferences_table (dump_file,
3309 "Dereference table before propagation:\n",
3312 propagate_dereference_distances ();
3314 if (dump_file && (dump_flags & TDF_DETAILS))
3315 dump_dereferences_table (dump_file,
3316 "Dereference table after propagation:\n",
3319 for (i = 0; i < func_param_count; i++)
3321 struct access *repr = VEC_index (access_p, representatives, i);
3322 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3324 if (!repr || no_accesses_p (repr))
3329 if ((repr->offset + repr->size) > bb_dereferences[idx])
3330 repr->grp_not_necessarilly_dereferenced = 1;
3331 repr = repr->next_grp;
3337 /* Return the representative access for the parameter declaration PARM if it is
3338 a scalar passed by reference which is not written to and the pointer value
3339 is not used directly. Thus, if it is legal to dereference it in the caller
3340 and we can rule out modifications through aliases, such parameter should be
3341 turned into one passed by value. Return NULL otherwise. */
3343 static struct access *
3344 unmodified_by_ref_scalar_representative (tree parm)
3346 int i, access_count;
3347 struct access *repr;
3348 VEC (access_p, heap) *access_vec;
3350 access_vec = get_base_access_vector (parm);
3351 gcc_assert (access_vec);
3352 repr = VEC_index (access_p, access_vec, 0);
3355 repr->group_representative = repr;
3357 access_count = VEC_length (access_p, access_vec);
3358 for (i = 1; i < access_count; i++)
3360 struct access *access = VEC_index (access_p, access_vec, i);
3363 access->group_representative = repr;
3364 access->next_sibling = repr->next_sibling;
3365 repr->next_sibling = access;
3369 repr->grp_scalar_ptr = 1;
3373 /* Return true iff this access precludes IPA-SRA of the parameter it is
3377 access_precludes_ipa_sra_p (struct access *access)
3379 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3380 is incompatible assign in a call statement (and possibly even in asm
3381 statements). This can be relaxed by using a new temporary but only for
3382 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3383 intraprocedural SRA we deal with this by keeping the old aggregate around,
3384 something we cannot do in IPA-SRA.) */
3386 && (is_gimple_call (access->stmt)
3387 || gimple_code (access->stmt) == GIMPLE_ASM))
3394 /* Sort collected accesses for parameter PARM, identify representatives for
3395 each accessed region and link them together. Return NULL if there are
3396 different but overlapping accesses, return the special ptr value meaning
3397 there are no accesses for this parameter if that is the case and return the
3398 first representative otherwise. Set *RO_GRP if there is a group of accesses
3399 with only read (i.e. no write) accesses. */
3401 static struct access *
3402 splice_param_accesses (tree parm, bool *ro_grp)
3404 int i, j, access_count, group_count;
3405 int agg_size, total_size = 0;
3406 struct access *access, *res, **prev_acc_ptr = &res;
3407 VEC (access_p, heap) *access_vec;
3409 access_vec = get_base_access_vector (parm);
3411 return &no_accesses_representant;
3412 access_count = VEC_length (access_p, access_vec);
3414 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
3415 compare_access_positions);
3420 while (i < access_count)
3423 access = VEC_index (access_p, access_vec, i);
3424 modification = access->write;
3425 if (access_precludes_ipa_sra_p (access))
3428 /* Access is about to become group representative unless we find some
3429 nasty overlap which would preclude us from breaking this parameter
3433 while (j < access_count)
3435 struct access *ac2 = VEC_index (access_p, access_vec, j);
3436 if (ac2->offset != access->offset)
3438 /* All or nothing law for parameters. */
3439 if (access->offset + access->size > ac2->offset)
3444 else if (ac2->size != access->size)
3447 if (access_precludes_ipa_sra_p (ac2))
3450 modification |= ac2->write;
3451 ac2->group_representative = access;
3452 ac2->next_sibling = access->next_sibling;
3453 access->next_sibling = ac2;
3458 access->grp_maybe_modified = modification;
3461 *prev_acc_ptr = access;
3462 prev_acc_ptr = &access->next_grp;
3463 total_size += access->size;
3467 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3468 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3470 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3471 if (total_size >= agg_size)
3474 gcc_assert (group_count > 0);
3478 /* Decide whether parameters with representative accesses given by REPR should
3479 be reduced into components. */
3482 decide_one_param_reduction (struct access *repr)
3484 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3489 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3490 gcc_assert (cur_parm_size > 0);
3492 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3495 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3500 agg_size = cur_parm_size;
3506 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3507 print_generic_expr (dump_file, parm, 0);
3508 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3509 for (acc = repr; acc; acc = acc->next_grp)
3510 dump_access (dump_file, acc, true);
3514 new_param_count = 0;
3516 for (; repr; repr = repr->next_grp)
3518 gcc_assert (parm == repr->base);
3521 if (!by_ref || (!repr->grp_maybe_modified
3522 && !repr->grp_not_necessarilly_dereferenced))
3523 total_size += repr->size;
3525 total_size += cur_parm_size;
3528 gcc_assert (new_param_count > 0);
3530 if (optimize_function_for_size_p (cfun))
3531 parm_size_limit = cur_parm_size;
3533 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3536 if (total_size < agg_size
3537 && total_size <= parm_size_limit)
3540 fprintf (dump_file, " ....will be split into %i components\n",
3542 return new_param_count;
3548 /* The order of the following enums is important, we need to do extra work for
3549 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3550 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3551 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3553 /* Identify representatives of all accesses to all candidate parameters for
3554 IPA-SRA. Return result based on what representatives have been found. */
3556 static enum ipa_splicing_result
3557 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3559 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3561 struct access *repr;
3563 *representatives = VEC_alloc (access_p, heap, func_param_count);
3565 for (parm = DECL_ARGUMENTS (current_function_decl);
3567 parm = TREE_CHAIN (parm))
3569 if (is_unused_scalar_param (parm))
3571 VEC_quick_push (access_p, *representatives,
3572 &no_accesses_representant);
3573 if (result == NO_GOOD_ACCESS)
3574 result = UNUSED_PARAMS;
3576 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3577 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3578 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3580 repr = unmodified_by_ref_scalar_representative (parm);
3581 VEC_quick_push (access_p, *representatives, repr);
3583 result = UNMODIF_BY_REF_ACCESSES;
3585 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3587 bool ro_grp = false;
3588 repr = splice_param_accesses (parm, &ro_grp);
3589 VEC_quick_push (access_p, *representatives, repr);
3591 if (repr && !no_accesses_p (repr))
3593 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3596 result = UNMODIF_BY_REF_ACCESSES;
3597 else if (result < MODIF_BY_REF_ACCESSES)
3598 result = MODIF_BY_REF_ACCESSES;
3600 else if (result < BY_VAL_ACCESSES)
3601 result = BY_VAL_ACCESSES;
3603 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3604 result = UNUSED_PARAMS;
3607 VEC_quick_push (access_p, *representatives, NULL);
3610 if (result == NO_GOOD_ACCESS)
3612 VEC_free (access_p, heap, *representatives);
3613 *representatives = NULL;
3614 return NO_GOOD_ACCESS;
3620 /* Return the index of BASE in PARMS. Abort if it is not found. */
3623 get_param_index (tree base, VEC(tree, heap) *parms)
3627 len = VEC_length (tree, parms);
3628 for (i = 0; i < len; i++)
3629 if (VEC_index (tree, parms, i) == base)
3634 /* Convert the decisions made at the representative level into compact
3635 parameter adjustments. REPRESENTATIVES are pointers to first
3636 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3637 final number of adjustments. */
3639 static ipa_parm_adjustment_vec
3640 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3641 int adjustments_count)
3643 VEC (tree, heap) *parms;
3644 ipa_parm_adjustment_vec adjustments;
3648 gcc_assert (adjustments_count > 0);
3649 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3650 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3651 parm = DECL_ARGUMENTS (current_function_decl);
3652 for (i = 0; i < func_param_count; i++, parm = TREE_CHAIN (parm))
3654 struct access *repr = VEC_index (access_p, representatives, i);
3656 if (!repr || no_accesses_p (repr))
3658 struct ipa_parm_adjustment *adj;
3660 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3661 memset (adj, 0, sizeof (*adj));
3662 adj->base_index = get_param_index (parm, parms);
3665 adj->copy_param = 1;
3667 adj->remove_param = 1;
3671 struct ipa_parm_adjustment *adj;
3672 int index = get_param_index (parm, parms);
3674 for (; repr; repr = repr->next_grp)
3676 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3677 memset (adj, 0, sizeof (*adj));
3678 gcc_assert (repr->base == parm);
3679 adj->base_index = index;
3680 adj->base = repr->base;
3681 adj->type = repr->type;
3682 adj->offset = repr->offset;
3683 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3684 && (repr->grp_maybe_modified
3685 || repr->grp_not_necessarilly_dereferenced));
3690 VEC_free (tree, heap, parms);
3694 /* Analyze the collected accesses and produce a plan what to do with the
3695 parameters in the form of adjustments, NULL meaning nothing. */
3697 static ipa_parm_adjustment_vec
3698 analyze_all_param_acesses (void)
3700 enum ipa_splicing_result repr_state;
3701 bool proceed = false;
3702 int i, adjustments_count = 0;
3703 VEC (access_p, heap) *representatives;
3704 ipa_parm_adjustment_vec adjustments;
3706 repr_state = splice_all_param_accesses (&representatives);
3707 if (repr_state == NO_GOOD_ACCESS)
3710 /* If there are any parameters passed by reference which are not modified
3711 directly, we need to check whether they can be modified indirectly. */
3712 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3714 analyze_caller_dereference_legality (representatives);
3715 analyze_modified_params (representatives);
3718 for (i = 0; i < func_param_count; i++)
3720 struct access *repr = VEC_index (access_p, representatives, i);
3722 if (repr && !no_accesses_p (repr))
3724 if (repr->grp_scalar_ptr)
3726 adjustments_count++;
3727 if (repr->grp_not_necessarilly_dereferenced
3728 || repr->grp_maybe_modified)
3729 VEC_replace (access_p, representatives, i, NULL);
3733 sra_stats.scalar_by_ref_to_by_val++;
3738 int new_components = decide_one_param_reduction (repr);
3740 if (new_components == 0)
3742 VEC_replace (access_p, representatives, i, NULL);
3743 adjustments_count++;
3747 adjustments_count += new_components;
3748 sra_stats.aggregate_params_reduced++;
3749 sra_stats.param_reductions_created += new_components;
3756 if (no_accesses_p (repr))
3759 sra_stats.deleted_unused_parameters++;
3761 adjustments_count++;
3765 if (!proceed && dump_file)
3766 fprintf (dump_file, "NOT proceeding to change params.\n");
3769 adjustments = turn_representatives_into_adjustments (representatives,
3774 VEC_free (access_p, heap, representatives);
3778 /* If a parameter replacement identified by ADJ does not yet exist in the form
3779 of declaration, create it and record it, otherwise return the previously
3783 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
3786 if (!adj->new_ssa_base)
3788 char *pretty_name = make_fancy_name (adj->base);
3790 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
3791 DECL_NAME (repl) = get_identifier (pretty_name);
3792 obstack_free (&name_obstack, pretty_name);
3795 add_referenced_var (repl);
3796 adj->new_ssa_base = repl;
3799 repl = adj->new_ssa_base;
3803 /* Find the first adjustment for a particular parameter BASE in a vector of
3804 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3807 static struct ipa_parm_adjustment *
3808 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
3812 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3813 for (i = 0; i < len; i++)
3815 struct ipa_parm_adjustment *adj;
3817 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3818 if (!adj->copy_param && adj->base == base)
3825 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3826 removed because its value is not used, replace the SSA_NAME with a one
3827 relating to a created VAR_DECL together all of its uses and return true.
3828 ADJUSTMENTS is a pointer to an adjustments vector. */
3831 replace_removed_params_ssa_names (gimple stmt,
3832 ipa_parm_adjustment_vec adjustments)
3834 struct ipa_parm_adjustment *adj;
3835 tree lhs, decl, repl, name;
3837 if (gimple_code (stmt) == GIMPLE_PHI)
3838 lhs = gimple_phi_result (stmt);
3839 else if (is_gimple_assign (stmt))
3840 lhs = gimple_assign_lhs (stmt);
3841 else if (is_gimple_call (stmt))
3842 lhs = gimple_call_lhs (stmt);
3846 if (TREE_CODE (lhs) != SSA_NAME)
3848 decl = SSA_NAME_VAR (lhs);
3849 if (TREE_CODE (decl) != PARM_DECL)
3852 adj = get_adjustment_for_base (adjustments, decl);
3856 repl = get_replaced_param_substitute (adj);
3857 name = make_ssa_name (repl, stmt);
3861 fprintf (dump_file, "replacing an SSA name of a removed param ");
3862 print_generic_expr (dump_file, lhs, 0);
3863 fprintf (dump_file, " with ");
3864 print_generic_expr (dump_file, name, 0);
3865 fprintf (dump_file, "\n");
3868 if (is_gimple_assign (stmt))
3869 gimple_assign_set_lhs (stmt, name);
3870 else if (is_gimple_call (stmt))
3871 gimple_call_set_lhs (stmt, name);
3873 gimple_phi_set_result (stmt, name);
3875 replace_uses_by (lhs, name);
3879 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
3880 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
3881 specifies whether the function should care about type incompatibility the
3882 current and new expressions. If it is false, the function will leave
3883 incompatibility issues to the caller. Return true iff the expression
3887 sra_ipa_modify_expr (tree *expr, bool convert,
3888 ipa_parm_adjustment_vec adjustments)
3891 struct ipa_parm_adjustment *adj, *cand = NULL;
3892 HOST_WIDE_INT offset, size, max_size;
3895 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3897 if (TREE_CODE (*expr) == BIT_FIELD_REF
3898 || TREE_CODE (*expr) == IMAGPART_EXPR
3899 || TREE_CODE (*expr) == REALPART_EXPR)
3901 expr = &TREE_OPERAND (*expr, 0);
3905 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
3906 if (!base || size == -1 || max_size == -1)
3909 if (INDIRECT_REF_P (base))
3910 base = TREE_OPERAND (base, 0);
3912 base = get_ssa_base_param (base);
3913 if (!base || TREE_CODE (base) != PARM_DECL)
3916 for (i = 0; i < len; i++)
3918 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3920 if (adj->base == base &&
3921 (adj->offset == offset || adj->remove_param))
3927 if (!cand || cand->copy_param || cand->remove_param)
3933 src = build1 (INDIRECT_REF, TREE_TYPE (TREE_TYPE (cand->reduction)),
3935 folded = gimple_fold_indirect_ref (src);
3940 src = cand->reduction;
3942 if (dump_file && (dump_flags & TDF_DETAILS))
3944 fprintf (dump_file, "About to replace expr ");
3945 print_generic_expr (dump_file, *expr, 0);
3946 fprintf (dump_file, " with ");
3947 print_generic_expr (dump_file, src, 0);
3948 fprintf (dump_file, "\n");
3951 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
3953 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
3961 /* If the statement pointed to by STMT_PTR contains any expressions that need
3962 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
3963 potential type incompatibilities (GSI is used to accommodate conversion
3964 statements and must point to the statement). Return true iff the statement
3968 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
3969 ipa_parm_adjustment_vec adjustments)
3971 gimple stmt = *stmt_ptr;
3972 tree *lhs_p, *rhs_p;
3975 if (!gimple_assign_single_p (stmt))
3978 rhs_p = gimple_assign_rhs1_ptr (stmt);
3979 lhs_p = gimple_assign_lhs_ptr (stmt);
3981 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
3982 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
3985 tree new_rhs = NULL_TREE;
3987 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
3989 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
3991 /* V_C_Es of constructors can cause trouble (PR 42714). */
3992 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
3993 *rhs_p = fold_convert (TREE_TYPE (*lhs_p), integer_zero_node);
3995 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
3998 new_rhs = fold_build1_loc (gimple_location (stmt),
3999 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4002 else if (REFERENCE_CLASS_P (*rhs_p)
4003 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4004 && !is_gimple_reg (*lhs_p))
4005 /* This can happen when an assignment in between two single field
4006 structures is turned into an assignment in between two pointers to
4007 scalars (PR 42237). */
4012 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4013 true, GSI_SAME_STMT);
4015 gimple_assign_set_rhs_from_tree (gsi, tmp);
4024 /* Traverse the function body and all modifications as described in
4028 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4034 gimple_stmt_iterator gsi;
4035 bool bb_changed = false;
4037 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4038 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4040 gsi = gsi_start_bb (bb);
4041 while (!gsi_end_p (gsi))
4043 gimple stmt = gsi_stmt (gsi);
4044 bool modified = false;
4048 switch (gimple_code (stmt))
4051 t = gimple_return_retval_ptr (stmt);
4052 if (*t != NULL_TREE)
4053 modified |= sra_ipa_modify_expr (t, true, adjustments);
4057 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4058 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4062 /* Operands must be processed before the lhs. */
4063 for (i = 0; i < gimple_call_num_args (stmt); i++)
4065 t = gimple_call_arg_ptr (stmt, i);
4066 modified |= sra_ipa_modify_expr (t, true, adjustments);
4069 if (gimple_call_lhs (stmt))
4071 t = gimple_call_lhs_ptr (stmt);
4072 modified |= sra_ipa_modify_expr (t, false, adjustments);
4073 modified |= replace_removed_params_ssa_names (stmt,
4079 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4081 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4082 modified |= sra_ipa_modify_expr (t, true, adjustments);
4084 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4086 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4087 modified |= sra_ipa_modify_expr (t, false, adjustments);
4099 maybe_clean_eh_stmt (stmt);
4104 gimple_purge_dead_eh_edges (bb);
4108 /* Call gimple_debug_bind_reset_value on all debug statements describing
4109 gimple register parameters that are being removed or replaced. */
4112 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4116 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4117 for (i = 0; i < len; i++)
4119 struct ipa_parm_adjustment *adj;
4120 imm_use_iterator ui;
4124 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4125 if (adj->copy_param || !is_gimple_reg (adj->base))
4127 name = gimple_default_def (cfun, adj->base);
4130 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4132 /* All other users must have been removed by
4133 ipa_sra_modify_function_body. */
4134 gcc_assert (is_gimple_debug (stmt));
4135 gimple_debug_bind_reset_value (stmt);
4141 /* Return true iff all callers have at least as many actual arguments as there
4142 are formal parameters in the current function. */
4145 all_callers_have_enough_arguments_p (struct cgraph_node *node)
4147 struct cgraph_edge *cs;
4148 for (cs = node->callers; cs; cs = cs->next_caller)
4149 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4156 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4159 convert_callers (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4161 tree old_cur_fndecl = current_function_decl;
4162 struct cgraph_edge *cs;
4163 basic_block this_block;
4164 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4166 for (cs = node->callers; cs; cs = cs->next_caller)
4168 current_function_decl = cs->caller->decl;
4169 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4172 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4173 cs->caller->uid, cs->callee->uid,
4174 cgraph_node_name (cs->caller),
4175 cgraph_node_name (cs->callee));
4177 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4182 for (cs = node->callers; cs; cs = cs->next_caller)
4183 if (!bitmap_bit_p (recomputed_callers, cs->caller->uid))
4185 compute_inline_parameters (cs->caller);
4186 bitmap_set_bit (recomputed_callers, cs->caller->uid);
4188 BITMAP_FREE (recomputed_callers);
4190 current_function_decl = old_cur_fndecl;
4192 if (!encountered_recursive_call)
4195 FOR_EACH_BB (this_block)
4197 gimple_stmt_iterator gsi;
4199 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4201 gimple stmt = gsi_stmt (gsi);
4203 if (gimple_code (stmt) != GIMPLE_CALL)
4205 call_fndecl = gimple_call_fndecl (stmt);
4206 if (call_fndecl && cgraph_get_node (call_fndecl) == node)
4209 fprintf (dump_file, "Adjusting recursive call");
4210 ipa_modify_call_arguments (NULL, stmt, adjustments);
4218 /* Create an abstract origin declaration for OLD_DECL and make it an abstract
4219 origin of the provided decl so that there are preserved parameters for debug
4223 create_abstract_origin (tree old_decl)
4225 if (!DECL_ABSTRACT_ORIGIN (old_decl))
4227 tree new_decl = copy_node (old_decl);
4229 DECL_ABSTRACT (new_decl) = 1;
4230 SET_DECL_ASSEMBLER_NAME (new_decl, NULL_TREE);
4231 SET_DECL_RTL (new_decl, NULL);
4232 DECL_STRUCT_FUNCTION (new_decl) = NULL;
4233 DECL_ARTIFICIAL (old_decl) = 1;
4234 DECL_ABSTRACT_ORIGIN (old_decl) = new_decl;
4238 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4239 as given in ADJUSTMENTS. */
4242 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4244 struct cgraph_node *alias;
4245 for (alias = node->same_body; alias; alias = alias->next)
4246 ipa_modify_formal_parameters (alias->decl, adjustments, "ISRA");
4247 /* current_function_decl must be handled last, after same_body aliases,
4248 as following functions will use what it computed. */
4249 create_abstract_origin (current_function_decl);
4250 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4251 ipa_sra_modify_function_body (adjustments);
4252 sra_ipa_reset_debug_stmts (adjustments);
4253 convert_callers (node, adjustments);
4254 cgraph_make_node_local (node);
4258 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4259 attributes, return true otherwise. NODE is the cgraph node of the current
4263 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4265 if (!cgraph_node_can_be_local_p (node))
4268 fprintf (dump_file, "Function not local to this compilation unit.\n");
4272 if (DECL_VIRTUAL_P (current_function_decl))
4275 fprintf (dump_file, "Function is a virtual method.\n");
4279 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4280 && node->global.size >= MAX_INLINE_INSNS_AUTO)
4283 fprintf (dump_file, "Function too big to be made truly local.\n");
4291 "Function has no callers in this compilation unit.\n");
4298 fprintf (dump_file, "Function uses stdarg. \n");
4302 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4308 /* Perform early interprocedural SRA. */
4311 ipa_early_sra (void)
4313 struct cgraph_node *node = cgraph_node (current_function_decl);
4314 ipa_parm_adjustment_vec adjustments;
4317 if (!ipa_sra_preliminary_function_checks (node))
4321 sra_mode = SRA_MODE_EARLY_IPA;
4323 if (!find_param_candidates ())
4326 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4330 if (!all_callers_have_enough_arguments_p (node))
4333 fprintf (dump_file, "There are callers with insufficient number of "
4338 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4340 * last_basic_block_for_function (cfun));
4341 final_bbs = BITMAP_ALLOC (NULL);
4344 if (encountered_apply_args)
4347 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4351 if (encountered_unchangable_recursive_call)
4354 fprintf (dump_file, "Function calls itself with insufficient "
4355 "number of arguments.\n");
4359 adjustments = analyze_all_param_acesses ();
4363 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4365 modify_function (node, adjustments);
4366 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4367 ret = TODO_update_ssa;
4369 statistics_counter_event (cfun, "Unused parameters deleted",
4370 sra_stats.deleted_unused_parameters);
4371 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4372 sra_stats.scalar_by_ref_to_by_val);
4373 statistics_counter_event (cfun, "Aggregate parameters broken up",
4374 sra_stats.aggregate_params_reduced);
4375 statistics_counter_event (cfun, "Aggregate parameter components created",
4376 sra_stats.param_reductions_created);
4379 BITMAP_FREE (final_bbs);
4380 free (bb_dereferences);
4382 sra_deinitialize ();
4386 /* Return if early ipa sra shall be performed. */
4388 ipa_early_sra_gate (void)
4390 return flag_ipa_sra;
4393 struct gimple_opt_pass pass_early_ipa_sra =
4397 "eipa_sra", /* name */
4398 ipa_early_sra_gate, /* gate */
4399 ipa_early_sra, /* execute */
4402 0, /* static_pass_number */
4403 TV_IPA_SRA, /* tv_id */
4404 0, /* properties_required */
4405 0, /* properties_provided */
4406 0, /* properties_destroyed */
4407 0, /* todo_flags_start */
4408 TODO_dump_func | TODO_dump_cgraph /* todo_flags_finish */