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"
83 #include "tree-flow.h"
85 #include "tree-pretty-print.h"
86 #include "statistics.h"
87 #include "tree-dump.h"
93 #include "tree-inline.h"
94 #include "gimple-pretty-print.h"
96 /* Enumeration of all aggregate reductions we can do. */
97 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
98 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
99 SRA_MODE_INTRA }; /* late intraprocedural SRA */
101 /* Global variable describing which aggregate reduction we are performing at
103 static enum sra_mode sra_mode;
107 /* ACCESS represents each access to an aggregate variable (as a whole or a
108 part). It can also represent a group of accesses that refer to exactly the
109 same fragment of an aggregate (i.e. those that have exactly the same offset
110 and size). Such representatives for a single aggregate, once determined,
111 are linked in a linked list and have the group fields set.
113 Moreover, when doing intraprocedural SRA, a tree is built from those
114 representatives (by the means of first_child and next_sibling pointers), in
115 which all items in a subtree are "within" the root, i.e. their offset is
116 greater or equal to offset of the root and offset+size is smaller or equal
117 to offset+size of the root. Children of an access are sorted by offset.
119 Note that accesses to parts of vector and complex number types always
120 represented by an access to the whole complex number or a vector. It is a
121 duty of the modifying functions to replace them appropriately. */
125 /* Values returned by `get_ref_base_and_extent' for each component reference
126 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
127 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
128 HOST_WIDE_INT offset;
132 /* Expression. It is context dependent so do not use it to create new
133 expressions to access the original aggregate. See PR 42154 for a
139 /* The statement this access belongs to. */
142 /* Next group representative for this aggregate. */
143 struct access *next_grp;
145 /* Pointer to the group representative. Pointer to itself if the struct is
146 the representative. */
147 struct access *group_representative;
149 /* If this access has any children (in terms of the definition above), this
150 points to the first one. */
151 struct access *first_child;
153 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
154 described above. In IPA-SRA this is a pointer to the next access
155 belonging to the same group (having the same representative). */
156 struct access *next_sibling;
158 /* Pointers to the first and last element in the linked list of assign
160 struct assign_link *first_link, *last_link;
162 /* Pointer to the next access in the work queue. */
163 struct access *next_queued;
165 /* Replacement variable for this access "region." Never to be accessed
166 directly, always only by the means of get_access_replacement() and only
167 when grp_to_be_replaced flag is set. */
168 tree replacement_decl;
170 /* Is this particular access write access? */
173 /* Is this access an artificial one created to scalarize some record
175 unsigned total_scalarization : 1;
177 /* Is this access currently in the work queue? */
178 unsigned grp_queued : 1;
180 /* Does this group contain a write access? This flag is propagated down the
182 unsigned grp_write : 1;
184 /* Does this group contain a read access? This flag is propagated down the
186 unsigned grp_read : 1;
188 /* Does this group contain a read access that comes from an assignment
189 statement? This flag is propagated down the access tree. */
190 unsigned grp_assignment_read : 1;
192 /* Other passes of the analysis use this bit to make function
193 analyze_access_subtree create scalar replacements for this group if
195 unsigned grp_hint : 1;
197 /* Is the subtree rooted in this access fully covered by scalar
199 unsigned grp_covered : 1;
201 /* If set to true, this access and all below it in an access tree must not be
203 unsigned grp_unscalarizable_region : 1;
205 /* Whether data have been written to parts of the aggregate covered by this
206 access which is not to be scalarized. This flag is propagated up in the
208 unsigned grp_unscalarized_data : 1;
210 /* Does this access and/or group contain a write access through a
212 unsigned grp_partial_lhs : 1;
214 /* Set when a scalar replacement should be created for this variable. We do
215 the decision and creation at different places because create_tmp_var
216 cannot be called from within FOR_EACH_REFERENCED_VAR. */
217 unsigned grp_to_be_replaced : 1;
219 /* Is it possible that the group refers to data which might be (directly or
220 otherwise) modified? */
221 unsigned grp_maybe_modified : 1;
223 /* Set when this is a representative of a pointer to scalar (i.e. by
224 reference) parameter which we consider for turning into a plain scalar
225 (i.e. a by value parameter). */
226 unsigned grp_scalar_ptr : 1;
228 /* Set when we discover that this pointer is not safe to dereference in the
230 unsigned grp_not_necessarilly_dereferenced : 1;
233 typedef struct access *access_p;
235 DEF_VEC_P (access_p);
236 DEF_VEC_ALLOC_P (access_p, heap);
238 /* Alloc pool for allocating access structures. */
239 static alloc_pool access_pool;
241 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
242 are used to propagate subaccesses from rhs to lhs as long as they don't
243 conflict with what is already there. */
246 struct access *lacc, *racc;
247 struct assign_link *next;
250 /* Alloc pool for allocating assign link structures. */
251 static alloc_pool link_pool;
253 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
254 static struct pointer_map_t *base_access_vec;
256 /* Bitmap of candidates. */
257 static bitmap candidate_bitmap;
259 /* Bitmap of candidates which we should try to entirely scalarize away and
260 those which cannot be (because they are and need be used as a whole). */
261 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
263 /* Obstack for creation of fancy names. */
264 static struct obstack name_obstack;
266 /* Head of a linked list of accesses that need to have its subaccesses
267 propagated to their assignment counterparts. */
268 static struct access *work_queue_head;
270 /* Number of parameters of the analyzed function when doing early ipa SRA. */
271 static int func_param_count;
273 /* scan_function sets the following to true if it encounters a call to
274 __builtin_apply_args. */
275 static bool encountered_apply_args;
277 /* Set by scan_function when it finds a recursive call. */
278 static bool encountered_recursive_call;
280 /* Set by scan_function when it finds a recursive call with less actual
281 arguments than formal parameters.. */
282 static bool encountered_unchangable_recursive_call;
284 /* This is a table in which for each basic block and parameter there is a
285 distance (offset + size) in that parameter which is dereferenced and
286 accessed in that BB. */
287 static HOST_WIDE_INT *bb_dereferences;
288 /* Bitmap of BBs that can cause the function to "stop" progressing by
289 returning, throwing externally, looping infinitely or calling a function
290 which might abort etc.. */
291 static bitmap final_bbs;
293 /* Representative of no accesses at all. */
294 static struct access no_accesses_representant;
296 /* Predicate to test the special value. */
299 no_accesses_p (struct access *access)
301 return access == &no_accesses_representant;
304 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
305 representative fields are dumped, otherwise those which only describe the
306 individual access are. */
310 /* Number of processed aggregates is readily available in
311 analyze_all_variable_accesses and so is not stored here. */
313 /* Number of created scalar replacements. */
316 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
320 /* Number of statements created by generate_subtree_copies. */
323 /* Number of statements created by load_assign_lhs_subreplacements. */
326 /* Number of times sra_modify_assign has deleted a statement. */
329 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
330 RHS reparately due to type conversions or nonexistent matching
332 int separate_lhs_rhs_handling;
334 /* Number of parameters that were removed because they were unused. */
335 int deleted_unused_parameters;
337 /* Number of scalars passed as parameters by reference that have been
338 converted to be passed by value. */
339 int scalar_by_ref_to_by_val;
341 /* Number of aggregate parameters that were replaced by one or more of their
343 int aggregate_params_reduced;
345 /* Numbber of components created when splitting aggregate parameters. */
346 int param_reductions_created;
350 dump_access (FILE *f, struct access *access, bool grp)
352 fprintf (f, "access { ");
353 fprintf (f, "base = (%d)'", DECL_UID (access->base));
354 print_generic_expr (f, access->base, 0);
355 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
356 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
357 fprintf (f, ", expr = ");
358 print_generic_expr (f, access->expr, 0);
359 fprintf (f, ", type = ");
360 print_generic_expr (f, access->type, 0);
362 fprintf (f, ", grp_write = %d, total_scalarization = %d, "
363 "grp_read = %d, grp_hint = %d, grp_assignment_read = %d,"
364 "grp_covered = %d, grp_unscalarizable_region = %d, "
365 "grp_unscalarized_data = %d, grp_partial_lhs = %d, "
366 "grp_to_be_replaced = %d, grp_maybe_modified = %d, "
367 "grp_not_necessarilly_dereferenced = %d\n",
368 access->grp_write, access->total_scalarization,
369 access->grp_read, access->grp_hint, access->grp_assignment_read,
370 access->grp_covered, access->grp_unscalarizable_region,
371 access->grp_unscalarized_data, access->grp_partial_lhs,
372 access->grp_to_be_replaced, access->grp_maybe_modified,
373 access->grp_not_necessarilly_dereferenced);
375 fprintf (f, ", write = %d, total_scalarization = %d, "
376 "grp_partial_lhs = %d\n",
377 access->write, access->total_scalarization,
378 access->grp_partial_lhs);
381 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
384 dump_access_tree_1 (FILE *f, struct access *access, int level)
390 for (i = 0; i < level; i++)
391 fputs ("* ", dump_file);
393 dump_access (f, access, true);
395 if (access->first_child)
396 dump_access_tree_1 (f, access->first_child, level + 1);
398 access = access->next_sibling;
403 /* Dump all access trees for a variable, given the pointer to the first root in
407 dump_access_tree (FILE *f, struct access *access)
409 for (; access; access = access->next_grp)
410 dump_access_tree_1 (f, access, 0);
413 /* Return true iff ACC is non-NULL and has subaccesses. */
416 access_has_children_p (struct access *acc)
418 return acc && acc->first_child;
421 /* Return a vector of pointers to accesses for the variable given in BASE or
422 NULL if there is none. */
424 static VEC (access_p, heap) *
425 get_base_access_vector (tree base)
429 slot = pointer_map_contains (base_access_vec, base);
433 return *(VEC (access_p, heap) **) slot;
436 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
437 in ACCESS. Return NULL if it cannot be found. */
439 static struct access *
440 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
443 while (access && (access->offset != offset || access->size != size))
445 struct access *child = access->first_child;
447 while (child && (child->offset + child->size <= offset))
448 child = child->next_sibling;
455 /* Return the first group representative for DECL or NULL if none exists. */
457 static struct access *
458 get_first_repr_for_decl (tree base)
460 VEC (access_p, heap) *access_vec;
462 access_vec = get_base_access_vector (base);
466 return VEC_index (access_p, access_vec, 0);
469 /* Find an access representative for the variable BASE and given OFFSET and
470 SIZE. Requires that access trees have already been built. Return NULL if
471 it cannot be found. */
473 static struct access *
474 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
477 struct access *access;
479 access = get_first_repr_for_decl (base);
480 while (access && (access->offset + access->size <= offset))
481 access = access->next_grp;
485 return find_access_in_subtree (access, offset, size);
488 /* Add LINK to the linked list of assign links of RACC. */
490 add_link_to_rhs (struct access *racc, struct assign_link *link)
492 gcc_assert (link->racc == racc);
494 if (!racc->first_link)
496 gcc_assert (!racc->last_link);
497 racc->first_link = link;
500 racc->last_link->next = link;
502 racc->last_link = link;
506 /* Move all link structures in their linked list in OLD_RACC to the linked list
509 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
511 if (!old_racc->first_link)
513 gcc_assert (!old_racc->last_link);
517 if (new_racc->first_link)
519 gcc_assert (!new_racc->last_link->next);
520 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
522 new_racc->last_link->next = old_racc->first_link;
523 new_racc->last_link = old_racc->last_link;
527 gcc_assert (!new_racc->last_link);
529 new_racc->first_link = old_racc->first_link;
530 new_racc->last_link = old_racc->last_link;
532 old_racc->first_link = old_racc->last_link = NULL;
535 /* Add ACCESS to the work queue (which is actually a stack). */
538 add_access_to_work_queue (struct access *access)
540 if (!access->grp_queued)
542 gcc_assert (!access->next_queued);
543 access->next_queued = work_queue_head;
544 access->grp_queued = 1;
545 work_queue_head = access;
549 /* Pop an access from the work queue, and return it, assuming there is one. */
551 static struct access *
552 pop_access_from_work_queue (void)
554 struct access *access = work_queue_head;
556 work_queue_head = access->next_queued;
557 access->next_queued = NULL;
558 access->grp_queued = 0;
563 /* Allocate necessary structures. */
566 sra_initialize (void)
568 candidate_bitmap = BITMAP_ALLOC (NULL);
569 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
570 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
571 gcc_obstack_init (&name_obstack);
572 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
573 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
574 base_access_vec = pointer_map_create ();
575 memset (&sra_stats, 0, sizeof (sra_stats));
576 encountered_apply_args = false;
577 encountered_recursive_call = false;
578 encountered_unchangable_recursive_call = false;
581 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
584 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
585 void *data ATTRIBUTE_UNUSED)
587 VEC (access_p, heap) *access_vec;
588 access_vec = (VEC (access_p, heap) *) *value;
589 VEC_free (access_p, heap, access_vec);
594 /* Deallocate all general structures. */
597 sra_deinitialize (void)
599 BITMAP_FREE (candidate_bitmap);
600 BITMAP_FREE (should_scalarize_away_bitmap);
601 BITMAP_FREE (cannot_scalarize_away_bitmap);
602 free_alloc_pool (access_pool);
603 free_alloc_pool (link_pool);
604 obstack_free (&name_obstack, NULL);
606 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
607 pointer_map_destroy (base_access_vec);
610 /* Remove DECL from candidates for SRA and write REASON to the dump file if
613 disqualify_candidate (tree decl, const char *reason)
615 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
617 if (dump_file && (dump_flags & TDF_DETAILS))
619 fprintf (dump_file, "! Disqualifying ");
620 print_generic_expr (dump_file, decl, 0);
621 fprintf (dump_file, " - %s\n", reason);
625 /* Return true iff the type contains a field or an element which does not allow
629 type_internals_preclude_sra_p (tree type)
634 switch (TREE_CODE (type))
638 case QUAL_UNION_TYPE:
639 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
640 if (TREE_CODE (fld) == FIELD_DECL)
642 tree ft = TREE_TYPE (fld);
644 if (TREE_THIS_VOLATILE (fld)
645 || !DECL_FIELD_OFFSET (fld) || !DECL_SIZE (fld)
646 || !host_integerp (DECL_FIELD_OFFSET (fld), 1)
647 || !host_integerp (DECL_SIZE (fld), 1))
650 if (AGGREGATE_TYPE_P (ft)
651 && type_internals_preclude_sra_p (ft))
658 et = TREE_TYPE (type);
660 if (AGGREGATE_TYPE_P (et))
661 return type_internals_preclude_sra_p (et);
670 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
671 base variable if it is. Return T if it is not an SSA_NAME. */
674 get_ssa_base_param (tree t)
676 if (TREE_CODE (t) == SSA_NAME)
678 if (SSA_NAME_IS_DEFAULT_DEF (t))
679 return SSA_NAME_VAR (t);
686 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
687 belongs to, unless the BB has already been marked as a potentially
691 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
693 basic_block bb = gimple_bb (stmt);
694 int idx, parm_index = 0;
697 if (bitmap_bit_p (final_bbs, bb->index))
700 for (parm = DECL_ARGUMENTS (current_function_decl);
701 parm && parm != base;
702 parm = DECL_CHAIN (parm))
705 gcc_assert (parm_index < func_param_count);
707 idx = bb->index * func_param_count + parm_index;
708 if (bb_dereferences[idx] < dist)
709 bb_dereferences[idx] = dist;
712 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
713 the three fields. Also add it to the vector of accesses corresponding to
714 the base. Finally, return the new access. */
716 static struct access *
717 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
719 VEC (access_p, heap) *vec;
720 struct access *access;
723 access = (struct access *) pool_alloc (access_pool);
724 memset (access, 0, sizeof (struct access));
726 access->offset = offset;
729 slot = pointer_map_contains (base_access_vec, base);
731 vec = (VEC (access_p, heap) *) *slot;
733 vec = VEC_alloc (access_p, heap, 32);
735 VEC_safe_push (access_p, heap, vec, access);
737 *((struct VEC (access_p,heap) **)
738 pointer_map_insert (base_access_vec, base)) = vec;
743 /* Create and insert access for EXPR. Return created access, or NULL if it is
746 static struct access *
747 create_access (tree expr, gimple stmt, bool write)
749 struct access *access;
750 HOST_WIDE_INT offset, size, max_size;
752 bool ptr, unscalarizable_region = false;
754 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
756 if (sra_mode == SRA_MODE_EARLY_IPA
757 && TREE_CODE (base) == MEM_REF)
759 base = get_ssa_base_param (TREE_OPERAND (base, 0));
767 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
770 if (sra_mode == SRA_MODE_EARLY_IPA)
772 if (size < 0 || size != max_size)
774 disqualify_candidate (base, "Encountered a variable sized access.");
777 if ((offset % BITS_PER_UNIT) != 0 || (size % BITS_PER_UNIT) != 0)
779 disqualify_candidate (base,
780 "Encountered an acces not aligned to a byte.");
785 mark_parm_dereference (base, offset + size, stmt);
789 if (size != max_size)
792 unscalarizable_region = true;
796 disqualify_candidate (base, "Encountered an unconstrained access.");
801 access = create_access_1 (base, offset, size);
803 access->type = TREE_TYPE (expr);
804 access->write = write;
805 access->grp_unscalarizable_region = unscalarizable_region;
812 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
813 register types or (recursively) records with only these two kinds of fields.
814 It also returns false if any of these records has a zero-size field as its
815 last field or has a bit-field. */
818 type_consists_of_records_p (tree type)
821 bool last_fld_has_zero_size = false;
823 if (TREE_CODE (type) != RECORD_TYPE)
826 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
827 if (TREE_CODE (fld) == FIELD_DECL)
829 tree ft = TREE_TYPE (fld);
831 if (DECL_BIT_FIELD (fld))
834 if (!is_gimple_reg_type (ft)
835 && !type_consists_of_records_p (ft))
838 last_fld_has_zero_size = tree_low_cst (DECL_SIZE (fld), 1) == 0;
841 if (last_fld_has_zero_size)
847 /* Create total_scalarization accesses for all scalar type fields in DECL that
848 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
849 must be the top-most VAR_DECL representing the variable, OFFSET must be the
850 offset of DECL within BASE. REF must be the memory reference expression for
854 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
857 tree fld, decl_type = TREE_TYPE (decl);
859 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
860 if (TREE_CODE (fld) == FIELD_DECL)
862 HOST_WIDE_INT pos = offset + int_bit_position (fld);
863 tree ft = TREE_TYPE (fld);
864 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
867 if (is_gimple_reg_type (ft))
869 struct access *access;
872 size = tree_low_cst (DECL_SIZE (fld), 1);
873 access = create_access_1 (base, pos, size);
876 access->total_scalarization = 1;
877 /* Accesses for intraprocedural SRA can have their stmt NULL. */
880 completely_scalarize_record (base, fld, pos, nref);
885 /* Search the given tree for a declaration by skipping handled components and
886 exclude it from the candidates. */
889 disqualify_base_of_expr (tree t, const char *reason)
891 t = get_base_address (t);
892 if (sra_mode == SRA_MODE_EARLY_IPA
893 && TREE_CODE (t) == MEM_REF)
894 t = get_ssa_base_param (TREE_OPERAND (t, 0));
897 disqualify_candidate (t, reason);
900 /* Scan expression EXPR and create access structures for all accesses to
901 candidates for scalarization. Return the created access or NULL if none is
904 static struct access *
905 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
907 struct access *ret = NULL;
910 if (TREE_CODE (expr) == BIT_FIELD_REF
911 || TREE_CODE (expr) == IMAGPART_EXPR
912 || TREE_CODE (expr) == REALPART_EXPR)
914 expr = TREE_OPERAND (expr, 0);
920 /* We need to dive through V_C_Es in order to get the size of its parameter
921 and not the result type. Ada produces such statements. We are also
922 capable of handling the topmost V_C_E but not any of those buried in other
923 handled components. */
924 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
925 expr = TREE_OPERAND (expr, 0);
927 if (contains_view_convert_expr_p (expr))
929 disqualify_base_of_expr (expr, "V_C_E under a different handled "
934 switch (TREE_CODE (expr))
937 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
938 && sra_mode != SRA_MODE_EARLY_IPA)
946 case ARRAY_RANGE_REF:
947 ret = create_access (expr, stmt, write);
954 if (write && partial_ref && ret)
955 ret->grp_partial_lhs = 1;
960 /* Scan expression EXPR and create access structures for all accesses to
961 candidates for scalarization. Return true if any access has been inserted.
962 STMT must be the statement from which the expression is taken, WRITE must be
963 true if the expression is a store and false otherwise. */
966 build_access_from_expr (tree expr, gimple stmt, bool write)
968 struct access *access;
970 access = build_access_from_expr_1 (expr, stmt, write);
973 /* This means the aggregate is accesses as a whole in a way other than an
974 assign statement and thus cannot be removed even if we had a scalar
975 replacement for everything. */
976 if (cannot_scalarize_away_bitmap)
977 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
983 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
984 modes in which it matters, return true iff they have been disqualified. RHS
985 may be NULL, in that case ignore it. If we scalarize an aggregate in
986 intra-SRA we may need to add statements after each statement. This is not
987 possible if a statement unconditionally has to end the basic block. */
989 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
991 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
992 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
994 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
996 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1002 /* Scan expressions occuring in STMT, create access structures for all accesses
1003 to candidates for scalarization and remove those candidates which occur in
1004 statements or expressions that prevent them from being split apart. Return
1005 true if any access has been inserted. */
1008 build_accesses_from_assign (gimple stmt)
1011 struct access *lacc, *racc;
1013 if (!gimple_assign_single_p (stmt))
1016 lhs = gimple_assign_lhs (stmt);
1017 rhs = gimple_assign_rhs1 (stmt);
1019 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1022 racc = build_access_from_expr_1 (rhs, stmt, false);
1023 lacc = build_access_from_expr_1 (lhs, stmt, true);
1027 racc->grp_assignment_read = 1;
1028 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1029 && !is_gimple_reg_type (racc->type))
1030 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1034 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1035 && !lacc->grp_unscalarizable_region
1036 && !racc->grp_unscalarizable_region
1037 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1038 /* FIXME: Turn the following line into an assert after PR 40058 is
1040 && lacc->size == racc->size
1041 && useless_type_conversion_p (lacc->type, racc->type))
1043 struct assign_link *link;
1045 link = (struct assign_link *) pool_alloc (link_pool);
1046 memset (link, 0, sizeof (struct assign_link));
1051 add_link_to_rhs (racc, link);
1054 return lacc || racc;
1057 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1058 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1061 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1062 void *data ATTRIBUTE_UNUSED)
1064 op = get_base_address (op);
1067 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1072 /* Return true iff callsite CALL has at least as many actual arguments as there
1073 are formal parameters of the function currently processed by IPA-SRA. */
1076 callsite_has_enough_arguments_p (gimple call)
1078 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1081 /* Scan function and look for interesting expressions and create access
1082 structures for them. Return true iff any access is created. */
1085 scan_function (void)
1092 gimple_stmt_iterator gsi;
1093 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1095 gimple stmt = gsi_stmt (gsi);
1099 if (final_bbs && stmt_can_throw_external (stmt))
1100 bitmap_set_bit (final_bbs, bb->index);
1101 switch (gimple_code (stmt))
1104 t = gimple_return_retval (stmt);
1106 ret |= build_access_from_expr (t, stmt, false);
1108 bitmap_set_bit (final_bbs, bb->index);
1112 ret |= build_accesses_from_assign (stmt);
1116 for (i = 0; i < gimple_call_num_args (stmt); i++)
1117 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1120 if (sra_mode == SRA_MODE_EARLY_IPA)
1122 tree dest = gimple_call_fndecl (stmt);
1123 int flags = gimple_call_flags (stmt);
1127 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1128 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1129 encountered_apply_args = true;
1130 if (cgraph_get_node (dest)
1131 == cgraph_get_node (current_function_decl))
1133 encountered_recursive_call = true;
1134 if (!callsite_has_enough_arguments_p (stmt))
1135 encountered_unchangable_recursive_call = true;
1140 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1141 bitmap_set_bit (final_bbs, bb->index);
1144 t = gimple_call_lhs (stmt);
1145 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1146 ret |= build_access_from_expr (t, stmt, true);
1150 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1153 bitmap_set_bit (final_bbs, bb->index);
1155 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1157 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1158 ret |= build_access_from_expr (t, stmt, false);
1160 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1162 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1163 ret |= build_access_from_expr (t, stmt, true);
1176 /* Helper of QSORT function. There are pointers to accesses in the array. An
1177 access is considered smaller than another if it has smaller offset or if the
1178 offsets are the same but is size is bigger. */
1181 compare_access_positions (const void *a, const void *b)
1183 const access_p *fp1 = (const access_p *) a;
1184 const access_p *fp2 = (const access_p *) b;
1185 const access_p f1 = *fp1;
1186 const access_p f2 = *fp2;
1188 if (f1->offset != f2->offset)
1189 return f1->offset < f2->offset ? -1 : 1;
1191 if (f1->size == f2->size)
1193 if (f1->type == f2->type)
1195 /* Put any non-aggregate type before any aggregate type. */
1196 else if (!is_gimple_reg_type (f1->type)
1197 && is_gimple_reg_type (f2->type))
1199 else if (is_gimple_reg_type (f1->type)
1200 && !is_gimple_reg_type (f2->type))
1202 /* Put any complex or vector type before any other scalar type. */
1203 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1204 && TREE_CODE (f1->type) != VECTOR_TYPE
1205 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1206 || TREE_CODE (f2->type) == VECTOR_TYPE))
1208 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1209 || TREE_CODE (f1->type) == VECTOR_TYPE)
1210 && TREE_CODE (f2->type) != COMPLEX_TYPE
1211 && TREE_CODE (f2->type) != VECTOR_TYPE)
1213 /* Put the integral type with the bigger precision first. */
1214 else if (INTEGRAL_TYPE_P (f1->type)
1215 && INTEGRAL_TYPE_P (f2->type))
1216 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1217 /* Put any integral type with non-full precision last. */
1218 else if (INTEGRAL_TYPE_P (f1->type)
1219 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1220 != TYPE_PRECISION (f1->type)))
1222 else if (INTEGRAL_TYPE_P (f2->type)
1223 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1224 != TYPE_PRECISION (f2->type)))
1226 /* Stabilize the sort. */
1227 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1230 /* We want the bigger accesses first, thus the opposite operator in the next
1232 return f1->size > f2->size ? -1 : 1;
1236 /* Append a name of the declaration to the name obstack. A helper function for
1240 make_fancy_decl_name (tree decl)
1244 tree name = DECL_NAME (decl);
1246 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1247 IDENTIFIER_LENGTH (name));
1250 sprintf (buffer, "D%u", DECL_UID (decl));
1251 obstack_grow (&name_obstack, buffer, strlen (buffer));
1255 /* Helper for make_fancy_name. */
1258 make_fancy_name_1 (tree expr)
1265 make_fancy_decl_name (expr);
1269 switch (TREE_CODE (expr))
1272 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1273 obstack_1grow (&name_obstack, '$');
1274 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1278 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1279 obstack_1grow (&name_obstack, '$');
1280 /* Arrays with only one element may not have a constant as their
1282 index = TREE_OPERAND (expr, 1);
1283 if (TREE_CODE (index) != INTEGER_CST)
1285 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1286 obstack_grow (&name_obstack, buffer, strlen (buffer));
1290 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1294 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1295 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1297 obstack_1grow (&name_obstack, '$');
1298 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1299 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1300 obstack_grow (&name_obstack, buffer, strlen (buffer));
1307 gcc_unreachable (); /* we treat these as scalars. */
1314 /* Create a human readable name for replacement variable of ACCESS. */
1317 make_fancy_name (tree expr)
1319 make_fancy_name_1 (expr);
1320 obstack_1grow (&name_obstack, '\0');
1321 return XOBFINISH (&name_obstack, char *);
1324 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1325 EXP_TYPE at the given OFFSET. If BASE is something for which
1326 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1327 to insert new statements either before or below the current one as specified
1328 by INSERT_AFTER. This function is not capable of handling bitfields. */
1331 build_ref_for_offset (tree base, HOST_WIDE_INT offset,
1332 tree exp_type, gimple_stmt_iterator *gsi,
1335 tree prev_base = base;
1337 location_t loc = EXPR_LOCATION (base);
1338 HOST_WIDE_INT base_offset;
1340 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1342 base = get_addr_base_and_unit_offset (base, &base_offset);
1344 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1345 offset such as array[var_index]. */
1351 gcc_checking_assert (gsi);
1352 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1353 add_referenced_var (tmp);
1354 tmp = make_ssa_name (tmp, NULL);
1355 addr = build_fold_addr_expr (unshare_expr (prev_base));
1356 stmt = gimple_build_assign (tmp, addr);
1357 SSA_NAME_DEF_STMT (tmp) = stmt;
1359 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1361 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1364 off = build_int_cst (reference_alias_ptr_type (prev_base),
1365 offset / BITS_PER_UNIT);
1368 else if (TREE_CODE (base) == MEM_REF)
1370 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1371 base_offset + offset / BITS_PER_UNIT);
1372 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off, 0);
1373 base = unshare_expr (TREE_OPERAND (base, 0));
1377 off = build_int_cst (reference_alias_ptr_type (base),
1378 base_offset + offset / BITS_PER_UNIT);
1379 base = build_fold_addr_expr (unshare_expr (base));
1382 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1385 /* Construct a memory reference to a part of an aggregate BASE at the given
1386 OFFSET and of the same type as MODEL. In case this is a reference to a
1387 bit-field, the function will replicate the last component_ref of model's
1388 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1389 build_ref_for_offset. */
1392 build_ref_for_model (tree base, HOST_WIDE_INT offset,
1393 struct access *model, gimple_stmt_iterator *gsi,
1396 if (TREE_CODE (model->expr) == COMPONENT_REF
1397 && DECL_BIT_FIELD (TREE_OPERAND (model->expr, 1)))
1399 /* This access represents a bit-field. */
1402 offset -= int_bit_position (TREE_OPERAND (model->expr, 1));
1403 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1404 t = build_ref_for_offset (base, offset, exp_type, gsi, insert_after);
1405 return fold_build3_loc (EXPR_LOCATION (base), COMPONENT_REF,
1406 model->type, t, TREE_OPERAND (model->expr, 1),
1410 return build_ref_for_offset (base, offset, model->type, gsi, insert_after);
1413 /* Construct a memory reference consisting of component_refs and array_refs to
1414 a part of an aggregate *RES (which is of type TYPE). The requested part
1415 should have type EXP_TYPE at be the given OFFSET. This function might not
1416 succeed, it returns true when it does and only then *RES points to something
1417 meaningful. This function should be used only to build expressions that we
1418 might need to present to user (e.g. in warnings). In all other situations,
1419 build_ref_for_model or build_ref_for_offset should be used instead. */
1422 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1428 tree tr_size, index, minidx;
1429 HOST_WIDE_INT el_size;
1431 if (offset == 0 && exp_type
1432 && types_compatible_p (exp_type, type))
1435 switch (TREE_CODE (type))
1438 case QUAL_UNION_TYPE:
1440 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1442 HOST_WIDE_INT pos, size;
1443 tree expr, *expr_ptr;
1445 if (TREE_CODE (fld) != FIELD_DECL)
1448 pos = int_bit_position (fld);
1449 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1450 tr_size = DECL_SIZE (fld);
1451 if (!tr_size || !host_integerp (tr_size, 1))
1453 size = tree_low_cst (tr_size, 1);
1459 else if (pos > offset || (pos + size) <= offset)
1462 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1465 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1466 offset - pos, exp_type))
1475 tr_size = TYPE_SIZE (TREE_TYPE (type));
1476 if (!tr_size || !host_integerp (tr_size, 1))
1478 el_size = tree_low_cst (tr_size, 1);
1480 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1481 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1483 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1484 if (!integer_zerop (minidx))
1485 index = int_const_binop (PLUS_EXPR, index, minidx, 0);
1486 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1487 NULL_TREE, NULL_TREE);
1488 offset = offset % el_size;
1489 type = TREE_TYPE (type);
1504 /* Return true iff TYPE is stdarg va_list type. */
1507 is_va_list_type (tree type)
1509 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1512 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1513 those with type which is suitable for scalarization. */
1516 find_var_candidates (void)
1519 referenced_var_iterator rvi;
1522 FOR_EACH_REFERENCED_VAR (var, rvi)
1524 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1526 type = TREE_TYPE (var);
1528 if (!AGGREGATE_TYPE_P (type)
1529 || needs_to_live_in_memory (var)
1530 || TREE_THIS_VOLATILE (var)
1531 || !COMPLETE_TYPE_P (type)
1532 || !host_integerp (TYPE_SIZE (type), 1)
1533 || tree_low_cst (TYPE_SIZE (type), 1) == 0
1534 || type_internals_preclude_sra_p (type)
1535 /* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1536 we also want to schedule it rather late. Thus we ignore it in
1538 || (sra_mode == SRA_MODE_EARLY_INTRA
1539 && is_va_list_type (type)))
1542 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1544 if (dump_file && (dump_flags & TDF_DETAILS))
1546 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1547 print_generic_expr (dump_file, var, 0);
1548 fprintf (dump_file, "\n");
1556 /* Sort all accesses for the given variable, check for partial overlaps and
1557 return NULL if there are any. If there are none, pick a representative for
1558 each combination of offset and size and create a linked list out of them.
1559 Return the pointer to the first representative and make sure it is the first
1560 one in the vector of accesses. */
1562 static struct access *
1563 sort_and_splice_var_accesses (tree var)
1565 int i, j, access_count;
1566 struct access *res, **prev_acc_ptr = &res;
1567 VEC (access_p, heap) *access_vec;
1569 HOST_WIDE_INT low = -1, high = 0;
1571 access_vec = get_base_access_vector (var);
1574 access_count = VEC_length (access_p, access_vec);
1576 /* Sort by <OFFSET, SIZE>. */
1577 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
1578 compare_access_positions);
1581 while (i < access_count)
1583 struct access *access = VEC_index (access_p, access_vec, i);
1584 bool grp_write = access->write;
1585 bool grp_read = !access->write;
1586 bool grp_assignment_read = access->grp_assignment_read;
1587 bool multiple_reads = false;
1588 bool total_scalarization = access->total_scalarization;
1589 bool grp_partial_lhs = access->grp_partial_lhs;
1590 bool first_scalar = is_gimple_reg_type (access->type);
1591 bool unscalarizable_region = access->grp_unscalarizable_region;
1593 if (first || access->offset >= high)
1596 low = access->offset;
1597 high = access->offset + access->size;
1599 else if (access->offset > low && access->offset + access->size > high)
1602 gcc_assert (access->offset >= low
1603 && access->offset + access->size <= high);
1606 while (j < access_count)
1608 struct access *ac2 = VEC_index (access_p, access_vec, j);
1609 if (ac2->offset != access->offset || ac2->size != access->size)
1616 multiple_reads = true;
1620 grp_assignment_read |= ac2->grp_assignment_read;
1621 grp_partial_lhs |= ac2->grp_partial_lhs;
1622 unscalarizable_region |= ac2->grp_unscalarizable_region;
1623 total_scalarization |= ac2->total_scalarization;
1624 relink_to_new_repr (access, ac2);
1626 /* If there are both aggregate-type and scalar-type accesses with
1627 this combination of size and offset, the comparison function
1628 should have put the scalars first. */
1629 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1630 ac2->group_representative = access;
1636 access->group_representative = access;
1637 access->grp_write = grp_write;
1638 access->grp_read = grp_read;
1639 access->grp_assignment_read = grp_assignment_read;
1640 access->grp_hint = multiple_reads || total_scalarization;
1641 access->grp_partial_lhs = grp_partial_lhs;
1642 access->grp_unscalarizable_region = unscalarizable_region;
1643 if (access->first_link)
1644 add_access_to_work_queue (access);
1646 *prev_acc_ptr = access;
1647 prev_acc_ptr = &access->next_grp;
1650 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1654 /* Create a variable for the given ACCESS which determines the type, name and a
1655 few other properties. Return the variable declaration and store it also to
1656 ACCESS->replacement. */
1659 create_access_replacement (struct access *access, bool rename)
1663 repl = create_tmp_var (access->type, "SR");
1665 add_referenced_var (repl);
1667 mark_sym_for_renaming (repl);
1669 if (!access->grp_partial_lhs
1670 && (TREE_CODE (access->type) == COMPLEX_TYPE
1671 || TREE_CODE (access->type) == VECTOR_TYPE))
1672 DECL_GIMPLE_REG_P (repl) = 1;
1674 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1675 DECL_ARTIFICIAL (repl) = 1;
1676 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1678 if (DECL_NAME (access->base)
1679 && !DECL_IGNORED_P (access->base)
1680 && !DECL_ARTIFICIAL (access->base))
1682 char *pretty_name = make_fancy_name (access->expr);
1683 tree debug_expr = unshare_expr (access->expr), d;
1685 DECL_NAME (repl) = get_identifier (pretty_name);
1686 obstack_free (&name_obstack, pretty_name);
1688 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1689 as DECL_DEBUG_EXPR isn't considered when looking for still
1690 used SSA_NAMEs and thus they could be freed. All debug info
1691 generation cares is whether something is constant or variable
1692 and that get_ref_base_and_extent works properly on the
1694 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1695 switch (TREE_CODE (d))
1698 case ARRAY_RANGE_REF:
1699 if (TREE_OPERAND (d, 1)
1700 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1701 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1702 if (TREE_OPERAND (d, 3)
1703 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1704 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1707 if (TREE_OPERAND (d, 2)
1708 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1709 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1714 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1715 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1716 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1719 TREE_NO_WARNING (repl) = 1;
1723 fprintf (dump_file, "Created a replacement for ");
1724 print_generic_expr (dump_file, access->base, 0);
1725 fprintf (dump_file, " offset: %u, size: %u: ",
1726 (unsigned) access->offset, (unsigned) access->size);
1727 print_generic_expr (dump_file, repl, 0);
1728 fprintf (dump_file, "\n");
1730 sra_stats.replacements++;
1735 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1738 get_access_replacement (struct access *access)
1740 gcc_assert (access->grp_to_be_replaced);
1742 if (!access->replacement_decl)
1743 access->replacement_decl = create_access_replacement (access, true);
1744 return access->replacement_decl;
1747 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1748 not mark it for renaming. */
1751 get_unrenamed_access_replacement (struct access *access)
1753 gcc_assert (!access->grp_to_be_replaced);
1755 if (!access->replacement_decl)
1756 access->replacement_decl = create_access_replacement (access, false);
1757 return access->replacement_decl;
1761 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1762 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1763 to it is not "within" the root. Return false iff some accesses partially
1767 build_access_subtree (struct access **access)
1769 struct access *root = *access, *last_child = NULL;
1770 HOST_WIDE_INT limit = root->offset + root->size;
1772 *access = (*access)->next_grp;
1773 while (*access && (*access)->offset + (*access)->size <= limit)
1776 root->first_child = *access;
1778 last_child->next_sibling = *access;
1779 last_child = *access;
1781 if (!build_access_subtree (access))
1785 if (*access && (*access)->offset < limit)
1791 /* Build a tree of access representatives, ACCESS is the pointer to the first
1792 one, others are linked in a list by the next_grp field. Return false iff
1793 some accesses partially overlap. */
1796 build_access_trees (struct access *access)
1800 struct access *root = access;
1802 if (!build_access_subtree (&access))
1804 root->next_grp = access;
1809 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1813 expr_with_var_bounded_array_refs_p (tree expr)
1815 while (handled_component_p (expr))
1817 if (TREE_CODE (expr) == ARRAY_REF
1818 && !host_integerp (array_ref_low_bound (expr), 0))
1820 expr = TREE_OPERAND (expr, 0);
1825 enum mark_read_status { SRA_MR_NOT_READ, SRA_MR_READ, SRA_MR_ASSIGN_READ};
1827 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1828 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1829 sorts of access flags appropriately along the way, notably always set
1830 grp_read and grp_assign_read according to MARK_READ and grp_write when
1831 MARK_WRITE is true. */
1834 analyze_access_subtree (struct access *root, bool allow_replacements,
1835 enum mark_read_status mark_read, bool mark_write)
1837 struct access *child;
1838 HOST_WIDE_INT limit = root->offset + root->size;
1839 HOST_WIDE_INT covered_to = root->offset;
1840 bool scalar = is_gimple_reg_type (root->type);
1841 bool hole = false, sth_created = false;
1842 bool direct_read = root->grp_read;
1844 if (mark_read == SRA_MR_ASSIGN_READ)
1847 root->grp_assignment_read = 1;
1849 if (mark_read == SRA_MR_READ)
1851 else if (root->grp_assignment_read)
1852 mark_read = SRA_MR_ASSIGN_READ;
1853 else if (root->grp_read)
1854 mark_read = SRA_MR_READ;
1857 root->grp_write = true;
1858 else if (root->grp_write)
1861 if (root->grp_unscalarizable_region)
1862 allow_replacements = false;
1864 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
1865 allow_replacements = false;
1867 for (child = root->first_child; child; child = child->next_sibling)
1869 if (!hole && child->offset < covered_to)
1872 covered_to += child->size;
1874 sth_created |= analyze_access_subtree (child,
1875 allow_replacements && !scalar,
1876 mark_read, mark_write);
1878 root->grp_unscalarized_data |= child->grp_unscalarized_data;
1879 hole |= !child->grp_covered;
1882 if (allow_replacements && scalar && !root->first_child
1884 || (root->grp_write && (direct_read || root->grp_assignment_read))))
1886 if (dump_file && (dump_flags & TDF_DETAILS))
1888 fprintf (dump_file, "Marking ");
1889 print_generic_expr (dump_file, root->base, 0);
1890 fprintf (dump_file, " offset: %u, size: %u: ",
1891 (unsigned) root->offset, (unsigned) root->size);
1892 fprintf (dump_file, " to be replaced.\n");
1895 root->grp_to_be_replaced = 1;
1899 else if (covered_to < limit)
1902 if (sth_created && !hole)
1904 root->grp_covered = 1;
1907 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
1908 root->grp_unscalarized_data = 1; /* not covered and written to */
1914 /* Analyze all access trees linked by next_grp by the means of
1915 analyze_access_subtree. */
1917 analyze_access_trees (struct access *access)
1923 if (analyze_access_subtree (access, true, SRA_MR_NOT_READ, false))
1925 access = access->next_grp;
1931 /* Return true iff a potential new child of LACC at offset OFFSET and with size
1932 SIZE would conflict with an already existing one. If exactly such a child
1933 already exists in LACC, store a pointer to it in EXACT_MATCH. */
1936 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
1937 HOST_WIDE_INT size, struct access **exact_match)
1939 struct access *child;
1941 for (child = lacc->first_child; child; child = child->next_sibling)
1943 if (child->offset == norm_offset && child->size == size)
1945 *exact_match = child;
1949 if (child->offset < norm_offset + size
1950 && child->offset + child->size > norm_offset)
1957 /* Create a new child access of PARENT, with all properties just like MODEL
1958 except for its offset and with its grp_write false and grp_read true.
1959 Return the new access or NULL if it cannot be created. Note that this access
1960 is created long after all splicing and sorting, it's not located in any
1961 access vector and is automatically a representative of its group. */
1963 static struct access *
1964 create_artificial_child_access (struct access *parent, struct access *model,
1965 HOST_WIDE_INT new_offset)
1967 struct access *access;
1968 struct access **child;
1969 tree expr = parent->base;
1971 gcc_assert (!model->grp_unscalarizable_region);
1972 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
1976 access = (struct access *) pool_alloc (access_pool);
1977 memset (access, 0, sizeof (struct access));
1978 access->base = parent->base;
1979 access->expr = expr;
1980 access->offset = new_offset;
1981 access->size = model->size;
1982 access->type = model->type;
1983 access->grp_write = true;
1984 access->grp_read = false;
1986 child = &parent->first_child;
1987 while (*child && (*child)->offset < new_offset)
1988 child = &(*child)->next_sibling;
1990 access->next_sibling = *child;
1997 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
1998 true if any new subaccess was created. Additionally, if RACC is a scalar
1999 access but LACC is not, change the type of the latter, if possible. */
2002 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2004 struct access *rchild;
2005 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2008 if (is_gimple_reg_type (lacc->type)
2009 || lacc->grp_unscalarizable_region
2010 || racc->grp_unscalarizable_region)
2013 if (!lacc->first_child && !racc->first_child
2014 && is_gimple_reg_type (racc->type))
2016 tree t = lacc->base;
2018 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t), lacc->offset,
2022 lacc->type = racc->type;
2027 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2029 struct access *new_acc = NULL;
2030 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2032 if (rchild->grp_unscalarizable_region)
2035 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2040 rchild->grp_hint = 1;
2041 new_acc->grp_hint |= new_acc->grp_read;
2042 if (rchild->first_child)
2043 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2048 rchild->grp_hint = 1;
2049 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2053 if (racc->first_child)
2054 propagate_subaccesses_across_link (new_acc, rchild);
2061 /* Propagate all subaccesses across assignment links. */
2064 propagate_all_subaccesses (void)
2066 while (work_queue_head)
2068 struct access *racc = pop_access_from_work_queue ();
2069 struct assign_link *link;
2071 gcc_assert (racc->first_link);
2073 for (link = racc->first_link; link; link = link->next)
2075 struct access *lacc = link->lacc;
2077 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2079 lacc = lacc->group_representative;
2080 if (propagate_subaccesses_across_link (lacc, racc)
2081 && lacc->first_link)
2082 add_access_to_work_queue (lacc);
2087 /* Go through all accesses collected throughout the (intraprocedural) analysis
2088 stage, exclude overlapping ones, identify representatives and build trees
2089 out of them, making decisions about scalarization on the way. Return true
2090 iff there are any to-be-scalarized variables after this stage. */
2093 analyze_all_variable_accesses (void)
2096 bitmap tmp = BITMAP_ALLOC (NULL);
2098 unsigned i, max_total_scalarization_size;
2100 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2101 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2103 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2104 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2105 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2107 tree var = referenced_var (i);
2109 if (TREE_CODE (var) == VAR_DECL
2110 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2111 <= max_total_scalarization_size)
2112 && type_consists_of_records_p (TREE_TYPE (var)))
2114 completely_scalarize_record (var, var, 0, var);
2115 if (dump_file && (dump_flags & TDF_DETAILS))
2117 fprintf (dump_file, "Will attempt to totally scalarize ");
2118 print_generic_expr (dump_file, var, 0);
2119 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2124 bitmap_copy (tmp, candidate_bitmap);
2125 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2127 tree var = referenced_var (i);
2128 struct access *access;
2130 access = sort_and_splice_var_accesses (var);
2131 if (!access || !build_access_trees (access))
2132 disqualify_candidate (var,
2133 "No or inhibitingly overlapping accesses.");
2136 propagate_all_subaccesses ();
2138 bitmap_copy (tmp, candidate_bitmap);
2139 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2141 tree var = referenced_var (i);
2142 struct access *access = get_first_repr_for_decl (var);
2144 if (analyze_access_trees (access))
2147 if (dump_file && (dump_flags & TDF_DETAILS))
2149 fprintf (dump_file, "\nAccess trees for ");
2150 print_generic_expr (dump_file, var, 0);
2151 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2152 dump_access_tree (dump_file, access);
2153 fprintf (dump_file, "\n");
2157 disqualify_candidate (var, "No scalar replacements to be created.");
2164 statistics_counter_event (cfun, "Scalarized aggregates", res);
2171 /* Generate statements copying scalar replacements of accesses within a subtree
2172 into or out of AGG. ACCESS is the first child of the root of the subtree to
2173 be processed. AGG is an aggregate type expression (can be a declaration but
2174 does not have to be, it can for example also be a mem_ref or a series of
2175 handled components). TOP_OFFSET is the offset of the processed subtree
2176 which has to be subtracted from offsets of individual accesses to get
2177 corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2178 replacements in the interval <start_offset, start_offset + chunk_size>,
2179 otherwise copy all. GSI is a statement iterator used to place the new
2180 statements. WRITE should be true when the statements should write from AGG
2181 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2182 statements will be added after the current statement in GSI, they will be
2183 added before the statement otherwise. */
2186 generate_subtree_copies (struct access *access, tree agg,
2187 HOST_WIDE_INT top_offset,
2188 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2189 gimple_stmt_iterator *gsi, bool write,
2194 if (chunk_size && access->offset >= start_offset + chunk_size)
2197 if (access->grp_to_be_replaced
2199 || access->offset + access->size > start_offset))
2201 tree expr, repl = get_access_replacement (access);
2204 expr = build_ref_for_model (agg, access->offset - top_offset,
2205 access, gsi, insert_after);
2209 if (access->grp_partial_lhs)
2210 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2212 insert_after ? GSI_NEW_STMT
2214 stmt = gimple_build_assign (repl, expr);
2218 TREE_NO_WARNING (repl) = 1;
2219 if (access->grp_partial_lhs)
2220 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2222 insert_after ? GSI_NEW_STMT
2224 stmt = gimple_build_assign (expr, repl);
2228 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2230 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2232 sra_stats.subtree_copies++;
2235 if (access->first_child)
2236 generate_subtree_copies (access->first_child, agg, top_offset,
2237 start_offset, chunk_size, gsi,
2238 write, insert_after);
2240 access = access->next_sibling;
2245 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2246 the root of the subtree to be processed. GSI is the statement iterator used
2247 for inserting statements which are added after the current statement if
2248 INSERT_AFTER is true or before it otherwise. */
2251 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2255 struct access *child;
2257 if (access->grp_to_be_replaced)
2261 stmt = gimple_build_assign (get_access_replacement (access),
2262 fold_convert (access->type,
2263 integer_zero_node));
2265 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2267 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2271 for (child = access->first_child; child; child = child->next_sibling)
2272 init_subtree_with_zero (child, gsi, insert_after);
2275 /* Search for an access representative for the given expression EXPR and
2276 return it or NULL if it cannot be found. */
2278 static struct access *
2279 get_access_for_expr (tree expr)
2281 HOST_WIDE_INT offset, size, max_size;
2284 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2285 a different size than the size of its argument and we need the latter
2287 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2288 expr = TREE_OPERAND (expr, 0);
2290 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2291 if (max_size == -1 || !DECL_P (base))
2294 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2297 return get_var_base_offset_size_access (base, offset, max_size);
2300 /* Replace the expression EXPR with a scalar replacement if there is one and
2301 generate other statements to do type conversion or subtree copying if
2302 necessary. GSI is used to place newly created statements, WRITE is true if
2303 the expression is being written to (it is on a LHS of a statement or output
2304 in an assembly statement). */
2307 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2309 struct access *access;
2312 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2315 expr = &TREE_OPERAND (*expr, 0);
2320 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2321 expr = &TREE_OPERAND (*expr, 0);
2322 access = get_access_for_expr (*expr);
2325 type = TREE_TYPE (*expr);
2327 if (access->grp_to_be_replaced)
2329 tree repl = get_access_replacement (access);
2330 /* If we replace a non-register typed access simply use the original
2331 access expression to extract the scalar component afterwards.
2332 This happens if scalarizing a function return value or parameter
2333 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2334 gcc.c-torture/compile/20011217-1.c.
2336 We also want to use this when accessing a complex or vector which can
2337 be accessed as a different type too, potentially creating a need for
2338 type conversion (see PR42196) and when scalarized unions are involved
2339 in assembler statements (see PR42398). */
2340 if (!useless_type_conversion_p (type, access->type))
2344 ref = build_ref_for_model (access->base, access->offset, access,
2351 if (access->grp_partial_lhs)
2352 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2353 false, GSI_NEW_STMT);
2354 stmt = gimple_build_assign (repl, ref);
2355 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2361 if (access->grp_partial_lhs)
2362 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2363 true, GSI_SAME_STMT);
2364 stmt = gimple_build_assign (ref, repl);
2365 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2373 if (access->first_child)
2375 HOST_WIDE_INT start_offset, chunk_size;
2377 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2378 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2380 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2381 start_offset = access->offset
2382 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2385 start_offset = chunk_size = 0;
2387 generate_subtree_copies (access->first_child, access->base, 0,
2388 start_offset, chunk_size, gsi, write, write);
2393 /* Where scalar replacements of the RHS have been written to when a replacement
2394 of a LHS of an assigments cannot be direclty loaded from a replacement of
2396 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2397 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2398 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2400 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2401 base aggregate if there are unscalarized data or directly to LHS
2404 static enum unscalarized_data_handling
2405 handle_unscalarized_data_in_subtree (struct access *top_racc, tree lhs,
2406 gimple_stmt_iterator *gsi)
2408 if (top_racc->grp_unscalarized_data)
2410 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2412 return SRA_UDH_RIGHT;
2416 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2417 0, 0, gsi, false, false);
2418 return SRA_UDH_LEFT;
2423 /* Try to generate statements to load all sub-replacements in an access
2424 (sub)tree (LACC is the first child) from scalar replacements in the TOP_RACC
2425 (sub)tree. If that is not possible, refresh the TOP_RACC base aggregate and
2426 load the accesses from it. LEFT_OFFSET is the offset of the left whole
2427 subtree being copied, RIGHT_OFFSET is the same thing for the right subtree.
2428 NEW_GSI is stmt iterator used for statement insertions after the original
2429 assignment, OLD_GSI is used to insert statements before the assignment.
2430 *REFRESHED keeps the information whether we have needed to refresh
2431 replacements of the LHS and from which side of the assignments this takes
2435 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2436 HOST_WIDE_INT left_offset,
2437 HOST_WIDE_INT right_offset,
2438 gimple_stmt_iterator *old_gsi,
2439 gimple_stmt_iterator *new_gsi,
2440 enum unscalarized_data_handling *refreshed,
2443 location_t loc = EXPR_LOCATION (lacc->expr);
2446 if (lacc->grp_to_be_replaced)
2448 struct access *racc;
2449 HOST_WIDE_INT offset = lacc->offset - left_offset + right_offset;
2453 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2454 if (racc && racc->grp_to_be_replaced)
2456 rhs = get_access_replacement (racc);
2457 if (!useless_type_conversion_p (lacc->type, racc->type))
2458 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2462 /* No suitable access on the right hand side, need to load from
2463 the aggregate. See if we have to update it first... */
2464 if (*refreshed == SRA_UDH_NONE)
2465 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2468 if (*refreshed == SRA_UDH_LEFT)
2469 rhs = build_ref_for_model (lacc->base, lacc->offset, lacc,
2472 rhs = build_ref_for_model (top_racc->base, offset, lacc,
2476 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2477 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2479 sra_stats.subreplacements++;
2481 else if (*refreshed == SRA_UDH_NONE
2482 && lacc->grp_read && !lacc->grp_covered)
2483 *refreshed = handle_unscalarized_data_in_subtree (top_racc, lhs,
2486 if (lacc->first_child)
2487 load_assign_lhs_subreplacements (lacc->first_child, top_racc,
2488 left_offset, right_offset,
2489 old_gsi, new_gsi, refreshed, lhs);
2490 lacc = lacc->next_sibling;
2495 /* Result code for SRA assignment modification. */
2496 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2497 SRA_AM_MODIFIED, /* stmt changed but not
2499 SRA_AM_REMOVED }; /* stmt eliminated */
2501 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2502 to the assignment and GSI is the statement iterator pointing at it. Returns
2503 the same values as sra_modify_assign. */
2505 static enum assignment_mod_result
2506 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2508 tree lhs = gimple_assign_lhs (*stmt);
2511 acc = get_access_for_expr (lhs);
2515 if (VEC_length (constructor_elt,
2516 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2518 /* I have never seen this code path trigger but if it can happen the
2519 following should handle it gracefully. */
2520 if (access_has_children_p (acc))
2521 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2523 return SRA_AM_MODIFIED;
2526 if (acc->grp_covered)
2528 init_subtree_with_zero (acc, gsi, false);
2529 unlink_stmt_vdef (*stmt);
2530 gsi_remove (gsi, true);
2531 return SRA_AM_REMOVED;
2535 init_subtree_with_zero (acc, gsi, true);
2536 return SRA_AM_MODIFIED;
2540 /* Create and return a new suitable default definition SSA_NAME for RACC which
2541 is an access describing an uninitialized part of an aggregate that is being
2545 get_repl_default_def_ssa_name (struct access *racc)
2549 decl = get_unrenamed_access_replacement (racc);
2551 repl = gimple_default_def (cfun, decl);
2554 repl = make_ssa_name (decl, gimple_build_nop ());
2555 set_default_def (decl, repl);
2561 /* Examine both sides of the assignment statement pointed to by STMT, replace
2562 them with a scalare replacement if there is one and generate copying of
2563 replacements if scalarized aggregates have been used in the assignment. GSI
2564 is used to hold generated statements for type conversions and subtree
2567 static enum assignment_mod_result
2568 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2570 struct access *lacc, *racc;
2572 bool modify_this_stmt = false;
2573 bool force_gimple_rhs = false;
2574 location_t loc = gimple_location (*stmt);
2575 gimple_stmt_iterator orig_gsi = *gsi;
2577 if (!gimple_assign_single_p (*stmt))
2579 lhs = gimple_assign_lhs (*stmt);
2580 rhs = gimple_assign_rhs1 (*stmt);
2582 if (TREE_CODE (rhs) == CONSTRUCTOR)
2583 return sra_modify_constructor_assign (stmt, gsi);
2585 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2586 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2587 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2589 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2591 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2593 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2596 lacc = get_access_for_expr (lhs);
2597 racc = get_access_for_expr (rhs);
2601 if (lacc && lacc->grp_to_be_replaced)
2603 lhs = get_access_replacement (lacc);
2604 gimple_assign_set_lhs (*stmt, lhs);
2605 modify_this_stmt = true;
2606 if (lacc->grp_partial_lhs)
2607 force_gimple_rhs = true;
2611 if (racc && racc->grp_to_be_replaced)
2613 rhs = get_access_replacement (racc);
2614 modify_this_stmt = true;
2615 if (racc->grp_partial_lhs)
2616 force_gimple_rhs = true;
2620 if (modify_this_stmt)
2622 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2624 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2625 ??? This should move to fold_stmt which we simply should
2626 call after building a VIEW_CONVERT_EXPR here. */
2627 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2628 && !access_has_children_p (lacc))
2630 lhs = build_ref_for_offset (lhs, 0, TREE_TYPE (rhs), gsi, false);
2631 gimple_assign_set_lhs (*stmt, lhs);
2633 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2634 && !contains_view_convert_expr_p (rhs)
2635 && !access_has_children_p (racc))
2636 rhs = build_ref_for_offset (rhs, 0, TREE_TYPE (lhs), gsi, false);
2638 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2640 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
2642 if (is_gimple_reg_type (TREE_TYPE (lhs))
2643 && TREE_CODE (lhs) != SSA_NAME)
2644 force_gimple_rhs = true;
2649 /* From this point on, the function deals with assignments in between
2650 aggregates when at least one has scalar reductions of some of its
2651 components. There are three possible scenarios: Both the LHS and RHS have
2652 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2654 In the first case, we would like to load the LHS components from RHS
2655 components whenever possible. If that is not possible, we would like to
2656 read it directly from the RHS (after updating it by storing in it its own
2657 components). If there are some necessary unscalarized data in the LHS,
2658 those will be loaded by the original assignment too. If neither of these
2659 cases happen, the original statement can be removed. Most of this is done
2660 by load_assign_lhs_subreplacements.
2662 In the second case, we would like to store all RHS scalarized components
2663 directly into LHS and if they cover the aggregate completely, remove the
2664 statement too. In the third case, we want the LHS components to be loaded
2665 directly from the RHS (DSE will remove the original statement if it
2668 This is a bit complex but manageable when types match and when unions do
2669 not cause confusion in a way that we cannot really load a component of LHS
2670 from the RHS or vice versa (the access representing this level can have
2671 subaccesses that are accessible only through a different union field at a
2672 higher level - different from the one used in the examined expression).
2675 Therefore, I specially handle a fourth case, happening when there is a
2676 specific type cast or it is impossible to locate a scalarized subaccess on
2677 the other side of the expression. If that happens, I simply "refresh" the
2678 RHS by storing in it is scalarized components leave the original statement
2679 there to do the copying and then load the scalar replacements of the LHS.
2680 This is what the first branch does. */
2682 if (gimple_has_volatile_ops (*stmt)
2683 || contains_view_convert_expr_p (rhs)
2684 || contains_view_convert_expr_p (lhs))
2686 if (access_has_children_p (racc))
2687 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2689 if (access_has_children_p (lacc))
2690 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2692 sra_stats.separate_lhs_rhs_handling++;
2696 if (access_has_children_p (lacc) && access_has_children_p (racc))
2698 gimple_stmt_iterator orig_gsi = *gsi;
2699 enum unscalarized_data_handling refreshed;
2701 if (lacc->grp_read && !lacc->grp_covered)
2702 refreshed = handle_unscalarized_data_in_subtree (racc, lhs, gsi);
2704 refreshed = SRA_UDH_NONE;
2706 load_assign_lhs_subreplacements (lacc->first_child, racc,
2707 lacc->offset, racc->offset,
2708 &orig_gsi, gsi, &refreshed, lhs);
2709 if (refreshed != SRA_UDH_RIGHT)
2712 unlink_stmt_vdef (*stmt);
2713 gsi_remove (&orig_gsi, true);
2714 sra_stats.deleted++;
2715 return SRA_AM_REMOVED;
2722 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2726 fprintf (dump_file, "Removing load: ");
2727 print_gimple_stmt (dump_file, *stmt, 0, 0);
2730 if (TREE_CODE (lhs) == SSA_NAME)
2732 rhs = get_repl_default_def_ssa_name (racc);
2733 if (!useless_type_conversion_p (TREE_TYPE (lhs),
2735 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR,
2736 TREE_TYPE (lhs), rhs);
2740 if (racc->first_child)
2741 generate_subtree_copies (racc->first_child, lhs,
2742 racc->offset, 0, 0, gsi,
2745 gcc_assert (*stmt == gsi_stmt (*gsi));
2746 unlink_stmt_vdef (*stmt);
2747 gsi_remove (gsi, true);
2748 sra_stats.deleted++;
2749 return SRA_AM_REMOVED;
2752 else if (racc->first_child)
2753 generate_subtree_copies (racc->first_child, lhs,
2754 racc->offset, 0, 0, gsi, false, true);
2756 if (access_has_children_p (lacc))
2757 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2758 0, 0, gsi, true, true);
2762 /* This gimplification must be done after generate_subtree_copies, lest we
2763 insert the subtree copies in the middle of the gimplified sequence. */
2764 if (force_gimple_rhs)
2765 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2766 true, GSI_SAME_STMT);
2767 if (gimple_assign_rhs1 (*stmt) != rhs)
2769 modify_this_stmt = true;
2770 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2771 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2774 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2777 /* Traverse the function body and all modifications as decided in
2778 analyze_all_variable_accesses. Return true iff the CFG has been
2782 sra_modify_function_body (void)
2784 bool cfg_changed = false;
2789 gimple_stmt_iterator gsi = gsi_start_bb (bb);
2790 while (!gsi_end_p (gsi))
2792 gimple stmt = gsi_stmt (gsi);
2793 enum assignment_mod_result assign_result;
2794 bool modified = false, deleted = false;
2798 switch (gimple_code (stmt))
2801 t = gimple_return_retval_ptr (stmt);
2802 if (*t != NULL_TREE)
2803 modified |= sra_modify_expr (t, &gsi, false);
2807 assign_result = sra_modify_assign (&stmt, &gsi);
2808 modified |= assign_result == SRA_AM_MODIFIED;
2809 deleted = assign_result == SRA_AM_REMOVED;
2813 /* Operands must be processed before the lhs. */
2814 for (i = 0; i < gimple_call_num_args (stmt); i++)
2816 t = gimple_call_arg_ptr (stmt, i);
2817 modified |= sra_modify_expr (t, &gsi, false);
2820 if (gimple_call_lhs (stmt))
2822 t = gimple_call_lhs_ptr (stmt);
2823 modified |= sra_modify_expr (t, &gsi, true);
2828 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
2830 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
2831 modified |= sra_modify_expr (t, &gsi, false);
2833 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
2835 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
2836 modified |= sra_modify_expr (t, &gsi, true);
2847 if (maybe_clean_eh_stmt (stmt)
2848 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
2859 /* Generate statements initializing scalar replacements of parts of function
2863 initialize_parameter_reductions (void)
2865 gimple_stmt_iterator gsi;
2866 gimple_seq seq = NULL;
2869 for (parm = DECL_ARGUMENTS (current_function_decl);
2871 parm = DECL_CHAIN (parm))
2873 VEC (access_p, heap) *access_vec;
2874 struct access *access;
2876 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
2878 access_vec = get_base_access_vector (parm);
2884 seq = gimple_seq_alloc ();
2885 gsi = gsi_start (seq);
2888 for (access = VEC_index (access_p, access_vec, 0);
2890 access = access->next_grp)
2891 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true);
2895 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
2898 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
2899 it reveals there are components of some aggregates to be scalarized, it runs
2900 the required transformations. */
2902 perform_intra_sra (void)
2907 if (!find_var_candidates ())
2910 if (!scan_function ())
2913 if (!analyze_all_variable_accesses ())
2916 if (sra_modify_function_body ())
2917 ret = TODO_update_ssa | TODO_cleanup_cfg;
2919 ret = TODO_update_ssa;
2920 initialize_parameter_reductions ();
2922 statistics_counter_event (cfun, "Scalar replacements created",
2923 sra_stats.replacements);
2924 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
2925 statistics_counter_event (cfun, "Subtree copy stmts",
2926 sra_stats.subtree_copies);
2927 statistics_counter_event (cfun, "Subreplacement stmts",
2928 sra_stats.subreplacements);
2929 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
2930 statistics_counter_event (cfun, "Separate LHS and RHS handling",
2931 sra_stats.separate_lhs_rhs_handling);
2934 sra_deinitialize ();
2938 /* Perform early intraprocedural SRA. */
2940 early_intra_sra (void)
2942 sra_mode = SRA_MODE_EARLY_INTRA;
2943 return perform_intra_sra ();
2946 /* Perform "late" intraprocedural SRA. */
2948 late_intra_sra (void)
2950 sra_mode = SRA_MODE_INTRA;
2951 return perform_intra_sra ();
2956 gate_intra_sra (void)
2958 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
2962 struct gimple_opt_pass pass_sra_early =
2967 gate_intra_sra, /* gate */
2968 early_intra_sra, /* execute */
2971 0, /* static_pass_number */
2972 TV_TREE_SRA, /* tv_id */
2973 PROP_cfg | PROP_ssa, /* properties_required */
2974 0, /* properties_provided */
2975 0, /* properties_destroyed */
2976 0, /* todo_flags_start */
2980 | TODO_verify_ssa /* todo_flags_finish */
2984 struct gimple_opt_pass pass_sra =
2989 gate_intra_sra, /* gate */
2990 late_intra_sra, /* execute */
2993 0, /* static_pass_number */
2994 TV_TREE_SRA, /* tv_id */
2995 PROP_cfg | PROP_ssa, /* properties_required */
2996 0, /* properties_provided */
2997 0, /* properties_destroyed */
2998 TODO_update_address_taken, /* todo_flags_start */
3002 | TODO_verify_ssa /* todo_flags_finish */
3007 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3011 is_unused_scalar_param (tree parm)
3014 return (is_gimple_reg (parm)
3015 && (!(name = gimple_default_def (cfun, parm))
3016 || has_zero_uses (name)));
3019 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3020 examine whether there are any direct or otherwise infeasible ones. If so,
3021 return true, otherwise return false. PARM must be a gimple register with a
3022 non-NULL default definition. */
3025 ptr_parm_has_direct_uses (tree parm)
3027 imm_use_iterator ui;
3029 tree name = gimple_default_def (cfun, parm);
3032 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3035 use_operand_p use_p;
3037 if (is_gimple_debug (stmt))
3040 /* Valid uses include dereferences on the lhs and the rhs. */
3041 if (gimple_has_lhs (stmt))
3043 tree lhs = gimple_get_lhs (stmt);
3044 while (handled_component_p (lhs))
3045 lhs = TREE_OPERAND (lhs, 0);
3046 if (TREE_CODE (lhs) == MEM_REF
3047 && TREE_OPERAND (lhs, 0) == name
3048 && integer_zerop (TREE_OPERAND (lhs, 1))
3049 && types_compatible_p (TREE_TYPE (lhs),
3050 TREE_TYPE (TREE_TYPE (name))))
3053 if (gimple_assign_single_p (stmt))
3055 tree rhs = gimple_assign_rhs1 (stmt);
3056 while (handled_component_p (rhs))
3057 rhs = TREE_OPERAND (rhs, 0);
3058 if (TREE_CODE (rhs) == MEM_REF
3059 && TREE_OPERAND (rhs, 0) == name
3060 && integer_zerop (TREE_OPERAND (rhs, 1))
3061 && types_compatible_p (TREE_TYPE (rhs),
3062 TREE_TYPE (TREE_TYPE (name))))
3065 else if (is_gimple_call (stmt))
3068 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3070 tree arg = gimple_call_arg (stmt, i);
3071 while (handled_component_p (arg))
3072 arg = TREE_OPERAND (arg, 0);
3073 if (TREE_CODE (arg) == MEM_REF
3074 && TREE_OPERAND (arg, 0) == name
3075 && integer_zerop (TREE_OPERAND (arg, 1))
3076 && types_compatible_p (TREE_TYPE (arg),
3077 TREE_TYPE (TREE_TYPE (name))))
3082 /* If the number of valid uses does not match the number of
3083 uses in this stmt there is an unhandled use. */
3084 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3091 BREAK_FROM_IMM_USE_STMT (ui);
3097 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3098 them in candidate_bitmap. Note that these do not necessarily include
3099 parameter which are unused and thus can be removed. Return true iff any
3100 such candidate has been found. */
3103 find_param_candidates (void)
3109 for (parm = DECL_ARGUMENTS (current_function_decl);
3111 parm = DECL_CHAIN (parm))
3113 tree type = TREE_TYPE (parm);
3117 if (TREE_THIS_VOLATILE (parm)
3118 || TREE_ADDRESSABLE (parm)
3119 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3122 if (is_unused_scalar_param (parm))
3128 if (POINTER_TYPE_P (type))
3130 type = TREE_TYPE (type);
3132 if (TREE_CODE (type) == FUNCTION_TYPE
3133 || TYPE_VOLATILE (type)
3134 || (TREE_CODE (type) == ARRAY_TYPE
3135 && TYPE_NONALIASED_COMPONENT (type))
3136 || !is_gimple_reg (parm)
3137 || is_va_list_type (type)
3138 || ptr_parm_has_direct_uses (parm))
3141 else if (!AGGREGATE_TYPE_P (type))
3144 if (!COMPLETE_TYPE_P (type)
3145 || !host_integerp (TYPE_SIZE (type), 1)
3146 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3147 || (AGGREGATE_TYPE_P (type)
3148 && type_internals_preclude_sra_p (type)))
3151 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3153 if (dump_file && (dump_flags & TDF_DETAILS))
3155 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3156 print_generic_expr (dump_file, parm, 0);
3157 fprintf (dump_file, "\n");
3161 func_param_count = count;
3165 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3169 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3172 struct access *repr = (struct access *) data;
3174 repr->grp_maybe_modified = 1;
3178 /* Analyze what representatives (in linked lists accessible from
3179 REPRESENTATIVES) can be modified by side effects of statements in the
3180 current function. */
3183 analyze_modified_params (VEC (access_p, heap) *representatives)
3187 for (i = 0; i < func_param_count; i++)
3189 struct access *repr;
3191 for (repr = VEC_index (access_p, representatives, i);
3193 repr = repr->next_grp)
3195 struct access *access;
3199 if (no_accesses_p (repr))
3201 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3202 || repr->grp_maybe_modified)
3205 ao_ref_init (&ar, repr->expr);
3206 visited = BITMAP_ALLOC (NULL);
3207 for (access = repr; access; access = access->next_sibling)
3209 /* All accesses are read ones, otherwise grp_maybe_modified would
3210 be trivially set. */
3211 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3212 mark_maybe_modified, repr, &visited);
3213 if (repr->grp_maybe_modified)
3216 BITMAP_FREE (visited);
3221 /* Propagate distances in bb_dereferences in the opposite direction than the
3222 control flow edges, in each step storing the maximum of the current value
3223 and the minimum of all successors. These steps are repeated until the table
3224 stabilizes. Note that BBs which might terminate the functions (according to
3225 final_bbs bitmap) never updated in this way. */
3228 propagate_dereference_distances (void)
3230 VEC (basic_block, heap) *queue;
3233 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3234 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3237 VEC_quick_push (basic_block, queue, bb);
3241 while (!VEC_empty (basic_block, queue))
3245 bool change = false;
3248 bb = VEC_pop (basic_block, queue);
3251 if (bitmap_bit_p (final_bbs, bb->index))
3254 for (i = 0; i < func_param_count; i++)
3256 int idx = bb->index * func_param_count + i;
3258 HOST_WIDE_INT inh = 0;
3260 FOR_EACH_EDGE (e, ei, bb->succs)
3262 int succ_idx = e->dest->index * func_param_count + i;
3264 if (e->src == EXIT_BLOCK_PTR)
3270 inh = bb_dereferences [succ_idx];
3272 else if (bb_dereferences [succ_idx] < inh)
3273 inh = bb_dereferences [succ_idx];
3276 if (!first && bb_dereferences[idx] < inh)
3278 bb_dereferences[idx] = inh;
3283 if (change && !bitmap_bit_p (final_bbs, bb->index))
3284 FOR_EACH_EDGE (e, ei, bb->preds)
3289 e->src->aux = e->src;
3290 VEC_quick_push (basic_block, queue, e->src);
3294 VEC_free (basic_block, heap, queue);
3297 /* Dump a dereferences TABLE with heading STR to file F. */
3300 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3304 fprintf (dump_file, str);
3305 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3307 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3308 if (bb != EXIT_BLOCK_PTR)
3311 for (i = 0; i < func_param_count; i++)
3313 int idx = bb->index * func_param_count + i;
3314 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3319 fprintf (dump_file, "\n");
3322 /* Determine what (parts of) parameters passed by reference that are not
3323 assigned to are not certainly dereferenced in this function and thus the
3324 dereferencing cannot be safely moved to the caller without potentially
3325 introducing a segfault. Mark such REPRESENTATIVES as
3326 grp_not_necessarilly_dereferenced.
3328 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3329 part is calculated rather than simple booleans are calculated for each
3330 pointer parameter to handle cases when only a fraction of the whole
3331 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3334 The maximum dereference distances for each pointer parameter and BB are
3335 already stored in bb_dereference. This routine simply propagates these
3336 values upwards by propagate_dereference_distances and then compares the
3337 distances of individual parameters in the ENTRY BB to the equivalent
3338 distances of each representative of a (fraction of a) parameter. */
3341 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3345 if (dump_file && (dump_flags & TDF_DETAILS))
3346 dump_dereferences_table (dump_file,
3347 "Dereference table before propagation:\n",
3350 propagate_dereference_distances ();
3352 if (dump_file && (dump_flags & TDF_DETAILS))
3353 dump_dereferences_table (dump_file,
3354 "Dereference table after propagation:\n",
3357 for (i = 0; i < func_param_count; i++)
3359 struct access *repr = VEC_index (access_p, representatives, i);
3360 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3362 if (!repr || no_accesses_p (repr))
3367 if ((repr->offset + repr->size) > bb_dereferences[idx])
3368 repr->grp_not_necessarilly_dereferenced = 1;
3369 repr = repr->next_grp;
3375 /* Return the representative access for the parameter declaration PARM if it is
3376 a scalar passed by reference which is not written to and the pointer value
3377 is not used directly. Thus, if it is legal to dereference it in the caller
3378 and we can rule out modifications through aliases, such parameter should be
3379 turned into one passed by value. Return NULL otherwise. */
3381 static struct access *
3382 unmodified_by_ref_scalar_representative (tree parm)
3384 int i, access_count;
3385 struct access *repr;
3386 VEC (access_p, heap) *access_vec;
3388 access_vec = get_base_access_vector (parm);
3389 gcc_assert (access_vec);
3390 repr = VEC_index (access_p, access_vec, 0);
3393 repr->group_representative = repr;
3395 access_count = VEC_length (access_p, access_vec);
3396 for (i = 1; i < access_count; i++)
3398 struct access *access = VEC_index (access_p, access_vec, i);
3401 access->group_representative = repr;
3402 access->next_sibling = repr->next_sibling;
3403 repr->next_sibling = access;
3407 repr->grp_scalar_ptr = 1;
3411 /* Return true iff this access precludes IPA-SRA of the parameter it is
3415 access_precludes_ipa_sra_p (struct access *access)
3417 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3418 is incompatible assign in a call statement (and possibly even in asm
3419 statements). This can be relaxed by using a new temporary but only for
3420 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3421 intraprocedural SRA we deal with this by keeping the old aggregate around,
3422 something we cannot do in IPA-SRA.) */
3424 && (is_gimple_call (access->stmt)
3425 || gimple_code (access->stmt) == GIMPLE_ASM))
3432 /* Sort collected accesses for parameter PARM, identify representatives for
3433 each accessed region and link them together. Return NULL if there are
3434 different but overlapping accesses, return the special ptr value meaning
3435 there are no accesses for this parameter if that is the case and return the
3436 first representative otherwise. Set *RO_GRP if there is a group of accesses
3437 with only read (i.e. no write) accesses. */
3439 static struct access *
3440 splice_param_accesses (tree parm, bool *ro_grp)
3442 int i, j, access_count, group_count;
3443 int agg_size, total_size = 0;
3444 struct access *access, *res, **prev_acc_ptr = &res;
3445 VEC (access_p, heap) *access_vec;
3447 access_vec = get_base_access_vector (parm);
3449 return &no_accesses_representant;
3450 access_count = VEC_length (access_p, access_vec);
3452 qsort (VEC_address (access_p, access_vec), access_count, sizeof (access_p),
3453 compare_access_positions);
3458 while (i < access_count)
3461 access = VEC_index (access_p, access_vec, i);
3462 modification = access->write;
3463 if (access_precludes_ipa_sra_p (access))
3466 /* Access is about to become group representative unless we find some
3467 nasty overlap which would preclude us from breaking this parameter
3471 while (j < access_count)
3473 struct access *ac2 = VEC_index (access_p, access_vec, j);
3474 if (ac2->offset != access->offset)
3476 /* All or nothing law for parameters. */
3477 if (access->offset + access->size > ac2->offset)
3482 else if (ac2->size != access->size)
3485 if (access_precludes_ipa_sra_p (ac2))
3488 modification |= ac2->write;
3489 ac2->group_representative = access;
3490 ac2->next_sibling = access->next_sibling;
3491 access->next_sibling = ac2;
3496 access->grp_maybe_modified = modification;
3499 *prev_acc_ptr = access;
3500 prev_acc_ptr = &access->next_grp;
3501 total_size += access->size;
3505 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3506 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3508 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3509 if (total_size >= agg_size)
3512 gcc_assert (group_count > 0);
3516 /* Decide whether parameters with representative accesses given by REPR should
3517 be reduced into components. */
3520 decide_one_param_reduction (struct access *repr)
3522 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3527 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3528 gcc_assert (cur_parm_size > 0);
3530 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3533 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3538 agg_size = cur_parm_size;
3544 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3545 print_generic_expr (dump_file, parm, 0);
3546 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3547 for (acc = repr; acc; acc = acc->next_grp)
3548 dump_access (dump_file, acc, true);
3552 new_param_count = 0;
3554 for (; repr; repr = repr->next_grp)
3556 gcc_assert (parm == repr->base);
3559 if (!by_ref || (!repr->grp_maybe_modified
3560 && !repr->grp_not_necessarilly_dereferenced))
3561 total_size += repr->size;
3563 total_size += cur_parm_size;
3566 gcc_assert (new_param_count > 0);
3568 if (optimize_function_for_size_p (cfun))
3569 parm_size_limit = cur_parm_size;
3571 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3574 if (total_size < agg_size
3575 && total_size <= parm_size_limit)
3578 fprintf (dump_file, " ....will be split into %i components\n",
3580 return new_param_count;
3586 /* The order of the following enums is important, we need to do extra work for
3587 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3588 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3589 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3591 /* Identify representatives of all accesses to all candidate parameters for
3592 IPA-SRA. Return result based on what representatives have been found. */
3594 static enum ipa_splicing_result
3595 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3597 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3599 struct access *repr;
3601 *representatives = VEC_alloc (access_p, heap, func_param_count);
3603 for (parm = DECL_ARGUMENTS (current_function_decl);
3605 parm = DECL_CHAIN (parm))
3607 if (is_unused_scalar_param (parm))
3609 VEC_quick_push (access_p, *representatives,
3610 &no_accesses_representant);
3611 if (result == NO_GOOD_ACCESS)
3612 result = UNUSED_PARAMS;
3614 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3615 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3616 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3618 repr = unmodified_by_ref_scalar_representative (parm);
3619 VEC_quick_push (access_p, *representatives, repr);
3621 result = UNMODIF_BY_REF_ACCESSES;
3623 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3625 bool ro_grp = false;
3626 repr = splice_param_accesses (parm, &ro_grp);
3627 VEC_quick_push (access_p, *representatives, repr);
3629 if (repr && !no_accesses_p (repr))
3631 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3634 result = UNMODIF_BY_REF_ACCESSES;
3635 else if (result < MODIF_BY_REF_ACCESSES)
3636 result = MODIF_BY_REF_ACCESSES;
3638 else if (result < BY_VAL_ACCESSES)
3639 result = BY_VAL_ACCESSES;
3641 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3642 result = UNUSED_PARAMS;
3645 VEC_quick_push (access_p, *representatives, NULL);
3648 if (result == NO_GOOD_ACCESS)
3650 VEC_free (access_p, heap, *representatives);
3651 *representatives = NULL;
3652 return NO_GOOD_ACCESS;
3658 /* Return the index of BASE in PARMS. Abort if it is not found. */
3661 get_param_index (tree base, VEC(tree, heap) *parms)
3665 len = VEC_length (tree, parms);
3666 for (i = 0; i < len; i++)
3667 if (VEC_index (tree, parms, i) == base)
3672 /* Convert the decisions made at the representative level into compact
3673 parameter adjustments. REPRESENTATIVES are pointers to first
3674 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3675 final number of adjustments. */
3677 static ipa_parm_adjustment_vec
3678 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3679 int adjustments_count)
3681 VEC (tree, heap) *parms;
3682 ipa_parm_adjustment_vec adjustments;
3686 gcc_assert (adjustments_count > 0);
3687 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3688 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3689 parm = DECL_ARGUMENTS (current_function_decl);
3690 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
3692 struct access *repr = VEC_index (access_p, representatives, i);
3694 if (!repr || no_accesses_p (repr))
3696 struct ipa_parm_adjustment *adj;
3698 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3699 memset (adj, 0, sizeof (*adj));
3700 adj->base_index = get_param_index (parm, parms);
3703 adj->copy_param = 1;
3705 adj->remove_param = 1;
3709 struct ipa_parm_adjustment *adj;
3710 int index = get_param_index (parm, parms);
3712 for (; repr; repr = repr->next_grp)
3714 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3715 memset (adj, 0, sizeof (*adj));
3716 gcc_assert (repr->base == parm);
3717 adj->base_index = index;
3718 adj->base = repr->base;
3719 adj->type = repr->type;
3720 adj->offset = repr->offset;
3721 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3722 && (repr->grp_maybe_modified
3723 || repr->grp_not_necessarilly_dereferenced));
3728 VEC_free (tree, heap, parms);
3732 /* Analyze the collected accesses and produce a plan what to do with the
3733 parameters in the form of adjustments, NULL meaning nothing. */
3735 static ipa_parm_adjustment_vec
3736 analyze_all_param_acesses (void)
3738 enum ipa_splicing_result repr_state;
3739 bool proceed = false;
3740 int i, adjustments_count = 0;
3741 VEC (access_p, heap) *representatives;
3742 ipa_parm_adjustment_vec adjustments;
3744 repr_state = splice_all_param_accesses (&representatives);
3745 if (repr_state == NO_GOOD_ACCESS)
3748 /* If there are any parameters passed by reference which are not modified
3749 directly, we need to check whether they can be modified indirectly. */
3750 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3752 analyze_caller_dereference_legality (representatives);
3753 analyze_modified_params (representatives);
3756 for (i = 0; i < func_param_count; i++)
3758 struct access *repr = VEC_index (access_p, representatives, i);
3760 if (repr && !no_accesses_p (repr))
3762 if (repr->grp_scalar_ptr)
3764 adjustments_count++;
3765 if (repr->grp_not_necessarilly_dereferenced
3766 || repr->grp_maybe_modified)
3767 VEC_replace (access_p, representatives, i, NULL);
3771 sra_stats.scalar_by_ref_to_by_val++;
3776 int new_components = decide_one_param_reduction (repr);
3778 if (new_components == 0)
3780 VEC_replace (access_p, representatives, i, NULL);
3781 adjustments_count++;
3785 adjustments_count += new_components;
3786 sra_stats.aggregate_params_reduced++;
3787 sra_stats.param_reductions_created += new_components;
3794 if (no_accesses_p (repr))
3797 sra_stats.deleted_unused_parameters++;
3799 adjustments_count++;
3803 if (!proceed && dump_file)
3804 fprintf (dump_file, "NOT proceeding to change params.\n");
3807 adjustments = turn_representatives_into_adjustments (representatives,
3812 VEC_free (access_p, heap, representatives);
3816 /* If a parameter replacement identified by ADJ does not yet exist in the form
3817 of declaration, create it and record it, otherwise return the previously
3821 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
3824 if (!adj->new_ssa_base)
3826 char *pretty_name = make_fancy_name (adj->base);
3828 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
3829 DECL_NAME (repl) = get_identifier (pretty_name);
3830 obstack_free (&name_obstack, pretty_name);
3833 add_referenced_var (repl);
3834 adj->new_ssa_base = repl;
3837 repl = adj->new_ssa_base;
3841 /* Find the first adjustment for a particular parameter BASE in a vector of
3842 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3845 static struct ipa_parm_adjustment *
3846 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
3850 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3851 for (i = 0; i < len; i++)
3853 struct ipa_parm_adjustment *adj;
3855 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3856 if (!adj->copy_param && adj->base == base)
3863 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3864 removed because its value is not used, replace the SSA_NAME with a one
3865 relating to a created VAR_DECL together all of its uses and return true.
3866 ADJUSTMENTS is a pointer to an adjustments vector. */
3869 replace_removed_params_ssa_names (gimple stmt,
3870 ipa_parm_adjustment_vec adjustments)
3872 struct ipa_parm_adjustment *adj;
3873 tree lhs, decl, repl, name;
3875 if (gimple_code (stmt) == GIMPLE_PHI)
3876 lhs = gimple_phi_result (stmt);
3877 else if (is_gimple_assign (stmt))
3878 lhs = gimple_assign_lhs (stmt);
3879 else if (is_gimple_call (stmt))
3880 lhs = gimple_call_lhs (stmt);
3884 if (TREE_CODE (lhs) != SSA_NAME)
3886 decl = SSA_NAME_VAR (lhs);
3887 if (TREE_CODE (decl) != PARM_DECL)
3890 adj = get_adjustment_for_base (adjustments, decl);
3894 repl = get_replaced_param_substitute (adj);
3895 name = make_ssa_name (repl, stmt);
3899 fprintf (dump_file, "replacing an SSA name of a removed param ");
3900 print_generic_expr (dump_file, lhs, 0);
3901 fprintf (dump_file, " with ");
3902 print_generic_expr (dump_file, name, 0);
3903 fprintf (dump_file, "\n");
3906 if (is_gimple_assign (stmt))
3907 gimple_assign_set_lhs (stmt, name);
3908 else if (is_gimple_call (stmt))
3909 gimple_call_set_lhs (stmt, name);
3911 gimple_phi_set_result (stmt, name);
3913 replace_uses_by (lhs, name);
3914 release_ssa_name (lhs);
3918 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
3919 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
3920 specifies whether the function should care about type incompatibility the
3921 current and new expressions. If it is false, the function will leave
3922 incompatibility issues to the caller. Return true iff the expression
3926 sra_ipa_modify_expr (tree *expr, bool convert,
3927 ipa_parm_adjustment_vec adjustments)
3930 struct ipa_parm_adjustment *adj, *cand = NULL;
3931 HOST_WIDE_INT offset, size, max_size;
3934 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3936 if (TREE_CODE (*expr) == BIT_FIELD_REF
3937 || TREE_CODE (*expr) == IMAGPART_EXPR
3938 || TREE_CODE (*expr) == REALPART_EXPR)
3940 expr = &TREE_OPERAND (*expr, 0);
3944 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
3945 if (!base || size == -1 || max_size == -1)
3948 if (TREE_CODE (base) == MEM_REF)
3950 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
3951 base = TREE_OPERAND (base, 0);
3954 base = get_ssa_base_param (base);
3955 if (!base || TREE_CODE (base) != PARM_DECL)
3958 for (i = 0; i < len; i++)
3960 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3962 if (adj->base == base &&
3963 (adj->offset == offset || adj->remove_param))
3969 if (!cand || cand->copy_param || cand->remove_param)
3973 src = build_simple_mem_ref (cand->reduction);
3975 src = cand->reduction;
3977 if (dump_file && (dump_flags & TDF_DETAILS))
3979 fprintf (dump_file, "About to replace expr ");
3980 print_generic_expr (dump_file, *expr, 0);
3981 fprintf (dump_file, " with ");
3982 print_generic_expr (dump_file, src, 0);
3983 fprintf (dump_file, "\n");
3986 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
3988 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
3996 /* If the statement pointed to by STMT_PTR contains any expressions that need
3997 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
3998 potential type incompatibilities (GSI is used to accommodate conversion
3999 statements and must point to the statement). Return true iff the statement
4003 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4004 ipa_parm_adjustment_vec adjustments)
4006 gimple stmt = *stmt_ptr;
4007 tree *lhs_p, *rhs_p;
4010 if (!gimple_assign_single_p (stmt))
4013 rhs_p = gimple_assign_rhs1_ptr (stmt);
4014 lhs_p = gimple_assign_lhs_ptr (stmt);
4016 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4017 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4020 tree new_rhs = NULL_TREE;
4022 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4024 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4026 /* V_C_Es of constructors can cause trouble (PR 42714). */
4027 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4028 *rhs_p = fold_convert (TREE_TYPE (*lhs_p), integer_zero_node);
4030 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4033 new_rhs = fold_build1_loc (gimple_location (stmt),
4034 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4037 else if (REFERENCE_CLASS_P (*rhs_p)
4038 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4039 && !is_gimple_reg (*lhs_p))
4040 /* This can happen when an assignment in between two single field
4041 structures is turned into an assignment in between two pointers to
4042 scalars (PR 42237). */
4047 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4048 true, GSI_SAME_STMT);
4050 gimple_assign_set_rhs_from_tree (gsi, tmp);
4059 /* Traverse the function body and all modifications as described in
4060 ADJUSTMENTS. Return true iff the CFG has been changed. */
4063 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4065 bool cfg_changed = false;
4070 gimple_stmt_iterator gsi;
4072 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4073 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4075 gsi = gsi_start_bb (bb);
4076 while (!gsi_end_p (gsi))
4078 gimple stmt = gsi_stmt (gsi);
4079 bool modified = false;
4083 switch (gimple_code (stmt))
4086 t = gimple_return_retval_ptr (stmt);
4087 if (*t != NULL_TREE)
4088 modified |= sra_ipa_modify_expr (t, true, adjustments);
4092 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4093 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4097 /* Operands must be processed before the lhs. */
4098 for (i = 0; i < gimple_call_num_args (stmt); i++)
4100 t = gimple_call_arg_ptr (stmt, i);
4101 modified |= sra_ipa_modify_expr (t, true, adjustments);
4104 if (gimple_call_lhs (stmt))
4106 t = gimple_call_lhs_ptr (stmt);
4107 modified |= sra_ipa_modify_expr (t, false, adjustments);
4108 modified |= replace_removed_params_ssa_names (stmt,
4114 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4116 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4117 modified |= sra_ipa_modify_expr (t, true, adjustments);
4119 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4121 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4122 modified |= sra_ipa_modify_expr (t, false, adjustments);
4133 if (maybe_clean_eh_stmt (stmt)
4134 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4144 /* Call gimple_debug_bind_reset_value on all debug statements describing
4145 gimple register parameters that are being removed or replaced. */
4148 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4152 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4153 for (i = 0; i < len; i++)
4155 struct ipa_parm_adjustment *adj;
4156 imm_use_iterator ui;
4160 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4161 if (adj->copy_param || !is_gimple_reg (adj->base))
4163 name = gimple_default_def (cfun, adj->base);
4166 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4168 /* All other users must have been removed by
4169 ipa_sra_modify_function_body. */
4170 gcc_assert (is_gimple_debug (stmt));
4171 gimple_debug_bind_reset_value (stmt);
4177 /* Return true iff all callers have at least as many actual arguments as there
4178 are formal parameters in the current function. */
4181 all_callers_have_enough_arguments_p (struct cgraph_node *node)
4183 struct cgraph_edge *cs;
4184 for (cs = node->callers; cs; cs = cs->next_caller)
4185 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4192 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4195 convert_callers (struct cgraph_node *node, tree old_decl,
4196 ipa_parm_adjustment_vec adjustments)
4198 tree old_cur_fndecl = current_function_decl;
4199 struct cgraph_edge *cs;
4200 basic_block this_block;
4201 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4203 for (cs = node->callers; cs; cs = cs->next_caller)
4205 current_function_decl = cs->caller->decl;
4206 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4209 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4210 cs->caller->uid, cs->callee->uid,
4211 cgraph_node_name (cs->caller),
4212 cgraph_node_name (cs->callee));
4214 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4219 for (cs = node->callers; cs; cs = cs->next_caller)
4220 if (bitmap_set_bit (recomputed_callers, cs->caller->uid))
4221 compute_inline_parameters (cs->caller);
4222 BITMAP_FREE (recomputed_callers);
4224 current_function_decl = old_cur_fndecl;
4226 if (!encountered_recursive_call)
4229 FOR_EACH_BB (this_block)
4231 gimple_stmt_iterator gsi;
4233 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4235 gimple stmt = gsi_stmt (gsi);
4237 if (gimple_code (stmt) != GIMPLE_CALL)
4239 call_fndecl = gimple_call_fndecl (stmt);
4240 if (call_fndecl == old_decl)
4243 fprintf (dump_file, "Adjusting recursive call");
4244 gimple_call_set_fndecl (stmt, node->decl);
4245 ipa_modify_call_arguments (NULL, stmt, adjustments);
4253 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4254 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4257 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4259 struct cgraph_node *new_node;
4260 struct cgraph_edge *cs;
4262 VEC (cgraph_edge_p, heap) * redirect_callers;
4266 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
4268 redirect_callers = VEC_alloc (cgraph_edge_p, heap, node_callers);
4269 for (cs = node->callers; cs != NULL; cs = cs->next_caller)
4270 VEC_quick_push (cgraph_edge_p, redirect_callers, cs);
4272 rebuild_cgraph_edges ();
4274 current_function_decl = NULL_TREE;
4276 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4277 NULL, NULL, "isra");
4278 current_function_decl = new_node->decl;
4279 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4281 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4282 cfg_changed = ipa_sra_modify_function_body (adjustments);
4283 sra_ipa_reset_debug_stmts (adjustments);
4284 convert_callers (new_node, node->decl, adjustments);
4285 cgraph_make_node_local (new_node);
4289 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4290 attributes, return true otherwise. NODE is the cgraph node of the current
4294 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4296 if (!cgraph_node_can_be_local_p (node))
4299 fprintf (dump_file, "Function not local to this compilation unit.\n");
4303 if (!tree_versionable_function_p (node->decl))
4306 fprintf (dump_file, "Function is not versionable.\n");
4310 if (DECL_VIRTUAL_P (current_function_decl))
4313 fprintf (dump_file, "Function is a virtual method.\n");
4317 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4318 && node->global.size >= MAX_INLINE_INSNS_AUTO)
4321 fprintf (dump_file, "Function too big to be made truly local.\n");
4329 "Function has no callers in this compilation unit.\n");
4336 fprintf (dump_file, "Function uses stdarg. \n");
4340 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4346 /* Perform early interprocedural SRA. */
4349 ipa_early_sra (void)
4351 struct cgraph_node *node = cgraph_node (current_function_decl);
4352 ipa_parm_adjustment_vec adjustments;
4355 if (!ipa_sra_preliminary_function_checks (node))
4359 sra_mode = SRA_MODE_EARLY_IPA;
4361 if (!find_param_candidates ())
4364 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4368 if (!all_callers_have_enough_arguments_p (node))
4371 fprintf (dump_file, "There are callers with insufficient number of "
4376 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4378 * last_basic_block_for_function (cfun));
4379 final_bbs = BITMAP_ALLOC (NULL);
4382 if (encountered_apply_args)
4385 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4389 if (encountered_unchangable_recursive_call)
4392 fprintf (dump_file, "Function calls itself with insufficient "
4393 "number of arguments.\n");
4397 adjustments = analyze_all_param_acesses ();
4401 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4403 if (modify_function (node, adjustments))
4404 ret = TODO_update_ssa | TODO_cleanup_cfg;
4406 ret = TODO_update_ssa;
4407 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4409 statistics_counter_event (cfun, "Unused parameters deleted",
4410 sra_stats.deleted_unused_parameters);
4411 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4412 sra_stats.scalar_by_ref_to_by_val);
4413 statistics_counter_event (cfun, "Aggregate parameters broken up",
4414 sra_stats.aggregate_params_reduced);
4415 statistics_counter_event (cfun, "Aggregate parameter components created",
4416 sra_stats.param_reductions_created);
4419 BITMAP_FREE (final_bbs);
4420 free (bb_dereferences);
4422 sra_deinitialize ();
4426 /* Return if early ipa sra shall be performed. */
4428 ipa_early_sra_gate (void)
4430 return flag_ipa_sra && dbg_cnt (eipa_sra);
4433 struct gimple_opt_pass pass_early_ipa_sra =
4437 "eipa_sra", /* name */
4438 ipa_early_sra_gate, /* gate */
4439 ipa_early_sra, /* execute */
4442 0, /* static_pass_number */
4443 TV_IPA_SRA, /* tv_id */
4444 0, /* properties_required */
4445 0, /* properties_provided */
4446 0, /* properties_destroyed */
4447 0, /* todo_flags_start */
4448 TODO_dump_func | TODO_dump_cgraph /* todo_flags_finish */