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 #include "tree-inline.h"
93 #include "gimple-pretty-print.h"
95 /* Enumeration of all aggregate reductions we can do. */
96 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
97 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
98 SRA_MODE_INTRA }; /* late intraprocedural SRA */
100 /* Global variable describing which aggregate reduction we are performing at
102 static enum sra_mode sra_mode;
106 /* ACCESS represents each access to an aggregate variable (as a whole or a
107 part). It can also represent a group of accesses that refer to exactly the
108 same fragment of an aggregate (i.e. those that have exactly the same offset
109 and size). Such representatives for a single aggregate, once determined,
110 are linked in a linked list and have the group fields set.
112 Moreover, when doing intraprocedural SRA, a tree is built from those
113 representatives (by the means of first_child and next_sibling pointers), in
114 which all items in a subtree are "within" the root, i.e. their offset is
115 greater or equal to offset of the root and offset+size is smaller or equal
116 to offset+size of the root. Children of an access are sorted by offset.
118 Note that accesses to parts of vector and complex number types always
119 represented by an access to the whole complex number or a vector. It is a
120 duty of the modifying functions to replace them appropriately. */
124 /* Values returned by `get_ref_base_and_extent' for each component reference
125 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
126 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
127 HOST_WIDE_INT offset;
131 /* Expression. It is context dependent so do not use it to create new
132 expressions to access the original aggregate. See PR 42154 for a
138 /* The statement this access belongs to. */
141 /* Next group representative for this aggregate. */
142 struct access *next_grp;
144 /* Pointer to the group representative. Pointer to itself if the struct is
145 the representative. */
146 struct access *group_representative;
148 /* If this access has any children (in terms of the definition above), this
149 points to the first one. */
150 struct access *first_child;
152 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
153 described above. In IPA-SRA this is a pointer to the next access
154 belonging to the same group (having the same representative). */
155 struct access *next_sibling;
157 /* Pointers to the first and last element in the linked list of assign
159 struct assign_link *first_link, *last_link;
161 /* Pointer to the next access in the work queue. */
162 struct access *next_queued;
164 /* Replacement variable for this access "region." Never to be accessed
165 directly, always only by the means of get_access_replacement() and only
166 when grp_to_be_replaced flag is set. */
167 tree replacement_decl;
169 /* Is this particular access write access? */
172 /* Is this access an artificial one created to scalarize some record
174 unsigned total_scalarization : 1;
176 /* Is this access currently in the work queue? */
177 unsigned grp_queued : 1;
179 /* Does this group contain a write access? This flag is propagated down the
181 unsigned grp_write : 1;
183 /* Does this group contain a read access? This flag is propagated down the
185 unsigned grp_read : 1;
187 /* Does this group contain a read access that comes from an assignment
188 statement? This flag is propagated down the access tree. */
189 unsigned grp_assignment_read : 1;
191 /* Does this group contain a write access that comes from an assignment
192 statement? This flag is propagated down the access tree. */
193 unsigned grp_assignment_write : 1;
195 /* Other passes of the analysis use this bit to make function
196 analyze_access_subtree create scalar replacements for this group if
198 unsigned grp_hint : 1;
200 /* Is the subtree rooted in this access fully covered by scalar
202 unsigned grp_covered : 1;
204 /* If set to true, this access and all below it in an access tree must not be
206 unsigned grp_unscalarizable_region : 1;
208 /* Whether data have been written to parts of the aggregate covered by this
209 access which is not to be scalarized. This flag is propagated up in the
211 unsigned grp_unscalarized_data : 1;
213 /* Does this access and/or group contain a write access through a
215 unsigned grp_partial_lhs : 1;
217 /* Set when a scalar replacement should be created for this variable. We do
218 the decision and creation at different places because create_tmp_var
219 cannot be called from within FOR_EACH_REFERENCED_VAR. */
220 unsigned grp_to_be_replaced : 1;
222 /* Should TREE_NO_WARNING of a replacement be set? */
223 unsigned grp_no_warning : 1;
225 /* Is it possible that the group refers to data which might be (directly or
226 otherwise) modified? */
227 unsigned grp_maybe_modified : 1;
229 /* Set when this is a representative of a pointer to scalar (i.e. by
230 reference) parameter which we consider for turning into a plain scalar
231 (i.e. a by value parameter). */
232 unsigned grp_scalar_ptr : 1;
234 /* Set when we discover that this pointer is not safe to dereference in the
236 unsigned grp_not_necessarilly_dereferenced : 1;
239 typedef struct access *access_p;
241 DEF_VEC_P (access_p);
242 DEF_VEC_ALLOC_P (access_p, heap);
244 /* Alloc pool for allocating access structures. */
245 static alloc_pool access_pool;
247 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
248 are used to propagate subaccesses from rhs to lhs as long as they don't
249 conflict with what is already there. */
252 struct access *lacc, *racc;
253 struct assign_link *next;
256 /* Alloc pool for allocating assign link structures. */
257 static alloc_pool link_pool;
259 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
260 static struct pointer_map_t *base_access_vec;
262 /* Bitmap of candidates. */
263 static bitmap candidate_bitmap;
265 /* Bitmap of candidates which we should try to entirely scalarize away and
266 those which cannot be (because they are and need be used as a whole). */
267 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
269 /* Obstack for creation of fancy names. */
270 static struct obstack name_obstack;
272 /* Head of a linked list of accesses that need to have its subaccesses
273 propagated to their assignment counterparts. */
274 static struct access *work_queue_head;
276 /* Number of parameters of the analyzed function when doing early ipa SRA. */
277 static int func_param_count;
279 /* scan_function sets the following to true if it encounters a call to
280 __builtin_apply_args. */
281 static bool encountered_apply_args;
283 /* Set by scan_function when it finds a recursive call. */
284 static bool encountered_recursive_call;
286 /* Set by scan_function when it finds a recursive call with less actual
287 arguments than formal parameters.. */
288 static bool encountered_unchangable_recursive_call;
290 /* This is a table in which for each basic block and parameter there is a
291 distance (offset + size) in that parameter which is dereferenced and
292 accessed in that BB. */
293 static HOST_WIDE_INT *bb_dereferences;
294 /* Bitmap of BBs that can cause the function to "stop" progressing by
295 returning, throwing externally, looping infinitely or calling a function
296 which might abort etc.. */
297 static bitmap final_bbs;
299 /* Representative of no accesses at all. */
300 static struct access no_accesses_representant;
302 /* Predicate to test the special value. */
305 no_accesses_p (struct access *access)
307 return access == &no_accesses_representant;
310 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
311 representative fields are dumped, otherwise those which only describe the
312 individual access are. */
316 /* Number of processed aggregates is readily available in
317 analyze_all_variable_accesses and so is not stored here. */
319 /* Number of created scalar replacements. */
322 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
326 /* Number of statements created by generate_subtree_copies. */
329 /* Number of statements created by load_assign_lhs_subreplacements. */
332 /* Number of times sra_modify_assign has deleted a statement. */
335 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
336 RHS reparately due to type conversions or nonexistent matching
338 int separate_lhs_rhs_handling;
340 /* Number of parameters that were removed because they were unused. */
341 int deleted_unused_parameters;
343 /* Number of scalars passed as parameters by reference that have been
344 converted to be passed by value. */
345 int scalar_by_ref_to_by_val;
347 /* Number of aggregate parameters that were replaced by one or more of their
349 int aggregate_params_reduced;
351 /* Numbber of components created when splitting aggregate parameters. */
352 int param_reductions_created;
356 dump_access (FILE *f, struct access *access, bool grp)
358 fprintf (f, "access { ");
359 fprintf (f, "base = (%d)'", DECL_UID (access->base));
360 print_generic_expr (f, access->base, 0);
361 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
362 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
363 fprintf (f, ", expr = ");
364 print_generic_expr (f, access->expr, 0);
365 fprintf (f, ", type = ");
366 print_generic_expr (f, access->type, 0);
368 fprintf (f, ", grp_write = %d, total_scalarization = %d, "
369 "grp_read = %d, grp_hint = %d, grp_assignment_read = %d,"
370 "grp_assignment_write = %d, grp_covered = %d, "
371 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
372 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
373 "grp_maybe_modified = %d, "
374 "grp_not_necessarilly_dereferenced = %d\n",
375 access->grp_write, access->total_scalarization,
376 access->grp_read, access->grp_hint, access->grp_assignment_read,
377 access->grp_assignment_write, access->grp_covered,
378 access->grp_unscalarizable_region, access->grp_unscalarized_data,
379 access->grp_partial_lhs, access->grp_to_be_replaced,
380 access->grp_maybe_modified,
381 access->grp_not_necessarilly_dereferenced);
383 fprintf (f, ", write = %d, total_scalarization = %d, "
384 "grp_partial_lhs = %d\n",
385 access->write, access->total_scalarization,
386 access->grp_partial_lhs);
389 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
392 dump_access_tree_1 (FILE *f, struct access *access, int level)
398 for (i = 0; i < level; i++)
399 fputs ("* ", dump_file);
401 dump_access (f, access, true);
403 if (access->first_child)
404 dump_access_tree_1 (f, access->first_child, level + 1);
406 access = access->next_sibling;
411 /* Dump all access trees for a variable, given the pointer to the first root in
415 dump_access_tree (FILE *f, struct access *access)
417 for (; access; access = access->next_grp)
418 dump_access_tree_1 (f, access, 0);
421 /* Return true iff ACC is non-NULL and has subaccesses. */
424 access_has_children_p (struct access *acc)
426 return acc && acc->first_child;
429 /* Return a vector of pointers to accesses for the variable given in BASE or
430 NULL if there is none. */
432 static VEC (access_p, heap) *
433 get_base_access_vector (tree base)
437 slot = pointer_map_contains (base_access_vec, base);
441 return *(VEC (access_p, heap) **) slot;
444 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
445 in ACCESS. Return NULL if it cannot be found. */
447 static struct access *
448 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
451 while (access && (access->offset != offset || access->size != size))
453 struct access *child = access->first_child;
455 while (child && (child->offset + child->size <= offset))
456 child = child->next_sibling;
463 /* Return the first group representative for DECL or NULL if none exists. */
465 static struct access *
466 get_first_repr_for_decl (tree base)
468 VEC (access_p, heap) *access_vec;
470 access_vec = get_base_access_vector (base);
474 return VEC_index (access_p, access_vec, 0);
477 /* Find an access representative for the variable BASE and given OFFSET and
478 SIZE. Requires that access trees have already been built. Return NULL if
479 it cannot be found. */
481 static struct access *
482 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
485 struct access *access;
487 access = get_first_repr_for_decl (base);
488 while (access && (access->offset + access->size <= offset))
489 access = access->next_grp;
493 return find_access_in_subtree (access, offset, size);
496 /* Add LINK to the linked list of assign links of RACC. */
498 add_link_to_rhs (struct access *racc, struct assign_link *link)
500 gcc_assert (link->racc == racc);
502 if (!racc->first_link)
504 gcc_assert (!racc->last_link);
505 racc->first_link = link;
508 racc->last_link->next = link;
510 racc->last_link = link;
514 /* Move all link structures in their linked list in OLD_RACC to the linked list
517 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
519 if (!old_racc->first_link)
521 gcc_assert (!old_racc->last_link);
525 if (new_racc->first_link)
527 gcc_assert (!new_racc->last_link->next);
528 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
530 new_racc->last_link->next = old_racc->first_link;
531 new_racc->last_link = old_racc->last_link;
535 gcc_assert (!new_racc->last_link);
537 new_racc->first_link = old_racc->first_link;
538 new_racc->last_link = old_racc->last_link;
540 old_racc->first_link = old_racc->last_link = NULL;
543 /* Add ACCESS to the work queue (which is actually a stack). */
546 add_access_to_work_queue (struct access *access)
548 if (!access->grp_queued)
550 gcc_assert (!access->next_queued);
551 access->next_queued = work_queue_head;
552 access->grp_queued = 1;
553 work_queue_head = access;
557 /* Pop an access from the work queue, and return it, assuming there is one. */
559 static struct access *
560 pop_access_from_work_queue (void)
562 struct access *access = work_queue_head;
564 work_queue_head = access->next_queued;
565 access->next_queued = NULL;
566 access->grp_queued = 0;
571 /* Allocate necessary structures. */
574 sra_initialize (void)
576 candidate_bitmap = BITMAP_ALLOC (NULL);
577 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
578 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
579 gcc_obstack_init (&name_obstack);
580 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
581 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
582 base_access_vec = pointer_map_create ();
583 memset (&sra_stats, 0, sizeof (sra_stats));
584 encountered_apply_args = false;
585 encountered_recursive_call = false;
586 encountered_unchangable_recursive_call = false;
589 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
592 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
593 void *data ATTRIBUTE_UNUSED)
595 VEC (access_p, heap) *access_vec;
596 access_vec = (VEC (access_p, heap) *) *value;
597 VEC_free (access_p, heap, access_vec);
602 /* Deallocate all general structures. */
605 sra_deinitialize (void)
607 BITMAP_FREE (candidate_bitmap);
608 BITMAP_FREE (should_scalarize_away_bitmap);
609 BITMAP_FREE (cannot_scalarize_away_bitmap);
610 free_alloc_pool (access_pool);
611 free_alloc_pool (link_pool);
612 obstack_free (&name_obstack, NULL);
614 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
615 pointer_map_destroy (base_access_vec);
618 /* Remove DECL from candidates for SRA and write REASON to the dump file if
621 disqualify_candidate (tree decl, const char *reason)
623 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
625 if (dump_file && (dump_flags & TDF_DETAILS))
627 fprintf (dump_file, "! Disqualifying ");
628 print_generic_expr (dump_file, decl, 0);
629 fprintf (dump_file, " - %s\n", reason);
633 /* Return true iff the type contains a field or an element which does not allow
637 type_internals_preclude_sra_p (tree type)
642 switch (TREE_CODE (type))
646 case QUAL_UNION_TYPE:
647 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
648 if (TREE_CODE (fld) == FIELD_DECL)
650 tree ft = TREE_TYPE (fld);
652 if (TREE_THIS_VOLATILE (fld)
653 || !DECL_FIELD_OFFSET (fld) || !DECL_SIZE (fld)
654 || !host_integerp (DECL_FIELD_OFFSET (fld), 1)
655 || !host_integerp (DECL_SIZE (fld), 1)
656 || (AGGREGATE_TYPE_P (ft)
657 && int_bit_position (fld) % BITS_PER_UNIT != 0))
660 if (AGGREGATE_TYPE_P (ft)
661 && type_internals_preclude_sra_p (ft))
668 et = TREE_TYPE (type);
670 if (AGGREGATE_TYPE_P (et))
671 return type_internals_preclude_sra_p (et);
680 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
681 base variable if it is. Return T if it is not an SSA_NAME. */
684 get_ssa_base_param (tree t)
686 if (TREE_CODE (t) == SSA_NAME)
688 if (SSA_NAME_IS_DEFAULT_DEF (t))
689 return SSA_NAME_VAR (t);
696 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
697 belongs to, unless the BB has already been marked as a potentially
701 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
703 basic_block bb = gimple_bb (stmt);
704 int idx, parm_index = 0;
707 if (bitmap_bit_p (final_bbs, bb->index))
710 for (parm = DECL_ARGUMENTS (current_function_decl);
711 parm && parm != base;
712 parm = DECL_CHAIN (parm))
715 gcc_assert (parm_index < func_param_count);
717 idx = bb->index * func_param_count + parm_index;
718 if (bb_dereferences[idx] < dist)
719 bb_dereferences[idx] = dist;
722 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
723 the three fields. Also add it to the vector of accesses corresponding to
724 the base. Finally, return the new access. */
726 static struct access *
727 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
729 VEC (access_p, heap) *vec;
730 struct access *access;
733 access = (struct access *) pool_alloc (access_pool);
734 memset (access, 0, sizeof (struct access));
736 access->offset = offset;
739 slot = pointer_map_contains (base_access_vec, base);
741 vec = (VEC (access_p, heap) *) *slot;
743 vec = VEC_alloc (access_p, heap, 32);
745 VEC_safe_push (access_p, heap, vec, access);
747 *((struct VEC (access_p,heap) **)
748 pointer_map_insert (base_access_vec, base)) = vec;
753 /* Create and insert access for EXPR. Return created access, or NULL if it is
756 static struct access *
757 create_access (tree expr, gimple stmt, bool write)
759 struct access *access;
760 HOST_WIDE_INT offset, size, max_size;
762 bool ptr, unscalarizable_region = false;
764 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
766 if (sra_mode == SRA_MODE_EARLY_IPA
767 && TREE_CODE (base) == MEM_REF)
769 base = get_ssa_base_param (TREE_OPERAND (base, 0));
777 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
780 if (sra_mode == SRA_MODE_EARLY_IPA)
782 if (size < 0 || size != max_size)
784 disqualify_candidate (base, "Encountered a variable sized access.");
787 if (TREE_CODE (expr) == COMPONENT_REF
788 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
790 disqualify_candidate (base, "Encountered a bit-field access.");
793 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
796 mark_parm_dereference (base, offset + size, stmt);
800 if (size != max_size)
803 unscalarizable_region = true;
807 disqualify_candidate (base, "Encountered an unconstrained access.");
812 access = create_access_1 (base, offset, size);
814 access->type = TREE_TYPE (expr);
815 access->write = write;
816 access->grp_unscalarizable_region = unscalarizable_region;
823 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
824 register types or (recursively) records with only these two kinds of fields.
825 It also returns false if any of these records contains a bit-field. */
828 type_consists_of_records_p (tree type)
832 if (TREE_CODE (type) != RECORD_TYPE)
835 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
836 if (TREE_CODE (fld) == FIELD_DECL)
838 tree ft = TREE_TYPE (fld);
840 if (DECL_BIT_FIELD (fld))
843 if (!is_gimple_reg_type (ft)
844 && !type_consists_of_records_p (ft))
851 /* Create total_scalarization accesses for all scalar type fields in DECL that
852 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
853 must be the top-most VAR_DECL representing the variable, OFFSET must be the
854 offset of DECL within BASE. REF must be the memory reference expression for
858 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
861 tree fld, decl_type = TREE_TYPE (decl);
863 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
864 if (TREE_CODE (fld) == FIELD_DECL)
866 HOST_WIDE_INT pos = offset + int_bit_position (fld);
867 tree ft = TREE_TYPE (fld);
868 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
871 if (is_gimple_reg_type (ft))
873 struct access *access;
876 size = tree_low_cst (DECL_SIZE (fld), 1);
877 access = create_access_1 (base, pos, size);
880 access->total_scalarization = 1;
881 /* Accesses for intraprocedural SRA can have their stmt NULL. */
884 completely_scalarize_record (base, fld, pos, nref);
889 /* Search the given tree for a declaration by skipping handled components and
890 exclude it from the candidates. */
893 disqualify_base_of_expr (tree t, const char *reason)
895 t = get_base_address (t);
896 if (sra_mode == SRA_MODE_EARLY_IPA
897 && TREE_CODE (t) == MEM_REF)
898 t = get_ssa_base_param (TREE_OPERAND (t, 0));
901 disqualify_candidate (t, reason);
904 /* Scan expression EXPR and create access structures for all accesses to
905 candidates for scalarization. Return the created access or NULL if none is
908 static struct access *
909 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
911 struct access *ret = NULL;
914 if (TREE_CODE (expr) == BIT_FIELD_REF
915 || TREE_CODE (expr) == IMAGPART_EXPR
916 || TREE_CODE (expr) == REALPART_EXPR)
918 expr = TREE_OPERAND (expr, 0);
924 /* We need to dive through V_C_Es in order to get the size of its parameter
925 and not the result type. Ada produces such statements. We are also
926 capable of handling the topmost V_C_E but not any of those buried in other
927 handled components. */
928 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
929 expr = TREE_OPERAND (expr, 0);
931 if (contains_view_convert_expr_p (expr))
933 disqualify_base_of_expr (expr, "V_C_E under a different handled "
938 switch (TREE_CODE (expr))
941 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
942 && sra_mode != SRA_MODE_EARLY_IPA)
950 case ARRAY_RANGE_REF:
951 ret = create_access (expr, stmt, write);
958 if (write && partial_ref && ret)
959 ret->grp_partial_lhs = 1;
964 /* Scan expression EXPR and create access structures for all accesses to
965 candidates for scalarization. Return true if any access has been inserted.
966 STMT must be the statement from which the expression is taken, WRITE must be
967 true if the expression is a store and false otherwise. */
970 build_access_from_expr (tree expr, gimple stmt, bool write)
972 struct access *access;
974 access = build_access_from_expr_1 (expr, stmt, write);
977 /* This means the aggregate is accesses as a whole in a way other than an
978 assign statement and thus cannot be removed even if we had a scalar
979 replacement for everything. */
980 if (cannot_scalarize_away_bitmap)
981 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
987 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
988 modes in which it matters, return true iff they have been disqualified. RHS
989 may be NULL, in that case ignore it. If we scalarize an aggregate in
990 intra-SRA we may need to add statements after each statement. This is not
991 possible if a statement unconditionally has to end the basic block. */
993 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
995 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
996 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
998 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1000 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1006 /* Scan expressions occuring in STMT, create access structures for all accesses
1007 to candidates for scalarization and remove those candidates which occur in
1008 statements or expressions that prevent them from being split apart. Return
1009 true if any access has been inserted. */
1012 build_accesses_from_assign (gimple stmt)
1015 struct access *lacc, *racc;
1017 if (!gimple_assign_single_p (stmt))
1020 lhs = gimple_assign_lhs (stmt);
1021 rhs = gimple_assign_rhs1 (stmt);
1023 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1026 racc = build_access_from_expr_1 (rhs, stmt, false);
1027 lacc = build_access_from_expr_1 (lhs, stmt, true);
1030 lacc->grp_assignment_write = 1;
1034 racc->grp_assignment_read = 1;
1035 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1036 && !is_gimple_reg_type (racc->type))
1037 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1041 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1042 && !lacc->grp_unscalarizable_region
1043 && !racc->grp_unscalarizable_region
1044 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1045 /* FIXME: Turn the following line into an assert after PR 40058 is
1047 && lacc->size == racc->size
1048 && useless_type_conversion_p (lacc->type, racc->type))
1050 struct assign_link *link;
1052 link = (struct assign_link *) pool_alloc (link_pool);
1053 memset (link, 0, sizeof (struct assign_link));
1058 add_link_to_rhs (racc, link);
1061 return lacc || racc;
1064 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1065 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1068 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1069 void *data ATTRIBUTE_UNUSED)
1071 op = get_base_address (op);
1074 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1079 /* Return true iff callsite CALL has at least as many actual arguments as there
1080 are formal parameters of the function currently processed by IPA-SRA. */
1083 callsite_has_enough_arguments_p (gimple call)
1085 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1088 /* Scan function and look for interesting expressions and create access
1089 structures for them. Return true iff any access is created. */
1092 scan_function (void)
1099 gimple_stmt_iterator gsi;
1100 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1102 gimple stmt = gsi_stmt (gsi);
1106 if (final_bbs && stmt_can_throw_external (stmt))
1107 bitmap_set_bit (final_bbs, bb->index);
1108 switch (gimple_code (stmt))
1111 t = gimple_return_retval (stmt);
1113 ret |= build_access_from_expr (t, stmt, false);
1115 bitmap_set_bit (final_bbs, bb->index);
1119 ret |= build_accesses_from_assign (stmt);
1123 for (i = 0; i < gimple_call_num_args (stmt); i++)
1124 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1127 if (sra_mode == SRA_MODE_EARLY_IPA)
1129 tree dest = gimple_call_fndecl (stmt);
1130 int flags = gimple_call_flags (stmt);
1134 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1135 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1136 encountered_apply_args = true;
1137 if (cgraph_get_node (dest)
1138 == cgraph_get_node (current_function_decl))
1140 encountered_recursive_call = true;
1141 if (!callsite_has_enough_arguments_p (stmt))
1142 encountered_unchangable_recursive_call = true;
1147 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1148 bitmap_set_bit (final_bbs, bb->index);
1151 t = gimple_call_lhs (stmt);
1152 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1153 ret |= build_access_from_expr (t, stmt, true);
1157 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1160 bitmap_set_bit (final_bbs, bb->index);
1162 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1164 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1165 ret |= build_access_from_expr (t, stmt, false);
1167 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1169 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1170 ret |= build_access_from_expr (t, stmt, true);
1183 /* Helper of QSORT function. There are pointers to accesses in the array. An
1184 access is considered smaller than another if it has smaller offset or if the
1185 offsets are the same but is size is bigger. */
1188 compare_access_positions (const void *a, const void *b)
1190 const access_p *fp1 = (const access_p *) a;
1191 const access_p *fp2 = (const access_p *) b;
1192 const access_p f1 = *fp1;
1193 const access_p f2 = *fp2;
1195 if (f1->offset != f2->offset)
1196 return f1->offset < f2->offset ? -1 : 1;
1198 if (f1->size == f2->size)
1200 if (f1->type == f2->type)
1202 /* Put any non-aggregate type before any aggregate type. */
1203 else if (!is_gimple_reg_type (f1->type)
1204 && is_gimple_reg_type (f2->type))
1206 else if (is_gimple_reg_type (f1->type)
1207 && !is_gimple_reg_type (f2->type))
1209 /* Put any complex or vector type before any other scalar type. */
1210 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1211 && TREE_CODE (f1->type) != VECTOR_TYPE
1212 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1213 || TREE_CODE (f2->type) == VECTOR_TYPE))
1215 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1216 || TREE_CODE (f1->type) == VECTOR_TYPE)
1217 && TREE_CODE (f2->type) != COMPLEX_TYPE
1218 && TREE_CODE (f2->type) != VECTOR_TYPE)
1220 /* Put the integral type with the bigger precision first. */
1221 else if (INTEGRAL_TYPE_P (f1->type)
1222 && INTEGRAL_TYPE_P (f2->type))
1223 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1224 /* Put any integral type with non-full precision last. */
1225 else if (INTEGRAL_TYPE_P (f1->type)
1226 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1227 != TYPE_PRECISION (f1->type)))
1229 else if (INTEGRAL_TYPE_P (f2->type)
1230 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1231 != TYPE_PRECISION (f2->type)))
1233 /* Stabilize the sort. */
1234 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1237 /* We want the bigger accesses first, thus the opposite operator in the next
1239 return f1->size > f2->size ? -1 : 1;
1243 /* Append a name of the declaration to the name obstack. A helper function for
1247 make_fancy_decl_name (tree decl)
1251 tree name = DECL_NAME (decl);
1253 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1254 IDENTIFIER_LENGTH (name));
1257 sprintf (buffer, "D%u", DECL_UID (decl));
1258 obstack_grow (&name_obstack, buffer, strlen (buffer));
1262 /* Helper for make_fancy_name. */
1265 make_fancy_name_1 (tree expr)
1272 make_fancy_decl_name (expr);
1276 switch (TREE_CODE (expr))
1279 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1280 obstack_1grow (&name_obstack, '$');
1281 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1285 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1286 obstack_1grow (&name_obstack, '$');
1287 /* Arrays with only one element may not have a constant as their
1289 index = TREE_OPERAND (expr, 1);
1290 if (TREE_CODE (index) != INTEGER_CST)
1292 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1293 obstack_grow (&name_obstack, buffer, strlen (buffer));
1297 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1301 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1302 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1304 obstack_1grow (&name_obstack, '$');
1305 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1306 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1307 obstack_grow (&name_obstack, buffer, strlen (buffer));
1314 gcc_unreachable (); /* we treat these as scalars. */
1321 /* Create a human readable name for replacement variable of ACCESS. */
1324 make_fancy_name (tree expr)
1326 make_fancy_name_1 (expr);
1327 obstack_1grow (&name_obstack, '\0');
1328 return XOBFINISH (&name_obstack, char *);
1331 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1332 EXP_TYPE at the given OFFSET. If BASE is something for which
1333 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1334 to insert new statements either before or below the current one as specified
1335 by INSERT_AFTER. This function is not capable of handling bitfields. */
1338 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1339 tree exp_type, gimple_stmt_iterator *gsi,
1342 tree prev_base = base;
1344 HOST_WIDE_INT base_offset;
1346 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1348 base = get_addr_base_and_unit_offset (base, &base_offset);
1350 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1351 offset such as array[var_index]. */
1357 gcc_checking_assert (gsi);
1358 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1359 add_referenced_var (tmp);
1360 tmp = make_ssa_name (tmp, NULL);
1361 addr = build_fold_addr_expr (unshare_expr (prev_base));
1362 stmt = gimple_build_assign (tmp, addr);
1363 gimple_set_location (stmt, loc);
1364 SSA_NAME_DEF_STMT (tmp) = stmt;
1366 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1368 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1371 off = build_int_cst (reference_alias_ptr_type (prev_base),
1372 offset / BITS_PER_UNIT);
1375 else if (TREE_CODE (base) == MEM_REF)
1377 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1378 base_offset + offset / BITS_PER_UNIT);
1379 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off, 0);
1380 base = unshare_expr (TREE_OPERAND (base, 0));
1384 off = build_int_cst (reference_alias_ptr_type (base),
1385 base_offset + offset / BITS_PER_UNIT);
1386 base = build_fold_addr_expr (unshare_expr (base));
1389 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1392 /* Construct a memory reference to a part of an aggregate BASE at the given
1393 OFFSET and of the same type as MODEL. In case this is a reference to a
1394 component, the function will replicate the last COMPONENT_REF of model's
1395 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1396 build_ref_for_offset. */
1399 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1400 struct access *model, gimple_stmt_iterator *gsi,
1403 if (TREE_CODE (model->expr) == COMPONENT_REF)
1406 offset -= int_bit_position (TREE_OPERAND (model->expr, 1));
1407 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1408 t = build_ref_for_offset (loc, base, offset, exp_type, gsi, insert_after);
1409 return fold_build3_loc (loc, COMPONENT_REF, model->type, t,
1410 TREE_OPERAND (model->expr, 1), NULL_TREE);
1413 return build_ref_for_offset (loc, base, offset, model->type,
1417 /* Construct a memory reference consisting of component_refs and array_refs to
1418 a part of an aggregate *RES (which is of type TYPE). The requested part
1419 should have type EXP_TYPE at be the given OFFSET. This function might not
1420 succeed, it returns true when it does and only then *RES points to something
1421 meaningful. This function should be used only to build expressions that we
1422 might need to present to user (e.g. in warnings). In all other situations,
1423 build_ref_for_model or build_ref_for_offset should be used instead. */
1426 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1432 tree tr_size, index, minidx;
1433 HOST_WIDE_INT el_size;
1435 if (offset == 0 && exp_type
1436 && types_compatible_p (exp_type, type))
1439 switch (TREE_CODE (type))
1442 case QUAL_UNION_TYPE:
1444 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1446 HOST_WIDE_INT pos, size;
1447 tree expr, *expr_ptr;
1449 if (TREE_CODE (fld) != FIELD_DECL)
1452 pos = int_bit_position (fld);
1453 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1454 tr_size = DECL_SIZE (fld);
1455 if (!tr_size || !host_integerp (tr_size, 1))
1457 size = tree_low_cst (tr_size, 1);
1463 else if (pos > offset || (pos + size) <= offset)
1466 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1469 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1470 offset - pos, exp_type))
1479 tr_size = TYPE_SIZE (TREE_TYPE (type));
1480 if (!tr_size || !host_integerp (tr_size, 1))
1482 el_size = tree_low_cst (tr_size, 1);
1484 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1485 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1487 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1488 if (!integer_zerop (minidx))
1489 index = int_const_binop (PLUS_EXPR, index, minidx, 0);
1490 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1491 NULL_TREE, NULL_TREE);
1492 offset = offset % el_size;
1493 type = TREE_TYPE (type);
1508 /* Return true iff TYPE is stdarg va_list type. */
1511 is_va_list_type (tree type)
1513 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1516 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1517 those with type which is suitable for scalarization. */
1520 find_var_candidates (void)
1523 referenced_var_iterator rvi;
1526 FOR_EACH_REFERENCED_VAR (var, rvi)
1528 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1530 type = TREE_TYPE (var);
1532 if (!AGGREGATE_TYPE_P (type)
1533 || needs_to_live_in_memory (var)
1534 || TREE_THIS_VOLATILE (var)
1535 || !COMPLETE_TYPE_P (type)
1536 || !host_integerp (TYPE_SIZE (type), 1)
1537 || tree_low_cst (TYPE_SIZE (type), 1) == 0
1538 || type_internals_preclude_sra_p (type)
1539 /* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1540 we also want to schedule it rather late. Thus we ignore it in
1542 || (sra_mode == SRA_MODE_EARLY_INTRA
1543 && is_va_list_type (type)))
1546 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1548 if (dump_file && (dump_flags & TDF_DETAILS))
1550 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1551 print_generic_expr (dump_file, var, 0);
1552 fprintf (dump_file, "\n");
1560 /* Sort all accesses for the given variable, check for partial overlaps and
1561 return NULL if there are any. If there are none, pick a representative for
1562 each combination of offset and size and create a linked list out of them.
1563 Return the pointer to the first representative and make sure it is the first
1564 one in the vector of accesses. */
1566 static struct access *
1567 sort_and_splice_var_accesses (tree var)
1569 int i, j, access_count;
1570 struct access *res, **prev_acc_ptr = &res;
1571 VEC (access_p, heap) *access_vec;
1573 HOST_WIDE_INT low = -1, high = 0;
1575 access_vec = get_base_access_vector (var);
1578 access_count = VEC_length (access_p, access_vec);
1580 /* Sort by <OFFSET, SIZE>. */
1581 VEC_qsort (access_p, access_vec, compare_access_positions);
1584 while (i < access_count)
1586 struct access *access = VEC_index (access_p, access_vec, i);
1587 bool grp_write = access->write;
1588 bool grp_read = !access->write;
1589 bool grp_assignment_read = access->grp_assignment_read;
1590 bool grp_assignment_write = access->grp_assignment_write;
1591 bool multiple_reads = false;
1592 bool total_scalarization = access->total_scalarization;
1593 bool grp_partial_lhs = access->grp_partial_lhs;
1594 bool first_scalar = is_gimple_reg_type (access->type);
1595 bool unscalarizable_region = access->grp_unscalarizable_region;
1597 if (first || access->offset >= high)
1600 low = access->offset;
1601 high = access->offset + access->size;
1603 else if (access->offset > low && access->offset + access->size > high)
1606 gcc_assert (access->offset >= low
1607 && access->offset + access->size <= high);
1610 while (j < access_count)
1612 struct access *ac2 = VEC_index (access_p, access_vec, j);
1613 if (ac2->offset != access->offset || ac2->size != access->size)
1620 multiple_reads = true;
1624 grp_assignment_read |= ac2->grp_assignment_read;
1625 grp_assignment_write |= ac2->grp_assignment_write;
1626 grp_partial_lhs |= ac2->grp_partial_lhs;
1627 unscalarizable_region |= ac2->grp_unscalarizable_region;
1628 total_scalarization |= ac2->total_scalarization;
1629 relink_to_new_repr (access, ac2);
1631 /* If there are both aggregate-type and scalar-type accesses with
1632 this combination of size and offset, the comparison function
1633 should have put the scalars first. */
1634 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1635 ac2->group_representative = access;
1641 access->group_representative = access;
1642 access->grp_write = grp_write;
1643 access->grp_read = grp_read;
1644 access->grp_assignment_read = grp_assignment_read;
1645 access->grp_assignment_write = grp_assignment_write;
1646 access->grp_hint = multiple_reads || total_scalarization;
1647 access->grp_partial_lhs = grp_partial_lhs;
1648 access->grp_unscalarizable_region = unscalarizable_region;
1649 if (access->first_link)
1650 add_access_to_work_queue (access);
1652 *prev_acc_ptr = access;
1653 prev_acc_ptr = &access->next_grp;
1656 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1660 /* Create a variable for the given ACCESS which determines the type, name and a
1661 few other properties. Return the variable declaration and store it also to
1662 ACCESS->replacement. */
1665 create_access_replacement (struct access *access, bool rename)
1669 repl = create_tmp_var (access->type, "SR");
1671 add_referenced_var (repl);
1673 mark_sym_for_renaming (repl);
1675 if (!access->grp_partial_lhs
1676 && (TREE_CODE (access->type) == COMPLEX_TYPE
1677 || TREE_CODE (access->type) == VECTOR_TYPE))
1678 DECL_GIMPLE_REG_P (repl) = 1;
1680 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1681 DECL_ARTIFICIAL (repl) = 1;
1682 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1684 if (DECL_NAME (access->base)
1685 && !DECL_IGNORED_P (access->base)
1686 && !DECL_ARTIFICIAL (access->base))
1688 char *pretty_name = make_fancy_name (access->expr);
1689 tree debug_expr = unshare_expr (access->expr), d;
1691 DECL_NAME (repl) = get_identifier (pretty_name);
1692 obstack_free (&name_obstack, pretty_name);
1694 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1695 as DECL_DEBUG_EXPR isn't considered when looking for still
1696 used SSA_NAMEs and thus they could be freed. All debug info
1697 generation cares is whether something is constant or variable
1698 and that get_ref_base_and_extent works properly on the
1700 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1701 switch (TREE_CODE (d))
1704 case ARRAY_RANGE_REF:
1705 if (TREE_OPERAND (d, 1)
1706 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1707 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1708 if (TREE_OPERAND (d, 3)
1709 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1710 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1713 if (TREE_OPERAND (d, 2)
1714 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1715 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1720 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1721 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1722 if (access->grp_no_warning)
1723 TREE_NO_WARNING (repl) = 1;
1725 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1728 TREE_NO_WARNING (repl) = 1;
1732 fprintf (dump_file, "Created a replacement for ");
1733 print_generic_expr (dump_file, access->base, 0);
1734 fprintf (dump_file, " offset: %u, size: %u: ",
1735 (unsigned) access->offset, (unsigned) access->size);
1736 print_generic_expr (dump_file, repl, 0);
1737 fprintf (dump_file, "\n");
1739 sra_stats.replacements++;
1744 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1747 get_access_replacement (struct access *access)
1749 gcc_assert (access->grp_to_be_replaced);
1751 if (!access->replacement_decl)
1752 access->replacement_decl = create_access_replacement (access, true);
1753 return access->replacement_decl;
1756 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1757 not mark it for renaming. */
1760 get_unrenamed_access_replacement (struct access *access)
1762 gcc_assert (!access->grp_to_be_replaced);
1764 if (!access->replacement_decl)
1765 access->replacement_decl = create_access_replacement (access, false);
1766 return access->replacement_decl;
1770 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1771 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1772 to it is not "within" the root. Return false iff some accesses partially
1776 build_access_subtree (struct access **access)
1778 struct access *root = *access, *last_child = NULL;
1779 HOST_WIDE_INT limit = root->offset + root->size;
1781 *access = (*access)->next_grp;
1782 while (*access && (*access)->offset + (*access)->size <= limit)
1785 root->first_child = *access;
1787 last_child->next_sibling = *access;
1788 last_child = *access;
1790 if (!build_access_subtree (access))
1794 if (*access && (*access)->offset < limit)
1800 /* Build a tree of access representatives, ACCESS is the pointer to the first
1801 one, others are linked in a list by the next_grp field. Return false iff
1802 some accesses partially overlap. */
1805 build_access_trees (struct access *access)
1809 struct access *root = access;
1811 if (!build_access_subtree (&access))
1813 root->next_grp = access;
1818 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1822 expr_with_var_bounded_array_refs_p (tree expr)
1824 while (handled_component_p (expr))
1826 if (TREE_CODE (expr) == ARRAY_REF
1827 && !host_integerp (array_ref_low_bound (expr), 0))
1829 expr = TREE_OPERAND (expr, 0);
1834 enum mark_rw_status { SRA_MRRW_NOTHING, SRA_MRRW_DIRECT, SRA_MRRW_ASSIGN};
1836 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1837 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1838 sorts of access flags appropriately along the way, notably always set
1839 grp_read and grp_assign_read according to MARK_READ and grp_write when
1842 Creating a replacement for a scalar access is considered beneficial if its
1843 grp_hint is set (this means we are either attempting total scalarization or
1844 there is more than one direct read access) or according to the following
1847 Access written to individually (once or more times)
1849 | Parent written to in an assignment statement
1851 | | Access read individually _once_
1853 | | | Parent read in an assignment statement
1855 | | | | Scalarize Comment
1856 -----------------------------------------------------------------------------
1857 0 0 0 0 No access for the scalar
1858 0 0 0 1 No access for the scalar
1859 0 0 1 0 No Single read - won't help
1860 0 0 1 1 No The same case
1861 0 1 0 0 No access for the scalar
1862 0 1 0 1 No access for the scalar
1863 0 1 1 0 Yes s = *g; return s.i;
1864 0 1 1 1 Yes The same case as above
1865 1 0 0 0 No Won't help
1866 1 0 0 1 Yes s.i = 1; *g = s;
1867 1 0 1 0 Yes s.i = 5; g = s.i;
1868 1 0 1 1 Yes The same case as above
1869 1 1 0 0 No Won't help.
1870 1 1 0 1 Yes s.i = 1; *g = s;
1871 1 1 1 0 Yes s = *g; return s.i;
1872 1 1 1 1 Yes Any of the above yeses */
1875 analyze_access_subtree (struct access *root, bool allow_replacements,
1876 enum mark_rw_status mark_read,
1877 enum mark_rw_status mark_write)
1879 struct access *child;
1880 HOST_WIDE_INT limit = root->offset + root->size;
1881 HOST_WIDE_INT covered_to = root->offset;
1882 bool scalar = is_gimple_reg_type (root->type);
1883 bool hole = false, sth_created = false;
1884 bool direct_read = root->grp_read;
1885 bool direct_write = root->grp_write;
1887 if (root->grp_assignment_read)
1888 mark_read = SRA_MRRW_ASSIGN;
1889 else if (mark_read == SRA_MRRW_ASSIGN)
1892 root->grp_assignment_read = 1;
1894 else if (mark_read == SRA_MRRW_DIRECT)
1896 else if (root->grp_read)
1897 mark_read = SRA_MRRW_DIRECT;
1899 if (root->grp_assignment_write)
1900 mark_write = SRA_MRRW_ASSIGN;
1901 else if (mark_write == SRA_MRRW_ASSIGN)
1903 root->grp_write = 1;
1904 root->grp_assignment_write = 1;
1906 else if (mark_write == SRA_MRRW_DIRECT)
1907 root->grp_write = 1;
1908 else if (root->grp_write)
1909 mark_write = SRA_MRRW_DIRECT;
1911 if (root->grp_unscalarizable_region)
1912 allow_replacements = false;
1914 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
1915 allow_replacements = false;
1917 for (child = root->first_child; child; child = child->next_sibling)
1919 if (!hole && child->offset < covered_to)
1922 covered_to += child->size;
1924 sth_created |= analyze_access_subtree (child,
1925 allow_replacements && !scalar,
1926 mark_read, mark_write);
1928 root->grp_unscalarized_data |= child->grp_unscalarized_data;
1929 hole |= !child->grp_covered;
1932 if (allow_replacements && scalar && !root->first_child
1934 || ((direct_write || root->grp_assignment_write)
1935 && (direct_read || root->grp_assignment_read))))
1937 if (dump_file && (dump_flags & TDF_DETAILS))
1939 fprintf (dump_file, "Marking ");
1940 print_generic_expr (dump_file, root->base, 0);
1941 fprintf (dump_file, " offset: %u, size: %u: ",
1942 (unsigned) root->offset, (unsigned) root->size);
1943 fprintf (dump_file, " to be replaced.\n");
1946 root->grp_to_be_replaced = 1;
1950 else if (covered_to < limit)
1953 if (sth_created && !hole)
1955 root->grp_covered = 1;
1958 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
1959 root->grp_unscalarized_data = 1; /* not covered and written to */
1965 /* Analyze all access trees linked by next_grp by the means of
1966 analyze_access_subtree. */
1968 analyze_access_trees (struct access *access)
1974 if (analyze_access_subtree (access, true,
1975 SRA_MRRW_NOTHING, SRA_MRRW_NOTHING))
1977 access = access->next_grp;
1983 /* Return true iff a potential new child of LACC at offset OFFSET and with size
1984 SIZE would conflict with an already existing one. If exactly such a child
1985 already exists in LACC, store a pointer to it in EXACT_MATCH. */
1988 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
1989 HOST_WIDE_INT size, struct access **exact_match)
1991 struct access *child;
1993 for (child = lacc->first_child; child; child = child->next_sibling)
1995 if (child->offset == norm_offset && child->size == size)
1997 *exact_match = child;
2001 if (child->offset < norm_offset + size
2002 && child->offset + child->size > norm_offset)
2009 /* Create a new child access of PARENT, with all properties just like MODEL
2010 except for its offset and with its grp_write false and grp_read true.
2011 Return the new access or NULL if it cannot be created. Note that this access
2012 is created long after all splicing and sorting, it's not located in any
2013 access vector and is automatically a representative of its group. */
2015 static struct access *
2016 create_artificial_child_access (struct access *parent, struct access *model,
2017 HOST_WIDE_INT new_offset)
2019 struct access *access;
2020 struct access **child;
2021 tree expr = parent->base;
2023 gcc_assert (!model->grp_unscalarizable_region);
2025 access = (struct access *) pool_alloc (access_pool);
2026 memset (access, 0, sizeof (struct access));
2027 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2030 access->grp_no_warning = true;
2031 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2032 new_offset, model, NULL, false);
2035 access->base = parent->base;
2036 access->expr = expr;
2037 access->offset = new_offset;
2038 access->size = model->size;
2039 access->type = model->type;
2040 access->grp_write = true;
2041 access->grp_read = false;
2043 child = &parent->first_child;
2044 while (*child && (*child)->offset < new_offset)
2045 child = &(*child)->next_sibling;
2047 access->next_sibling = *child;
2054 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2055 true if any new subaccess was created. Additionally, if RACC is a scalar
2056 access but LACC is not, change the type of the latter, if possible. */
2059 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2061 struct access *rchild;
2062 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2065 if (is_gimple_reg_type (lacc->type)
2066 || lacc->grp_unscalarizable_region
2067 || racc->grp_unscalarizable_region)
2070 if (!lacc->first_child && !racc->first_child
2071 && is_gimple_reg_type (racc->type))
2073 tree t = lacc->base;
2075 lacc->type = racc->type;
2076 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t), lacc->offset,
2081 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2082 lacc->base, lacc->offset,
2084 lacc->grp_no_warning = true;
2089 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2091 struct access *new_acc = NULL;
2092 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2094 if (rchild->grp_unscalarizable_region)
2097 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2102 rchild->grp_hint = 1;
2103 new_acc->grp_hint |= new_acc->grp_read;
2104 if (rchild->first_child)
2105 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2110 rchild->grp_hint = 1;
2111 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2115 if (racc->first_child)
2116 propagate_subaccesses_across_link (new_acc, rchild);
2123 /* Propagate all subaccesses across assignment links. */
2126 propagate_all_subaccesses (void)
2128 while (work_queue_head)
2130 struct access *racc = pop_access_from_work_queue ();
2131 struct assign_link *link;
2133 gcc_assert (racc->first_link);
2135 for (link = racc->first_link; link; link = link->next)
2137 struct access *lacc = link->lacc;
2139 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2141 lacc = lacc->group_representative;
2142 if (propagate_subaccesses_across_link (lacc, racc)
2143 && lacc->first_link)
2144 add_access_to_work_queue (lacc);
2149 /* Go through all accesses collected throughout the (intraprocedural) analysis
2150 stage, exclude overlapping ones, identify representatives and build trees
2151 out of them, making decisions about scalarization on the way. Return true
2152 iff there are any to-be-scalarized variables after this stage. */
2155 analyze_all_variable_accesses (void)
2158 bitmap tmp = BITMAP_ALLOC (NULL);
2160 unsigned i, max_total_scalarization_size;
2162 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2163 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2165 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2166 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2167 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2169 tree var = referenced_var (i);
2171 if (TREE_CODE (var) == VAR_DECL
2172 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2173 <= max_total_scalarization_size)
2174 && type_consists_of_records_p (TREE_TYPE (var)))
2176 completely_scalarize_record (var, var, 0, var);
2177 if (dump_file && (dump_flags & TDF_DETAILS))
2179 fprintf (dump_file, "Will attempt to totally scalarize ");
2180 print_generic_expr (dump_file, var, 0);
2181 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2186 bitmap_copy (tmp, candidate_bitmap);
2187 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2189 tree var = referenced_var (i);
2190 struct access *access;
2192 access = sort_and_splice_var_accesses (var);
2193 if (!access || !build_access_trees (access))
2194 disqualify_candidate (var,
2195 "No or inhibitingly overlapping accesses.");
2198 propagate_all_subaccesses ();
2200 bitmap_copy (tmp, candidate_bitmap);
2201 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2203 tree var = referenced_var (i);
2204 struct access *access = get_first_repr_for_decl (var);
2206 if (analyze_access_trees (access))
2209 if (dump_file && (dump_flags & TDF_DETAILS))
2211 fprintf (dump_file, "\nAccess trees for ");
2212 print_generic_expr (dump_file, var, 0);
2213 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2214 dump_access_tree (dump_file, access);
2215 fprintf (dump_file, "\n");
2219 disqualify_candidate (var, "No scalar replacements to be created.");
2226 statistics_counter_event (cfun, "Scalarized aggregates", res);
2233 /* Generate statements copying scalar replacements of accesses within a subtree
2234 into or out of AGG. ACCESS, all its children, siblings and their children
2235 are to be processed. AGG is an aggregate type expression (can be a
2236 declaration but does not have to be, it can for example also be a mem_ref or
2237 a series of handled components). TOP_OFFSET is the offset of the processed
2238 subtree which has to be subtracted from offsets of individual accesses to
2239 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2240 replacements in the interval <start_offset, start_offset + chunk_size>,
2241 otherwise copy all. GSI is a statement iterator used to place the new
2242 statements. WRITE should be true when the statements should write from AGG
2243 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2244 statements will be added after the current statement in GSI, they will be
2245 added before the statement otherwise. */
2248 generate_subtree_copies (struct access *access, tree agg,
2249 HOST_WIDE_INT top_offset,
2250 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2251 gimple_stmt_iterator *gsi, bool write,
2252 bool insert_after, location_t loc)
2256 if (chunk_size && access->offset >= start_offset + chunk_size)
2259 if (access->grp_to_be_replaced
2261 || access->offset + access->size > start_offset))
2263 tree expr, repl = get_access_replacement (access);
2266 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2267 access, gsi, insert_after);
2271 if (access->grp_partial_lhs)
2272 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2274 insert_after ? GSI_NEW_STMT
2276 stmt = gimple_build_assign (repl, expr);
2280 TREE_NO_WARNING (repl) = 1;
2281 if (access->grp_partial_lhs)
2282 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2284 insert_after ? GSI_NEW_STMT
2286 stmt = gimple_build_assign (expr, repl);
2288 gimple_set_location (stmt, loc);
2291 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2293 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2295 sra_stats.subtree_copies++;
2298 if (access->first_child)
2299 generate_subtree_copies (access->first_child, agg, top_offset,
2300 start_offset, chunk_size, gsi,
2301 write, insert_after, loc);
2303 access = access->next_sibling;
2308 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2309 the root of the subtree to be processed. GSI is the statement iterator used
2310 for inserting statements which are added after the current statement if
2311 INSERT_AFTER is true or before it otherwise. */
2314 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2315 bool insert_after, location_t loc)
2318 struct access *child;
2320 if (access->grp_to_be_replaced)
2324 stmt = gimple_build_assign (get_access_replacement (access),
2325 build_zero_cst (access->type));
2327 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2329 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2331 gimple_set_location (stmt, loc);
2334 for (child = access->first_child; child; child = child->next_sibling)
2335 init_subtree_with_zero (child, gsi, insert_after, loc);
2338 /* Search for an access representative for the given expression EXPR and
2339 return it or NULL if it cannot be found. */
2341 static struct access *
2342 get_access_for_expr (tree expr)
2344 HOST_WIDE_INT offset, size, max_size;
2347 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2348 a different size than the size of its argument and we need the latter
2350 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2351 expr = TREE_OPERAND (expr, 0);
2353 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2354 if (max_size == -1 || !DECL_P (base))
2357 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2360 return get_var_base_offset_size_access (base, offset, max_size);
2363 /* Replace the expression EXPR with a scalar replacement if there is one and
2364 generate other statements to do type conversion or subtree copying if
2365 necessary. GSI is used to place newly created statements, WRITE is true if
2366 the expression is being written to (it is on a LHS of a statement or output
2367 in an assembly statement). */
2370 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2373 struct access *access;
2376 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2379 expr = &TREE_OPERAND (*expr, 0);
2384 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2385 expr = &TREE_OPERAND (*expr, 0);
2386 access = get_access_for_expr (*expr);
2389 type = TREE_TYPE (*expr);
2391 loc = gimple_location (gsi_stmt (*gsi));
2392 if (access->grp_to_be_replaced)
2394 tree repl = get_access_replacement (access);
2395 /* If we replace a non-register typed access simply use the original
2396 access expression to extract the scalar component afterwards.
2397 This happens if scalarizing a function return value or parameter
2398 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2399 gcc.c-torture/compile/20011217-1.c.
2401 We also want to use this when accessing a complex or vector which can
2402 be accessed as a different type too, potentially creating a need for
2403 type conversion (see PR42196) and when scalarized unions are involved
2404 in assembler statements (see PR42398). */
2405 if (!useless_type_conversion_p (type, access->type))
2409 ref = build_ref_for_model (loc, access->base, access->offset, access,
2416 if (access->grp_partial_lhs)
2417 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2418 false, GSI_NEW_STMT);
2419 stmt = gimple_build_assign (repl, ref);
2420 gimple_set_location (stmt, loc);
2421 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2427 if (access->grp_partial_lhs)
2428 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2429 true, GSI_SAME_STMT);
2430 stmt = gimple_build_assign (ref, repl);
2431 gimple_set_location (stmt, loc);
2432 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2440 if (access->first_child)
2442 HOST_WIDE_INT start_offset, chunk_size;
2444 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2445 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2447 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2448 start_offset = access->offset
2449 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2452 start_offset = chunk_size = 0;
2454 generate_subtree_copies (access->first_child, access->base, 0,
2455 start_offset, chunk_size, gsi, write, write,
2461 /* Where scalar replacements of the RHS have been written to when a replacement
2462 of a LHS of an assigments cannot be direclty loaded from a replacement of
2464 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2465 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2466 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2468 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2469 base aggregate if there are unscalarized data or directly to LHS of the
2470 statement that is pointed to by GSI otherwise. */
2472 static enum unscalarized_data_handling
2473 handle_unscalarized_data_in_subtree (struct access *top_racc,
2474 gimple_stmt_iterator *gsi)
2476 if (top_racc->grp_unscalarized_data)
2478 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2480 gimple_location (gsi_stmt (*gsi)));
2481 return SRA_UDH_RIGHT;
2485 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2486 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2487 0, 0, gsi, false, false,
2488 gimple_location (gsi_stmt (*gsi)));
2489 return SRA_UDH_LEFT;
2494 /* Try to generate statements to load all sub-replacements in an access subtree
2495 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2496 If that is not possible, refresh the TOP_RACC base aggregate and load the
2497 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2498 copied. NEW_GSI is stmt iterator used for statement insertions after the
2499 original assignment, OLD_GSI is used to insert statements before the
2500 assignment. *REFRESHED keeps the information whether we have needed to
2501 refresh replacements of the LHS and from which side of the assignments this
2505 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2506 HOST_WIDE_INT left_offset,
2507 gimple_stmt_iterator *old_gsi,
2508 gimple_stmt_iterator *new_gsi,
2509 enum unscalarized_data_handling *refreshed)
2511 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2512 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2514 if (lacc->grp_to_be_replaced)
2516 struct access *racc;
2517 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2521 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2522 if (racc && racc->grp_to_be_replaced)
2524 rhs = get_access_replacement (racc);
2525 if (!useless_type_conversion_p (lacc->type, racc->type))
2526 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2530 /* No suitable access on the right hand side, need to load from
2531 the aggregate. See if we have to update it first... */
2532 if (*refreshed == SRA_UDH_NONE)
2533 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2536 if (*refreshed == SRA_UDH_LEFT)
2537 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2540 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2544 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2545 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2546 gimple_set_location (stmt, loc);
2548 sra_stats.subreplacements++;
2550 else if (*refreshed == SRA_UDH_NONE
2551 && lacc->grp_read && !lacc->grp_covered)
2552 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2555 if (lacc->first_child)
2556 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2557 old_gsi, new_gsi, refreshed);
2561 /* Result code for SRA assignment modification. */
2562 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2563 SRA_AM_MODIFIED, /* stmt changed but not
2565 SRA_AM_REMOVED }; /* stmt eliminated */
2567 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2568 to the assignment and GSI is the statement iterator pointing at it. Returns
2569 the same values as sra_modify_assign. */
2571 static enum assignment_mod_result
2572 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2574 tree lhs = gimple_assign_lhs (*stmt);
2578 acc = get_access_for_expr (lhs);
2582 loc = gimple_location (*stmt);
2583 if (VEC_length (constructor_elt,
2584 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2586 /* I have never seen this code path trigger but if it can happen the
2587 following should handle it gracefully. */
2588 if (access_has_children_p (acc))
2589 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2591 return SRA_AM_MODIFIED;
2594 if (acc->grp_covered)
2596 init_subtree_with_zero (acc, gsi, false, loc);
2597 unlink_stmt_vdef (*stmt);
2598 gsi_remove (gsi, true);
2599 return SRA_AM_REMOVED;
2603 init_subtree_with_zero (acc, gsi, true, loc);
2604 return SRA_AM_MODIFIED;
2608 /* Create and return a new suitable default definition SSA_NAME for RACC which
2609 is an access describing an uninitialized part of an aggregate that is being
2613 get_repl_default_def_ssa_name (struct access *racc)
2617 decl = get_unrenamed_access_replacement (racc);
2619 repl = gimple_default_def (cfun, decl);
2622 repl = make_ssa_name (decl, gimple_build_nop ());
2623 set_default_def (decl, repl);
2629 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2633 contains_bitfld_comp_ref_p (const_tree ref)
2635 while (handled_component_p (ref))
2637 if (TREE_CODE (ref) == COMPONENT_REF
2638 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2640 ref = TREE_OPERAND (ref, 0);
2646 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2647 bit-field field declaration somewhere in it. */
2650 contains_vce_or_bfcref_p (const_tree ref)
2652 while (handled_component_p (ref))
2654 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2655 || (TREE_CODE (ref) == COMPONENT_REF
2656 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2658 ref = TREE_OPERAND (ref, 0);
2664 /* Examine both sides of the assignment statement pointed to by STMT, replace
2665 them with a scalare replacement if there is one and generate copying of
2666 replacements if scalarized aggregates have been used in the assignment. GSI
2667 is used to hold generated statements for type conversions and subtree
2670 static enum assignment_mod_result
2671 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2673 struct access *lacc, *racc;
2675 bool modify_this_stmt = false;
2676 bool force_gimple_rhs = false;
2678 gimple_stmt_iterator orig_gsi = *gsi;
2680 if (!gimple_assign_single_p (*stmt))
2682 lhs = gimple_assign_lhs (*stmt);
2683 rhs = gimple_assign_rhs1 (*stmt);
2685 if (TREE_CODE (rhs) == CONSTRUCTOR)
2686 return sra_modify_constructor_assign (stmt, gsi);
2688 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2689 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2690 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2692 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2694 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2696 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2699 lacc = get_access_for_expr (lhs);
2700 racc = get_access_for_expr (rhs);
2704 loc = gimple_location (*stmt);
2705 if (lacc && lacc->grp_to_be_replaced)
2707 lhs = get_access_replacement (lacc);
2708 gimple_assign_set_lhs (*stmt, lhs);
2709 modify_this_stmt = true;
2710 if (lacc->grp_partial_lhs)
2711 force_gimple_rhs = true;
2715 if (racc && racc->grp_to_be_replaced)
2717 rhs = get_access_replacement (racc);
2718 modify_this_stmt = true;
2719 if (racc->grp_partial_lhs)
2720 force_gimple_rhs = true;
2724 if (modify_this_stmt)
2726 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2728 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2729 ??? This should move to fold_stmt which we simply should
2730 call after building a VIEW_CONVERT_EXPR here. */
2731 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2732 && !contains_bitfld_comp_ref_p (lhs)
2733 && !access_has_children_p (lacc))
2735 lhs = build_ref_for_offset (loc, lhs, 0, TREE_TYPE (rhs),
2737 gimple_assign_set_lhs (*stmt, lhs);
2739 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2740 && !contains_vce_or_bfcref_p (rhs)
2741 && !access_has_children_p (racc))
2742 rhs = build_ref_for_offset (loc, rhs, 0, TREE_TYPE (lhs),
2745 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2747 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
2749 if (is_gimple_reg_type (TREE_TYPE (lhs))
2750 && TREE_CODE (lhs) != SSA_NAME)
2751 force_gimple_rhs = true;
2756 /* From this point on, the function deals with assignments in between
2757 aggregates when at least one has scalar reductions of some of its
2758 components. There are three possible scenarios: Both the LHS and RHS have
2759 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2761 In the first case, we would like to load the LHS components from RHS
2762 components whenever possible. If that is not possible, we would like to
2763 read it directly from the RHS (after updating it by storing in it its own
2764 components). If there are some necessary unscalarized data in the LHS,
2765 those will be loaded by the original assignment too. If neither of these
2766 cases happen, the original statement can be removed. Most of this is done
2767 by load_assign_lhs_subreplacements.
2769 In the second case, we would like to store all RHS scalarized components
2770 directly into LHS and if they cover the aggregate completely, remove the
2771 statement too. In the third case, we want the LHS components to be loaded
2772 directly from the RHS (DSE will remove the original statement if it
2775 This is a bit complex but manageable when types match and when unions do
2776 not cause confusion in a way that we cannot really load a component of LHS
2777 from the RHS or vice versa (the access representing this level can have
2778 subaccesses that are accessible only through a different union field at a
2779 higher level - different from the one used in the examined expression).
2782 Therefore, I specially handle a fourth case, happening when there is a
2783 specific type cast or it is impossible to locate a scalarized subaccess on
2784 the other side of the expression. If that happens, I simply "refresh" the
2785 RHS by storing in it is scalarized components leave the original statement
2786 there to do the copying and then load the scalar replacements of the LHS.
2787 This is what the first branch does. */
2789 if (gimple_has_volatile_ops (*stmt)
2790 || contains_vce_or_bfcref_p (rhs)
2791 || contains_vce_or_bfcref_p (lhs))
2793 if (access_has_children_p (racc))
2794 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2795 gsi, false, false, loc);
2796 if (access_has_children_p (lacc))
2797 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2798 gsi, true, true, loc);
2799 sra_stats.separate_lhs_rhs_handling++;
2803 if (access_has_children_p (lacc) && access_has_children_p (racc))
2805 gimple_stmt_iterator orig_gsi = *gsi;
2806 enum unscalarized_data_handling refreshed;
2808 if (lacc->grp_read && !lacc->grp_covered)
2809 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
2811 refreshed = SRA_UDH_NONE;
2813 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
2814 &orig_gsi, gsi, &refreshed);
2815 if (refreshed != SRA_UDH_RIGHT)
2818 unlink_stmt_vdef (*stmt);
2819 gsi_remove (&orig_gsi, true);
2820 sra_stats.deleted++;
2821 return SRA_AM_REMOVED;
2828 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2832 fprintf (dump_file, "Removing load: ");
2833 print_gimple_stmt (dump_file, *stmt, 0, 0);
2836 if (TREE_CODE (lhs) == SSA_NAME)
2838 rhs = get_repl_default_def_ssa_name (racc);
2839 if (!useless_type_conversion_p (TREE_TYPE (lhs),
2841 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR,
2842 TREE_TYPE (lhs), rhs);
2846 if (racc->first_child)
2847 generate_subtree_copies (racc->first_child, lhs,
2848 racc->offset, 0, 0, gsi,
2851 gcc_assert (*stmt == gsi_stmt (*gsi));
2852 unlink_stmt_vdef (*stmt);
2853 gsi_remove (gsi, true);
2854 sra_stats.deleted++;
2855 return SRA_AM_REMOVED;
2858 else if (racc->first_child)
2859 generate_subtree_copies (racc->first_child, lhs, racc->offset,
2860 0, 0, gsi, false, true, loc);
2862 if (access_has_children_p (lacc))
2863 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2864 0, 0, gsi, true, true, loc);
2868 /* This gimplification must be done after generate_subtree_copies, lest we
2869 insert the subtree copies in the middle of the gimplified sequence. */
2870 if (force_gimple_rhs)
2871 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2872 true, GSI_SAME_STMT);
2873 if (gimple_assign_rhs1 (*stmt) != rhs)
2875 modify_this_stmt = true;
2876 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2877 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2880 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2883 /* Traverse the function body and all modifications as decided in
2884 analyze_all_variable_accesses. Return true iff the CFG has been
2888 sra_modify_function_body (void)
2890 bool cfg_changed = false;
2895 gimple_stmt_iterator gsi = gsi_start_bb (bb);
2896 while (!gsi_end_p (gsi))
2898 gimple stmt = gsi_stmt (gsi);
2899 enum assignment_mod_result assign_result;
2900 bool modified = false, deleted = false;
2904 switch (gimple_code (stmt))
2907 t = gimple_return_retval_ptr (stmt);
2908 if (*t != NULL_TREE)
2909 modified |= sra_modify_expr (t, &gsi, false);
2913 assign_result = sra_modify_assign (&stmt, &gsi);
2914 modified |= assign_result == SRA_AM_MODIFIED;
2915 deleted = assign_result == SRA_AM_REMOVED;
2919 /* Operands must be processed before the lhs. */
2920 for (i = 0; i < gimple_call_num_args (stmt); i++)
2922 t = gimple_call_arg_ptr (stmt, i);
2923 modified |= sra_modify_expr (t, &gsi, false);
2926 if (gimple_call_lhs (stmt))
2928 t = gimple_call_lhs_ptr (stmt);
2929 modified |= sra_modify_expr (t, &gsi, true);
2934 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
2936 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
2937 modified |= sra_modify_expr (t, &gsi, false);
2939 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
2941 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
2942 modified |= sra_modify_expr (t, &gsi, true);
2953 if (maybe_clean_eh_stmt (stmt)
2954 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
2965 /* Generate statements initializing scalar replacements of parts of function
2969 initialize_parameter_reductions (void)
2971 gimple_stmt_iterator gsi;
2972 gimple_seq seq = NULL;
2975 for (parm = DECL_ARGUMENTS (current_function_decl);
2977 parm = DECL_CHAIN (parm))
2979 VEC (access_p, heap) *access_vec;
2980 struct access *access;
2982 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
2984 access_vec = get_base_access_vector (parm);
2990 seq = gimple_seq_alloc ();
2991 gsi = gsi_start (seq);
2994 for (access = VEC_index (access_p, access_vec, 0);
2996 access = access->next_grp)
2997 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
2998 EXPR_LOCATION (parm));
3002 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3005 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3006 it reveals there are components of some aggregates to be scalarized, it runs
3007 the required transformations. */
3009 perform_intra_sra (void)
3014 if (!find_var_candidates ())
3017 if (!scan_function ())
3020 if (!analyze_all_variable_accesses ())
3023 if (sra_modify_function_body ())
3024 ret = TODO_update_ssa | TODO_cleanup_cfg;
3026 ret = TODO_update_ssa;
3027 initialize_parameter_reductions ();
3029 statistics_counter_event (cfun, "Scalar replacements created",
3030 sra_stats.replacements);
3031 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3032 statistics_counter_event (cfun, "Subtree copy stmts",
3033 sra_stats.subtree_copies);
3034 statistics_counter_event (cfun, "Subreplacement stmts",
3035 sra_stats.subreplacements);
3036 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3037 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3038 sra_stats.separate_lhs_rhs_handling);
3041 sra_deinitialize ();
3045 /* Perform early intraprocedural SRA. */
3047 early_intra_sra (void)
3049 sra_mode = SRA_MODE_EARLY_INTRA;
3050 return perform_intra_sra ();
3053 /* Perform "late" intraprocedural SRA. */
3055 late_intra_sra (void)
3057 sra_mode = SRA_MODE_INTRA;
3058 return perform_intra_sra ();
3063 gate_intra_sra (void)
3065 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3069 struct gimple_opt_pass pass_sra_early =
3074 gate_intra_sra, /* gate */
3075 early_intra_sra, /* execute */
3078 0, /* static_pass_number */
3079 TV_TREE_SRA, /* tv_id */
3080 PROP_cfg | PROP_ssa, /* properties_required */
3081 0, /* properties_provided */
3082 0, /* properties_destroyed */
3083 0, /* todo_flags_start */
3087 | TODO_verify_ssa /* todo_flags_finish */
3091 struct gimple_opt_pass pass_sra =
3096 gate_intra_sra, /* gate */
3097 late_intra_sra, /* execute */
3100 0, /* static_pass_number */
3101 TV_TREE_SRA, /* tv_id */
3102 PROP_cfg | PROP_ssa, /* properties_required */
3103 0, /* properties_provided */
3104 0, /* properties_destroyed */
3105 TODO_update_address_taken, /* todo_flags_start */
3109 | TODO_verify_ssa /* todo_flags_finish */
3114 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3118 is_unused_scalar_param (tree parm)
3121 return (is_gimple_reg (parm)
3122 && (!(name = gimple_default_def (cfun, parm))
3123 || has_zero_uses (name)));
3126 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3127 examine whether there are any direct or otherwise infeasible ones. If so,
3128 return true, otherwise return false. PARM must be a gimple register with a
3129 non-NULL default definition. */
3132 ptr_parm_has_direct_uses (tree parm)
3134 imm_use_iterator ui;
3136 tree name = gimple_default_def (cfun, parm);
3139 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3142 use_operand_p use_p;
3144 if (is_gimple_debug (stmt))
3147 /* Valid uses include dereferences on the lhs and the rhs. */
3148 if (gimple_has_lhs (stmt))
3150 tree lhs = gimple_get_lhs (stmt);
3151 while (handled_component_p (lhs))
3152 lhs = TREE_OPERAND (lhs, 0);
3153 if (TREE_CODE (lhs) == MEM_REF
3154 && TREE_OPERAND (lhs, 0) == name
3155 && integer_zerop (TREE_OPERAND (lhs, 1))
3156 && types_compatible_p (TREE_TYPE (lhs),
3157 TREE_TYPE (TREE_TYPE (name))))
3160 if (gimple_assign_single_p (stmt))
3162 tree rhs = gimple_assign_rhs1 (stmt);
3163 while (handled_component_p (rhs))
3164 rhs = TREE_OPERAND (rhs, 0);
3165 if (TREE_CODE (rhs) == MEM_REF
3166 && TREE_OPERAND (rhs, 0) == name
3167 && integer_zerop (TREE_OPERAND (rhs, 1))
3168 && types_compatible_p (TREE_TYPE (rhs),
3169 TREE_TYPE (TREE_TYPE (name))))
3172 else if (is_gimple_call (stmt))
3175 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3177 tree arg = gimple_call_arg (stmt, i);
3178 while (handled_component_p (arg))
3179 arg = TREE_OPERAND (arg, 0);
3180 if (TREE_CODE (arg) == MEM_REF
3181 && TREE_OPERAND (arg, 0) == name
3182 && integer_zerop (TREE_OPERAND (arg, 1))
3183 && types_compatible_p (TREE_TYPE (arg),
3184 TREE_TYPE (TREE_TYPE (name))))
3189 /* If the number of valid uses does not match the number of
3190 uses in this stmt there is an unhandled use. */
3191 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3198 BREAK_FROM_IMM_USE_STMT (ui);
3204 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3205 them in candidate_bitmap. Note that these do not necessarily include
3206 parameter which are unused and thus can be removed. Return true iff any
3207 such candidate has been found. */
3210 find_param_candidates (void)
3216 for (parm = DECL_ARGUMENTS (current_function_decl);
3218 parm = DECL_CHAIN (parm))
3220 tree type = TREE_TYPE (parm);
3224 if (TREE_THIS_VOLATILE (parm)
3225 || TREE_ADDRESSABLE (parm)
3226 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3229 if (is_unused_scalar_param (parm))
3235 if (POINTER_TYPE_P (type))
3237 type = TREE_TYPE (type);
3239 if (TREE_CODE (type) == FUNCTION_TYPE
3240 || TYPE_VOLATILE (type)
3241 || (TREE_CODE (type) == ARRAY_TYPE
3242 && TYPE_NONALIASED_COMPONENT (type))
3243 || !is_gimple_reg (parm)
3244 || is_va_list_type (type)
3245 || ptr_parm_has_direct_uses (parm))
3248 else if (!AGGREGATE_TYPE_P (type))
3251 if (!COMPLETE_TYPE_P (type)
3252 || !host_integerp (TYPE_SIZE (type), 1)
3253 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3254 || (AGGREGATE_TYPE_P (type)
3255 && type_internals_preclude_sra_p (type)))
3258 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3260 if (dump_file && (dump_flags & TDF_DETAILS))
3262 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3263 print_generic_expr (dump_file, parm, 0);
3264 fprintf (dump_file, "\n");
3268 func_param_count = count;
3272 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3276 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3279 struct access *repr = (struct access *) data;
3281 repr->grp_maybe_modified = 1;
3285 /* Analyze what representatives (in linked lists accessible from
3286 REPRESENTATIVES) can be modified by side effects of statements in the
3287 current function. */
3290 analyze_modified_params (VEC (access_p, heap) *representatives)
3294 for (i = 0; i < func_param_count; i++)
3296 struct access *repr;
3298 for (repr = VEC_index (access_p, representatives, i);
3300 repr = repr->next_grp)
3302 struct access *access;
3306 if (no_accesses_p (repr))
3308 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3309 || repr->grp_maybe_modified)
3312 ao_ref_init (&ar, repr->expr);
3313 visited = BITMAP_ALLOC (NULL);
3314 for (access = repr; access; access = access->next_sibling)
3316 /* All accesses are read ones, otherwise grp_maybe_modified would
3317 be trivially set. */
3318 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3319 mark_maybe_modified, repr, &visited);
3320 if (repr->grp_maybe_modified)
3323 BITMAP_FREE (visited);
3328 /* Propagate distances in bb_dereferences in the opposite direction than the
3329 control flow edges, in each step storing the maximum of the current value
3330 and the minimum of all successors. These steps are repeated until the table
3331 stabilizes. Note that BBs which might terminate the functions (according to
3332 final_bbs bitmap) never updated in this way. */
3335 propagate_dereference_distances (void)
3337 VEC (basic_block, heap) *queue;
3340 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3341 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3344 VEC_quick_push (basic_block, queue, bb);
3348 while (!VEC_empty (basic_block, queue))
3352 bool change = false;
3355 bb = VEC_pop (basic_block, queue);
3358 if (bitmap_bit_p (final_bbs, bb->index))
3361 for (i = 0; i < func_param_count; i++)
3363 int idx = bb->index * func_param_count + i;
3365 HOST_WIDE_INT inh = 0;
3367 FOR_EACH_EDGE (e, ei, bb->succs)
3369 int succ_idx = e->dest->index * func_param_count + i;
3371 if (e->src == EXIT_BLOCK_PTR)
3377 inh = bb_dereferences [succ_idx];
3379 else if (bb_dereferences [succ_idx] < inh)
3380 inh = bb_dereferences [succ_idx];
3383 if (!first && bb_dereferences[idx] < inh)
3385 bb_dereferences[idx] = inh;
3390 if (change && !bitmap_bit_p (final_bbs, bb->index))
3391 FOR_EACH_EDGE (e, ei, bb->preds)
3396 e->src->aux = e->src;
3397 VEC_quick_push (basic_block, queue, e->src);
3401 VEC_free (basic_block, heap, queue);
3404 /* Dump a dereferences TABLE with heading STR to file F. */
3407 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3411 fprintf (dump_file, str);
3412 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3414 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3415 if (bb != EXIT_BLOCK_PTR)
3418 for (i = 0; i < func_param_count; i++)
3420 int idx = bb->index * func_param_count + i;
3421 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3426 fprintf (dump_file, "\n");
3429 /* Determine what (parts of) parameters passed by reference that are not
3430 assigned to are not certainly dereferenced in this function and thus the
3431 dereferencing cannot be safely moved to the caller without potentially
3432 introducing a segfault. Mark such REPRESENTATIVES as
3433 grp_not_necessarilly_dereferenced.
3435 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3436 part is calculated rather than simple booleans are calculated for each
3437 pointer parameter to handle cases when only a fraction of the whole
3438 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3441 The maximum dereference distances for each pointer parameter and BB are
3442 already stored in bb_dereference. This routine simply propagates these
3443 values upwards by propagate_dereference_distances and then compares the
3444 distances of individual parameters in the ENTRY BB to the equivalent
3445 distances of each representative of a (fraction of a) parameter. */
3448 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3452 if (dump_file && (dump_flags & TDF_DETAILS))
3453 dump_dereferences_table (dump_file,
3454 "Dereference table before propagation:\n",
3457 propagate_dereference_distances ();
3459 if (dump_file && (dump_flags & TDF_DETAILS))
3460 dump_dereferences_table (dump_file,
3461 "Dereference table after propagation:\n",
3464 for (i = 0; i < func_param_count; i++)
3466 struct access *repr = VEC_index (access_p, representatives, i);
3467 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3469 if (!repr || no_accesses_p (repr))
3474 if ((repr->offset + repr->size) > bb_dereferences[idx])
3475 repr->grp_not_necessarilly_dereferenced = 1;
3476 repr = repr->next_grp;
3482 /* Return the representative access for the parameter declaration PARM if it is
3483 a scalar passed by reference which is not written to and the pointer value
3484 is not used directly. Thus, if it is legal to dereference it in the caller
3485 and we can rule out modifications through aliases, such parameter should be
3486 turned into one passed by value. Return NULL otherwise. */
3488 static struct access *
3489 unmodified_by_ref_scalar_representative (tree parm)
3491 int i, access_count;
3492 struct access *repr;
3493 VEC (access_p, heap) *access_vec;
3495 access_vec = get_base_access_vector (parm);
3496 gcc_assert (access_vec);
3497 repr = VEC_index (access_p, access_vec, 0);
3500 repr->group_representative = repr;
3502 access_count = VEC_length (access_p, access_vec);
3503 for (i = 1; i < access_count; i++)
3505 struct access *access = VEC_index (access_p, access_vec, i);
3508 access->group_representative = repr;
3509 access->next_sibling = repr->next_sibling;
3510 repr->next_sibling = access;
3514 repr->grp_scalar_ptr = 1;
3518 /* Return true iff this access precludes IPA-SRA of the parameter it is
3522 access_precludes_ipa_sra_p (struct access *access)
3524 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3525 is incompatible assign in a call statement (and possibly even in asm
3526 statements). This can be relaxed by using a new temporary but only for
3527 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3528 intraprocedural SRA we deal with this by keeping the old aggregate around,
3529 something we cannot do in IPA-SRA.) */
3531 && (is_gimple_call (access->stmt)
3532 || gimple_code (access->stmt) == GIMPLE_ASM))
3539 /* Sort collected accesses for parameter PARM, identify representatives for
3540 each accessed region and link them together. Return NULL if there are
3541 different but overlapping accesses, return the special ptr value meaning
3542 there are no accesses for this parameter if that is the case and return the
3543 first representative otherwise. Set *RO_GRP if there is a group of accesses
3544 with only read (i.e. no write) accesses. */
3546 static struct access *
3547 splice_param_accesses (tree parm, bool *ro_grp)
3549 int i, j, access_count, group_count;
3550 int agg_size, total_size = 0;
3551 struct access *access, *res, **prev_acc_ptr = &res;
3552 VEC (access_p, heap) *access_vec;
3554 access_vec = get_base_access_vector (parm);
3556 return &no_accesses_representant;
3557 access_count = VEC_length (access_p, access_vec);
3559 VEC_qsort (access_p, access_vec, compare_access_positions);
3564 while (i < access_count)
3568 access = VEC_index (access_p, access_vec, i);
3569 modification = access->write;
3570 if (access_precludes_ipa_sra_p (access))
3572 a1_alias_type = reference_alias_ptr_type (access->expr);
3574 /* Access is about to become group representative unless we find some
3575 nasty overlap which would preclude us from breaking this parameter
3579 while (j < access_count)
3581 struct access *ac2 = VEC_index (access_p, access_vec, j);
3582 if (ac2->offset != access->offset)
3584 /* All or nothing law for parameters. */
3585 if (access->offset + access->size > ac2->offset)
3590 else if (ac2->size != access->size)
3593 if (access_precludes_ipa_sra_p (ac2)
3594 || (ac2->type != access->type
3595 && (TREE_ADDRESSABLE (ac2->type)
3596 || TREE_ADDRESSABLE (access->type)))
3597 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3600 modification |= ac2->write;
3601 ac2->group_representative = access;
3602 ac2->next_sibling = access->next_sibling;
3603 access->next_sibling = ac2;
3608 access->grp_maybe_modified = modification;
3611 *prev_acc_ptr = access;
3612 prev_acc_ptr = &access->next_grp;
3613 total_size += access->size;
3617 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3618 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3620 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3621 if (total_size >= agg_size)
3624 gcc_assert (group_count > 0);
3628 /* Decide whether parameters with representative accesses given by REPR should
3629 be reduced into components. */
3632 decide_one_param_reduction (struct access *repr)
3634 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3639 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3640 gcc_assert (cur_parm_size > 0);
3642 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3645 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3650 agg_size = cur_parm_size;
3656 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3657 print_generic_expr (dump_file, parm, 0);
3658 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3659 for (acc = repr; acc; acc = acc->next_grp)
3660 dump_access (dump_file, acc, true);
3664 new_param_count = 0;
3666 for (; repr; repr = repr->next_grp)
3668 gcc_assert (parm == repr->base);
3671 if (!by_ref || (!repr->grp_maybe_modified
3672 && !repr->grp_not_necessarilly_dereferenced))
3673 total_size += repr->size;
3675 total_size += cur_parm_size;
3678 gcc_assert (new_param_count > 0);
3680 if (optimize_function_for_size_p (cfun))
3681 parm_size_limit = cur_parm_size;
3683 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3686 if (total_size < agg_size
3687 && total_size <= parm_size_limit)
3690 fprintf (dump_file, " ....will be split into %i components\n",
3692 return new_param_count;
3698 /* The order of the following enums is important, we need to do extra work for
3699 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3700 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3701 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3703 /* Identify representatives of all accesses to all candidate parameters for
3704 IPA-SRA. Return result based on what representatives have been found. */
3706 static enum ipa_splicing_result
3707 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3709 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3711 struct access *repr;
3713 *representatives = VEC_alloc (access_p, heap, func_param_count);
3715 for (parm = DECL_ARGUMENTS (current_function_decl);
3717 parm = DECL_CHAIN (parm))
3719 if (is_unused_scalar_param (parm))
3721 VEC_quick_push (access_p, *representatives,
3722 &no_accesses_representant);
3723 if (result == NO_GOOD_ACCESS)
3724 result = UNUSED_PARAMS;
3726 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3727 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3728 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3730 repr = unmodified_by_ref_scalar_representative (parm);
3731 VEC_quick_push (access_p, *representatives, repr);
3733 result = UNMODIF_BY_REF_ACCESSES;
3735 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3737 bool ro_grp = false;
3738 repr = splice_param_accesses (parm, &ro_grp);
3739 VEC_quick_push (access_p, *representatives, repr);
3741 if (repr && !no_accesses_p (repr))
3743 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3746 result = UNMODIF_BY_REF_ACCESSES;
3747 else if (result < MODIF_BY_REF_ACCESSES)
3748 result = MODIF_BY_REF_ACCESSES;
3750 else if (result < BY_VAL_ACCESSES)
3751 result = BY_VAL_ACCESSES;
3753 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3754 result = UNUSED_PARAMS;
3757 VEC_quick_push (access_p, *representatives, NULL);
3760 if (result == NO_GOOD_ACCESS)
3762 VEC_free (access_p, heap, *representatives);
3763 *representatives = NULL;
3764 return NO_GOOD_ACCESS;
3770 /* Return the index of BASE in PARMS. Abort if it is not found. */
3773 get_param_index (tree base, VEC(tree, heap) *parms)
3777 len = VEC_length (tree, parms);
3778 for (i = 0; i < len; i++)
3779 if (VEC_index (tree, parms, i) == base)
3784 /* Convert the decisions made at the representative level into compact
3785 parameter adjustments. REPRESENTATIVES are pointers to first
3786 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3787 final number of adjustments. */
3789 static ipa_parm_adjustment_vec
3790 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3791 int adjustments_count)
3793 VEC (tree, heap) *parms;
3794 ipa_parm_adjustment_vec adjustments;
3798 gcc_assert (adjustments_count > 0);
3799 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3800 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3801 parm = DECL_ARGUMENTS (current_function_decl);
3802 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
3804 struct access *repr = VEC_index (access_p, representatives, i);
3806 if (!repr || no_accesses_p (repr))
3808 struct ipa_parm_adjustment *adj;
3810 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3811 memset (adj, 0, sizeof (*adj));
3812 adj->base_index = get_param_index (parm, parms);
3815 adj->copy_param = 1;
3817 adj->remove_param = 1;
3821 struct ipa_parm_adjustment *adj;
3822 int index = get_param_index (parm, parms);
3824 for (; repr; repr = repr->next_grp)
3826 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3827 memset (adj, 0, sizeof (*adj));
3828 gcc_assert (repr->base == parm);
3829 adj->base_index = index;
3830 adj->base = repr->base;
3831 adj->type = repr->type;
3832 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
3833 adj->offset = repr->offset;
3834 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3835 && (repr->grp_maybe_modified
3836 || repr->grp_not_necessarilly_dereferenced));
3841 VEC_free (tree, heap, parms);
3845 /* Analyze the collected accesses and produce a plan what to do with the
3846 parameters in the form of adjustments, NULL meaning nothing. */
3848 static ipa_parm_adjustment_vec
3849 analyze_all_param_acesses (void)
3851 enum ipa_splicing_result repr_state;
3852 bool proceed = false;
3853 int i, adjustments_count = 0;
3854 VEC (access_p, heap) *representatives;
3855 ipa_parm_adjustment_vec adjustments;
3857 repr_state = splice_all_param_accesses (&representatives);
3858 if (repr_state == NO_GOOD_ACCESS)
3861 /* If there are any parameters passed by reference which are not modified
3862 directly, we need to check whether they can be modified indirectly. */
3863 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3865 analyze_caller_dereference_legality (representatives);
3866 analyze_modified_params (representatives);
3869 for (i = 0; i < func_param_count; i++)
3871 struct access *repr = VEC_index (access_p, representatives, i);
3873 if (repr && !no_accesses_p (repr))
3875 if (repr->grp_scalar_ptr)
3877 adjustments_count++;
3878 if (repr->grp_not_necessarilly_dereferenced
3879 || repr->grp_maybe_modified)
3880 VEC_replace (access_p, representatives, i, NULL);
3884 sra_stats.scalar_by_ref_to_by_val++;
3889 int new_components = decide_one_param_reduction (repr);
3891 if (new_components == 0)
3893 VEC_replace (access_p, representatives, i, NULL);
3894 adjustments_count++;
3898 adjustments_count += new_components;
3899 sra_stats.aggregate_params_reduced++;
3900 sra_stats.param_reductions_created += new_components;
3907 if (no_accesses_p (repr))
3910 sra_stats.deleted_unused_parameters++;
3912 adjustments_count++;
3916 if (!proceed && dump_file)
3917 fprintf (dump_file, "NOT proceeding to change params.\n");
3920 adjustments = turn_representatives_into_adjustments (representatives,
3925 VEC_free (access_p, heap, representatives);
3929 /* If a parameter replacement identified by ADJ does not yet exist in the form
3930 of declaration, create it and record it, otherwise return the previously
3934 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
3937 if (!adj->new_ssa_base)
3939 char *pretty_name = make_fancy_name (adj->base);
3941 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
3942 DECL_NAME (repl) = get_identifier (pretty_name);
3943 obstack_free (&name_obstack, pretty_name);
3946 add_referenced_var (repl);
3947 adj->new_ssa_base = repl;
3950 repl = adj->new_ssa_base;
3954 /* Find the first adjustment for a particular parameter BASE in a vector of
3955 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3958 static struct ipa_parm_adjustment *
3959 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
3963 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3964 for (i = 0; i < len; i++)
3966 struct ipa_parm_adjustment *adj;
3968 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3969 if (!adj->copy_param && adj->base == base)
3976 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3977 removed because its value is not used, replace the SSA_NAME with a one
3978 relating to a created VAR_DECL together all of its uses and return true.
3979 ADJUSTMENTS is a pointer to an adjustments vector. */
3982 replace_removed_params_ssa_names (gimple stmt,
3983 ipa_parm_adjustment_vec adjustments)
3985 struct ipa_parm_adjustment *adj;
3986 tree lhs, decl, repl, name;
3988 if (gimple_code (stmt) == GIMPLE_PHI)
3989 lhs = gimple_phi_result (stmt);
3990 else if (is_gimple_assign (stmt))
3991 lhs = gimple_assign_lhs (stmt);
3992 else if (is_gimple_call (stmt))
3993 lhs = gimple_call_lhs (stmt);
3997 if (TREE_CODE (lhs) != SSA_NAME)
3999 decl = SSA_NAME_VAR (lhs);
4000 if (TREE_CODE (decl) != PARM_DECL)
4003 adj = get_adjustment_for_base (adjustments, decl);
4007 repl = get_replaced_param_substitute (adj);
4008 name = make_ssa_name (repl, stmt);
4012 fprintf (dump_file, "replacing an SSA name of a removed param ");
4013 print_generic_expr (dump_file, lhs, 0);
4014 fprintf (dump_file, " with ");
4015 print_generic_expr (dump_file, name, 0);
4016 fprintf (dump_file, "\n");
4019 if (is_gimple_assign (stmt))
4020 gimple_assign_set_lhs (stmt, name);
4021 else if (is_gimple_call (stmt))
4022 gimple_call_set_lhs (stmt, name);
4024 gimple_phi_set_result (stmt, name);
4026 replace_uses_by (lhs, name);
4027 release_ssa_name (lhs);
4031 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4032 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4033 specifies whether the function should care about type incompatibility the
4034 current and new expressions. If it is false, the function will leave
4035 incompatibility issues to the caller. Return true iff the expression
4039 sra_ipa_modify_expr (tree *expr, bool convert,
4040 ipa_parm_adjustment_vec adjustments)
4043 struct ipa_parm_adjustment *adj, *cand = NULL;
4044 HOST_WIDE_INT offset, size, max_size;
4047 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4049 if (TREE_CODE (*expr) == BIT_FIELD_REF
4050 || TREE_CODE (*expr) == IMAGPART_EXPR
4051 || TREE_CODE (*expr) == REALPART_EXPR)
4053 expr = &TREE_OPERAND (*expr, 0);
4057 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4058 if (!base || size == -1 || max_size == -1)
4061 if (TREE_CODE (base) == MEM_REF)
4063 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4064 base = TREE_OPERAND (base, 0);
4067 base = get_ssa_base_param (base);
4068 if (!base || TREE_CODE (base) != PARM_DECL)
4071 for (i = 0; i < len; i++)
4073 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4075 if (adj->base == base &&
4076 (adj->offset == offset || adj->remove_param))
4082 if (!cand || cand->copy_param || cand->remove_param)
4086 src = build_simple_mem_ref (cand->reduction);
4088 src = cand->reduction;
4090 if (dump_file && (dump_flags & TDF_DETAILS))
4092 fprintf (dump_file, "About to replace expr ");
4093 print_generic_expr (dump_file, *expr, 0);
4094 fprintf (dump_file, " with ");
4095 print_generic_expr (dump_file, src, 0);
4096 fprintf (dump_file, "\n");
4099 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4101 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4109 /* If the statement pointed to by STMT_PTR contains any expressions that need
4110 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4111 potential type incompatibilities (GSI is used to accommodate conversion
4112 statements and must point to the statement). Return true iff the statement
4116 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4117 ipa_parm_adjustment_vec adjustments)
4119 gimple stmt = *stmt_ptr;
4120 tree *lhs_p, *rhs_p;
4123 if (!gimple_assign_single_p (stmt))
4126 rhs_p = gimple_assign_rhs1_ptr (stmt);
4127 lhs_p = gimple_assign_lhs_ptr (stmt);
4129 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4130 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4133 tree new_rhs = NULL_TREE;
4135 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4137 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4139 /* V_C_Es of constructors can cause trouble (PR 42714). */
4140 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4141 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4143 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4146 new_rhs = fold_build1_loc (gimple_location (stmt),
4147 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4150 else if (REFERENCE_CLASS_P (*rhs_p)
4151 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4152 && !is_gimple_reg (*lhs_p))
4153 /* This can happen when an assignment in between two single field
4154 structures is turned into an assignment in between two pointers to
4155 scalars (PR 42237). */
4160 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4161 true, GSI_SAME_STMT);
4163 gimple_assign_set_rhs_from_tree (gsi, tmp);
4172 /* Traverse the function body and all modifications as described in
4173 ADJUSTMENTS. Return true iff the CFG has been changed. */
4176 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4178 bool cfg_changed = false;
4183 gimple_stmt_iterator gsi;
4185 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4186 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4188 gsi = gsi_start_bb (bb);
4189 while (!gsi_end_p (gsi))
4191 gimple stmt = gsi_stmt (gsi);
4192 bool modified = false;
4196 switch (gimple_code (stmt))
4199 t = gimple_return_retval_ptr (stmt);
4200 if (*t != NULL_TREE)
4201 modified |= sra_ipa_modify_expr (t, true, adjustments);
4205 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4206 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4210 /* Operands must be processed before the lhs. */
4211 for (i = 0; i < gimple_call_num_args (stmt); i++)
4213 t = gimple_call_arg_ptr (stmt, i);
4214 modified |= sra_ipa_modify_expr (t, true, adjustments);
4217 if (gimple_call_lhs (stmt))
4219 t = gimple_call_lhs_ptr (stmt);
4220 modified |= sra_ipa_modify_expr (t, false, adjustments);
4221 modified |= replace_removed_params_ssa_names (stmt,
4227 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4229 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4230 modified |= sra_ipa_modify_expr (t, true, adjustments);
4232 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4234 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4235 modified |= sra_ipa_modify_expr (t, false, adjustments);
4246 if (maybe_clean_eh_stmt (stmt)
4247 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4257 /* Call gimple_debug_bind_reset_value on all debug statements describing
4258 gimple register parameters that are being removed or replaced. */
4261 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4265 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4266 for (i = 0; i < len; i++)
4268 struct ipa_parm_adjustment *adj;
4269 imm_use_iterator ui;
4273 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4274 if (adj->copy_param || !is_gimple_reg (adj->base))
4276 name = gimple_default_def (cfun, adj->base);
4279 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4281 /* All other users must have been removed by
4282 ipa_sra_modify_function_body. */
4283 gcc_assert (is_gimple_debug (stmt));
4284 gimple_debug_bind_reset_value (stmt);
4290 /* Return true iff all callers have at least as many actual arguments as there
4291 are formal parameters in the current function. */
4294 all_callers_have_enough_arguments_p (struct cgraph_node *node)
4296 struct cgraph_edge *cs;
4297 for (cs = node->callers; cs; cs = cs->next_caller)
4298 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4305 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4308 convert_callers (struct cgraph_node *node, tree old_decl,
4309 ipa_parm_adjustment_vec adjustments)
4311 tree old_cur_fndecl = current_function_decl;
4312 struct cgraph_edge *cs;
4313 basic_block this_block;
4314 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4316 for (cs = node->callers; cs; cs = cs->next_caller)
4318 current_function_decl = cs->caller->decl;
4319 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4322 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4323 cs->caller->uid, cs->callee->uid,
4324 cgraph_node_name (cs->caller),
4325 cgraph_node_name (cs->callee));
4327 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4332 for (cs = node->callers; cs; cs = cs->next_caller)
4333 if (bitmap_set_bit (recomputed_callers, cs->caller->uid))
4334 compute_inline_parameters (cs->caller);
4335 BITMAP_FREE (recomputed_callers);
4337 current_function_decl = old_cur_fndecl;
4339 if (!encountered_recursive_call)
4342 FOR_EACH_BB (this_block)
4344 gimple_stmt_iterator gsi;
4346 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4348 gimple stmt = gsi_stmt (gsi);
4350 if (gimple_code (stmt) != GIMPLE_CALL)
4352 call_fndecl = gimple_call_fndecl (stmt);
4353 if (call_fndecl == old_decl)
4356 fprintf (dump_file, "Adjusting recursive call");
4357 gimple_call_set_fndecl (stmt, node->decl);
4358 ipa_modify_call_arguments (NULL, stmt, adjustments);
4366 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4367 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4370 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4372 struct cgraph_node *new_node;
4373 struct cgraph_edge *cs;
4375 VEC (cgraph_edge_p, heap) * redirect_callers;
4379 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
4381 redirect_callers = VEC_alloc (cgraph_edge_p, heap, node_callers);
4382 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
4383 VEC_quick_push (cgraph_edge_p, redirect_callers, cs);
4385 rebuild_cgraph_edges ();
4387 current_function_decl = NULL_TREE;
4389 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4390 NULL, NULL, "isra");
4391 current_function_decl = new_node->decl;
4392 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4394 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4395 cfg_changed = ipa_sra_modify_function_body (adjustments);
4396 sra_ipa_reset_debug_stmts (adjustments);
4397 convert_callers (new_node, node->decl, adjustments);
4398 cgraph_make_node_local (new_node);
4402 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4403 attributes, return true otherwise. NODE is the cgraph node of the current
4407 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4409 if (!cgraph_node_can_be_local_p (node))
4412 fprintf (dump_file, "Function not local to this compilation unit.\n");
4416 if (!tree_versionable_function_p (node->decl))
4419 fprintf (dump_file, "Function is not versionable.\n");
4423 if (DECL_VIRTUAL_P (current_function_decl))
4426 fprintf (dump_file, "Function is a virtual method.\n");
4430 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4431 && node->global.size >= MAX_INLINE_INSNS_AUTO)
4434 fprintf (dump_file, "Function too big to be made truly local.\n");
4442 "Function has no callers in this compilation unit.\n");
4449 fprintf (dump_file, "Function uses stdarg. \n");
4453 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4459 /* Perform early interprocedural SRA. */
4462 ipa_early_sra (void)
4464 struct cgraph_node *node = cgraph_node (current_function_decl);
4465 ipa_parm_adjustment_vec adjustments;
4468 if (!ipa_sra_preliminary_function_checks (node))
4472 sra_mode = SRA_MODE_EARLY_IPA;
4474 if (!find_param_candidates ())
4477 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4481 if (!all_callers_have_enough_arguments_p (node))
4484 fprintf (dump_file, "There are callers with insufficient number of "
4489 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4491 * last_basic_block_for_function (cfun));
4492 final_bbs = BITMAP_ALLOC (NULL);
4495 if (encountered_apply_args)
4498 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4502 if (encountered_unchangable_recursive_call)
4505 fprintf (dump_file, "Function calls itself with insufficient "
4506 "number of arguments.\n");
4510 adjustments = analyze_all_param_acesses ();
4514 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4516 if (modify_function (node, adjustments))
4517 ret = TODO_update_ssa | TODO_cleanup_cfg;
4519 ret = TODO_update_ssa;
4520 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4522 statistics_counter_event (cfun, "Unused parameters deleted",
4523 sra_stats.deleted_unused_parameters);
4524 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4525 sra_stats.scalar_by_ref_to_by_val);
4526 statistics_counter_event (cfun, "Aggregate parameters broken up",
4527 sra_stats.aggregate_params_reduced);
4528 statistics_counter_event (cfun, "Aggregate parameter components created",
4529 sra_stats.param_reductions_created);
4532 BITMAP_FREE (final_bbs);
4533 free (bb_dereferences);
4535 sra_deinitialize ();
4539 /* Return if early ipa sra shall be performed. */
4541 ipa_early_sra_gate (void)
4543 return flag_ipa_sra && dbg_cnt (eipa_sra);
4546 struct gimple_opt_pass pass_early_ipa_sra =
4550 "eipa_sra", /* name */
4551 ipa_early_sra_gate, /* gate */
4552 ipa_early_sra, /* execute */
4555 0, /* static_pass_number */
4556 TV_IPA_SRA, /* tv_id */
4557 0, /* properties_required */
4558 0, /* properties_provided */
4559 0, /* properties_destroyed */
4560 0, /* todo_flags_start */
4561 TODO_dump_func | TODO_dump_cgraph /* todo_flags_finish */