1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
4 Copyright (C) 2008, 2009, 2010 Free Software Foundation, Inc.
5 Contributed by Martin Jambor <mjambor@suse.cz>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it under
10 the terms of the GNU General Public License as published by the Free
11 Software Foundation; either version 3, or (at your option) any later
14 GCC is distributed in the hope that it will be useful, but WITHOUT ANY
15 WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING3. If not see
21 <http://www.gnu.org/licenses/>. */
23 /* This file implements Scalar Reduction of Aggregates (SRA). SRA is run
24 twice, once in the early stages of compilation (early SRA) and once in the
25 late stages (late SRA). The aim of both is to turn references to scalar
26 parts of aggregates into uses of independent scalar variables.
28 The two passes are nearly identical, the only difference is that early SRA
29 does not scalarize unions which are used as the result in a GIMPLE_RETURN
30 statement because together with inlining this can lead to weird type
33 Both passes operate in four stages:
35 1. The declarations that have properties which make them candidates for
36 scalarization are identified in function find_var_candidates(). The
37 candidates are stored in candidate_bitmap.
39 2. The function body is scanned. In the process, declarations which are
40 used in a manner that prevent their scalarization are removed from the
41 candidate bitmap. More importantly, for every access into an aggregate,
42 an access structure (struct access) is created by create_access() and
43 stored in a vector associated with the aggregate. Among other
44 information, the aggregate declaration, the offset and size of the access
45 and its type are stored in the structure.
47 On a related note, assign_link structures are created for every assign
48 statement between candidate aggregates and attached to the related
51 3. The vectors of accesses are analyzed. They are first sorted according to
52 their offset and size and then scanned for partially overlapping accesses
53 (i.e. those which overlap but one is not entirely within another). Such
54 an access disqualifies the whole aggregate from being scalarized.
56 If there is no such inhibiting overlap, a representative access structure
57 is chosen for every unique combination of offset and size. Afterwards,
58 the pass builds a set of trees from these structures, in which children
59 of an access are within their parent (in terms of offset and size).
61 Then accesses are propagated whenever possible (i.e. in cases when it
62 does not create a partially overlapping access) across assign_links from
63 the right hand side to the left hand side.
65 Then the set of trees for each declaration is traversed again and those
66 accesses which should be replaced by a scalar are identified.
68 4. The function is traversed again, and for every reference into an
69 aggregate that has some component which is about to be scalarized,
70 statements are amended and new statements are created as necessary.
71 Finally, if a parameter got scalarized, the scalar replacements are
72 initialized with values from respective parameter aggregates. */
76 #include "coretypes.h"
77 #include "alloc-pool.h"
82 #include "tree-flow.h"
84 #include "tree-pretty-print.h"
85 #include "statistics.h"
86 #include "tree-dump.h"
92 #include "tree-inline.h"
93 #include "gimple-pretty-print.h"
94 #include "ipa-inline.h"
96 /* Enumeration of all aggregate reductions we can do. */
97 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
98 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
99 SRA_MODE_INTRA }; /* late intraprocedural SRA */
101 /* Global variable describing which aggregate reduction we are performing at
103 static enum sra_mode sra_mode;
107 /* ACCESS represents each access to an aggregate variable (as a whole or a
108 part). It can also represent a group of accesses that refer to exactly the
109 same fragment of an aggregate (i.e. those that have exactly the same offset
110 and size). Such representatives for a single aggregate, once determined,
111 are linked in a linked list and have the group fields set.
113 Moreover, when doing intraprocedural SRA, a tree is built from those
114 representatives (by the means of first_child and next_sibling pointers), in
115 which all items in a subtree are "within" the root, i.e. their offset is
116 greater or equal to offset of the root and offset+size is smaller or equal
117 to offset+size of the root. Children of an access are sorted by offset.
119 Note that accesses to parts of vector and complex number types always
120 represented by an access to the whole complex number or a vector. It is a
121 duty of the modifying functions to replace them appropriately. */
125 /* Values returned by `get_ref_base_and_extent' for each component reference
126 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
127 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
128 HOST_WIDE_INT offset;
132 /* Expression. It is context dependent so do not use it to create new
133 expressions to access the original aggregate. See PR 42154 for a
139 /* The statement this access belongs to. */
142 /* Next group representative for this aggregate. */
143 struct access *next_grp;
145 /* Pointer to the group representative. Pointer to itself if the struct is
146 the representative. */
147 struct access *group_representative;
149 /* If this access has any children (in terms of the definition above), this
150 points to the first one. */
151 struct access *first_child;
153 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
154 described above. In IPA-SRA this is a pointer to the next access
155 belonging to the same group (having the same representative). */
156 struct access *next_sibling;
158 /* Pointers to the first and last element in the linked list of assign
160 struct assign_link *first_link, *last_link;
162 /* Pointer to the next access in the work queue. */
163 struct access *next_queued;
165 /* Replacement variable for this access "region." Never to be accessed
166 directly, always only by the means of get_access_replacement() and only
167 when grp_to_be_replaced flag is set. */
168 tree replacement_decl;
170 /* Is this particular access write access? */
173 /* Is this access an artificial one created to scalarize some record
175 unsigned total_scalarization : 1;
177 /* Is this access an access to a non-addressable field? */
178 unsigned non_addressable : 1;
180 /* Is this access currently in the work queue? */
181 unsigned grp_queued : 1;
183 /* Does this group contain a write access? This flag is propagated down the
185 unsigned grp_write : 1;
187 /* Does this group contain a read access? This flag is propagated down the
189 unsigned grp_read : 1;
191 /* Does this group contain a read access that comes from an assignment
192 statement? This flag is propagated down the access tree. */
193 unsigned grp_assignment_read : 1;
195 /* Does this group contain a write access that comes from an assignment
196 statement? This flag is propagated down the access tree. */
197 unsigned grp_assignment_write : 1;
199 /* Does this group contain a read access through a scalar type? This flag is
200 not propagated in the access tree in any direction. */
201 unsigned grp_scalar_read : 1;
203 /* Does this group contain a write access through a scalar type? This flag
204 is not propagated in the access tree in any direction. */
205 unsigned grp_scalar_write : 1;
207 /* Other passes of the analysis use this bit to make function
208 analyze_access_subtree create scalar replacements for this group if
210 unsigned grp_hint : 1;
212 /* Is the subtree rooted in this access fully covered by scalar
214 unsigned grp_covered : 1;
216 /* If set to true, this access and all below it in an access tree must not be
218 unsigned grp_unscalarizable_region : 1;
220 /* Whether data have been written to parts of the aggregate covered by this
221 access which is not to be scalarized. This flag is propagated up in the
223 unsigned grp_unscalarized_data : 1;
225 /* Does this access and/or group contain a write access through a
227 unsigned grp_partial_lhs : 1;
229 /* Set when a scalar replacement should be created for this variable. We do
230 the decision and creation at different places because create_tmp_var
231 cannot be called from within FOR_EACH_REFERENCED_VAR. */
232 unsigned grp_to_be_replaced : 1;
234 /* Should TREE_NO_WARNING of a replacement be set? */
235 unsigned grp_no_warning : 1;
237 /* Is it possible that the group refers to data which might be (directly or
238 otherwise) modified? */
239 unsigned grp_maybe_modified : 1;
241 /* Set when this is a representative of a pointer to scalar (i.e. by
242 reference) parameter which we consider for turning into a plain scalar
243 (i.e. a by value parameter). */
244 unsigned grp_scalar_ptr : 1;
246 /* Set when we discover that this pointer is not safe to dereference in the
248 unsigned grp_not_necessarilly_dereferenced : 1;
251 typedef struct access *access_p;
253 DEF_VEC_P (access_p);
254 DEF_VEC_ALLOC_P (access_p, heap);
256 /* Alloc pool for allocating access structures. */
257 static alloc_pool access_pool;
259 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
260 are used to propagate subaccesses from rhs to lhs as long as they don't
261 conflict with what is already there. */
264 struct access *lacc, *racc;
265 struct assign_link *next;
268 /* Alloc pool for allocating assign link structures. */
269 static alloc_pool link_pool;
271 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
272 static struct pointer_map_t *base_access_vec;
274 /* Bitmap of candidates. */
275 static bitmap candidate_bitmap;
277 /* Bitmap of candidates which we should try to entirely scalarize away and
278 those which cannot be (because they are and need be used as a whole). */
279 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
281 /* Obstack for creation of fancy names. */
282 static struct obstack name_obstack;
284 /* Head of a linked list of accesses that need to have its subaccesses
285 propagated to their assignment counterparts. */
286 static struct access *work_queue_head;
288 /* Number of parameters of the analyzed function when doing early ipa SRA. */
289 static int func_param_count;
291 /* scan_function sets the following to true if it encounters a call to
292 __builtin_apply_args. */
293 static bool encountered_apply_args;
295 /* Set by scan_function when it finds a recursive call. */
296 static bool encountered_recursive_call;
298 /* Set by scan_function when it finds a recursive call with less actual
299 arguments than formal parameters.. */
300 static bool encountered_unchangable_recursive_call;
302 /* This is a table in which for each basic block and parameter there is a
303 distance (offset + size) in that parameter which is dereferenced and
304 accessed in that BB. */
305 static HOST_WIDE_INT *bb_dereferences;
306 /* Bitmap of BBs that can cause the function to "stop" progressing by
307 returning, throwing externally, looping infinitely or calling a function
308 which might abort etc.. */
309 static bitmap final_bbs;
311 /* Representative of no accesses at all. */
312 static struct access no_accesses_representant;
314 /* Predicate to test the special value. */
317 no_accesses_p (struct access *access)
319 return access == &no_accesses_representant;
322 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
323 representative fields are dumped, otherwise those which only describe the
324 individual access are. */
328 /* Number of processed aggregates is readily available in
329 analyze_all_variable_accesses and so is not stored here. */
331 /* Number of created scalar replacements. */
334 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
338 /* Number of statements created by generate_subtree_copies. */
341 /* Number of statements created by load_assign_lhs_subreplacements. */
344 /* Number of times sra_modify_assign has deleted a statement. */
347 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
348 RHS reparately due to type conversions or nonexistent matching
350 int separate_lhs_rhs_handling;
352 /* Number of parameters that were removed because they were unused. */
353 int deleted_unused_parameters;
355 /* Number of scalars passed as parameters by reference that have been
356 converted to be passed by value. */
357 int scalar_by_ref_to_by_val;
359 /* Number of aggregate parameters that were replaced by one or more of their
361 int aggregate_params_reduced;
363 /* Numbber of components created when splitting aggregate parameters. */
364 int param_reductions_created;
368 dump_access (FILE *f, struct access *access, bool grp)
370 fprintf (f, "access { ");
371 fprintf (f, "base = (%d)'", DECL_UID (access->base));
372 print_generic_expr (f, access->base, 0);
373 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
374 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
375 fprintf (f, ", expr = ");
376 print_generic_expr (f, access->expr, 0);
377 fprintf (f, ", type = ");
378 print_generic_expr (f, access->type, 0);
380 fprintf (f, ", total_scalarization = %d, grp_read = %d, grp_write = %d, "
381 "grp_assignment_read = %d, grp_assignment_write = %d, "
382 "grp_scalar_read = %d, grp_scalar_write = %d, "
383 "grp_hint = %d, grp_covered = %d, "
384 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
385 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
386 "grp_maybe_modified = %d, "
387 "grp_not_necessarilly_dereferenced = %d\n",
388 access->total_scalarization, access->grp_read, access->grp_write,
389 access->grp_assignment_read, access->grp_assignment_write,
390 access->grp_scalar_read, access->grp_scalar_write,
391 access->grp_hint, access->grp_covered,
392 access->grp_unscalarizable_region, access->grp_unscalarized_data,
393 access->grp_partial_lhs, access->grp_to_be_replaced,
394 access->grp_maybe_modified,
395 access->grp_not_necessarilly_dereferenced);
397 fprintf (f, ", write = %d, total_scalarization = %d, "
398 "grp_partial_lhs = %d\n",
399 access->write, access->total_scalarization,
400 access->grp_partial_lhs);
403 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
406 dump_access_tree_1 (FILE *f, struct access *access, int level)
412 for (i = 0; i < level; i++)
413 fputs ("* ", dump_file);
415 dump_access (f, access, true);
417 if (access->first_child)
418 dump_access_tree_1 (f, access->first_child, level + 1);
420 access = access->next_sibling;
425 /* Dump all access trees for a variable, given the pointer to the first root in
429 dump_access_tree (FILE *f, struct access *access)
431 for (; access; access = access->next_grp)
432 dump_access_tree_1 (f, access, 0);
435 /* Return true iff ACC is non-NULL and has subaccesses. */
438 access_has_children_p (struct access *acc)
440 return acc && acc->first_child;
443 /* Return a vector of pointers to accesses for the variable given in BASE or
444 NULL if there is none. */
446 static VEC (access_p, heap) *
447 get_base_access_vector (tree base)
451 slot = pointer_map_contains (base_access_vec, base);
455 return *(VEC (access_p, heap) **) slot;
458 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
459 in ACCESS. Return NULL if it cannot be found. */
461 static struct access *
462 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
465 while (access && (access->offset != offset || access->size != size))
467 struct access *child = access->first_child;
469 while (child && (child->offset + child->size <= offset))
470 child = child->next_sibling;
477 /* Return the first group representative for DECL or NULL if none exists. */
479 static struct access *
480 get_first_repr_for_decl (tree base)
482 VEC (access_p, heap) *access_vec;
484 access_vec = get_base_access_vector (base);
488 return VEC_index (access_p, access_vec, 0);
491 /* Find an access representative for the variable BASE and given OFFSET and
492 SIZE. Requires that access trees have already been built. Return NULL if
493 it cannot be found. */
495 static struct access *
496 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
499 struct access *access;
501 access = get_first_repr_for_decl (base);
502 while (access && (access->offset + access->size <= offset))
503 access = access->next_grp;
507 return find_access_in_subtree (access, offset, size);
510 /* Add LINK to the linked list of assign links of RACC. */
512 add_link_to_rhs (struct access *racc, struct assign_link *link)
514 gcc_assert (link->racc == racc);
516 if (!racc->first_link)
518 gcc_assert (!racc->last_link);
519 racc->first_link = link;
522 racc->last_link->next = link;
524 racc->last_link = link;
528 /* Move all link structures in their linked list in OLD_RACC to the linked list
531 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
533 if (!old_racc->first_link)
535 gcc_assert (!old_racc->last_link);
539 if (new_racc->first_link)
541 gcc_assert (!new_racc->last_link->next);
542 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
544 new_racc->last_link->next = old_racc->first_link;
545 new_racc->last_link = old_racc->last_link;
549 gcc_assert (!new_racc->last_link);
551 new_racc->first_link = old_racc->first_link;
552 new_racc->last_link = old_racc->last_link;
554 old_racc->first_link = old_racc->last_link = NULL;
557 /* Add ACCESS to the work queue (which is actually a stack). */
560 add_access_to_work_queue (struct access *access)
562 if (!access->grp_queued)
564 gcc_assert (!access->next_queued);
565 access->next_queued = work_queue_head;
566 access->grp_queued = 1;
567 work_queue_head = access;
571 /* Pop an access from the work queue, and return it, assuming there is one. */
573 static struct access *
574 pop_access_from_work_queue (void)
576 struct access *access = work_queue_head;
578 work_queue_head = access->next_queued;
579 access->next_queued = NULL;
580 access->grp_queued = 0;
585 /* Allocate necessary structures. */
588 sra_initialize (void)
590 candidate_bitmap = BITMAP_ALLOC (NULL);
591 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
592 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
593 gcc_obstack_init (&name_obstack);
594 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
595 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
596 base_access_vec = pointer_map_create ();
597 memset (&sra_stats, 0, sizeof (sra_stats));
598 encountered_apply_args = false;
599 encountered_recursive_call = false;
600 encountered_unchangable_recursive_call = false;
603 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
606 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
607 void *data ATTRIBUTE_UNUSED)
609 VEC (access_p, heap) *access_vec;
610 access_vec = (VEC (access_p, heap) *) *value;
611 VEC_free (access_p, heap, access_vec);
616 /* Deallocate all general structures. */
619 sra_deinitialize (void)
621 BITMAP_FREE (candidate_bitmap);
622 BITMAP_FREE (should_scalarize_away_bitmap);
623 BITMAP_FREE (cannot_scalarize_away_bitmap);
624 free_alloc_pool (access_pool);
625 free_alloc_pool (link_pool);
626 obstack_free (&name_obstack, NULL);
628 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
629 pointer_map_destroy (base_access_vec);
632 /* Remove DECL from candidates for SRA and write REASON to the dump file if
635 disqualify_candidate (tree decl, const char *reason)
637 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
639 if (dump_file && (dump_flags & TDF_DETAILS))
641 fprintf (dump_file, "! Disqualifying ");
642 print_generic_expr (dump_file, decl, 0);
643 fprintf (dump_file, " - %s\n", reason);
647 /* Return true iff the type contains a field or an element which does not allow
651 type_internals_preclude_sra_p (tree type)
656 switch (TREE_CODE (type))
660 case QUAL_UNION_TYPE:
661 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
662 if (TREE_CODE (fld) == FIELD_DECL)
664 tree ft = TREE_TYPE (fld);
666 if (TREE_THIS_VOLATILE (fld)
667 || !DECL_FIELD_OFFSET (fld) || !DECL_SIZE (fld)
668 || !host_integerp (DECL_FIELD_OFFSET (fld), 1)
669 || !host_integerp (DECL_SIZE (fld), 1)
670 || (AGGREGATE_TYPE_P (ft)
671 && int_bit_position (fld) % BITS_PER_UNIT != 0))
674 if (AGGREGATE_TYPE_P (ft)
675 && type_internals_preclude_sra_p (ft))
682 et = TREE_TYPE (type);
684 if (AGGREGATE_TYPE_P (et))
685 return type_internals_preclude_sra_p (et);
694 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
695 base variable if it is. Return T if it is not an SSA_NAME. */
698 get_ssa_base_param (tree t)
700 if (TREE_CODE (t) == SSA_NAME)
702 if (SSA_NAME_IS_DEFAULT_DEF (t))
703 return SSA_NAME_VAR (t);
710 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
711 belongs to, unless the BB has already been marked as a potentially
715 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
717 basic_block bb = gimple_bb (stmt);
718 int idx, parm_index = 0;
721 if (bitmap_bit_p (final_bbs, bb->index))
724 for (parm = DECL_ARGUMENTS (current_function_decl);
725 parm && parm != base;
726 parm = DECL_CHAIN (parm))
729 gcc_assert (parm_index < func_param_count);
731 idx = bb->index * func_param_count + parm_index;
732 if (bb_dereferences[idx] < dist)
733 bb_dereferences[idx] = dist;
736 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
737 the three fields. Also add it to the vector of accesses corresponding to
738 the base. Finally, return the new access. */
740 static struct access *
741 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
743 VEC (access_p, heap) *vec;
744 struct access *access;
747 access = (struct access *) pool_alloc (access_pool);
748 memset (access, 0, sizeof (struct access));
750 access->offset = offset;
753 slot = pointer_map_contains (base_access_vec, base);
755 vec = (VEC (access_p, heap) *) *slot;
757 vec = VEC_alloc (access_p, heap, 32);
759 VEC_safe_push (access_p, heap, vec, access);
761 *((struct VEC (access_p,heap) **)
762 pointer_map_insert (base_access_vec, base)) = vec;
767 /* Create and insert access for EXPR. Return created access, or NULL if it is
770 static struct access *
771 create_access (tree expr, gimple stmt, bool write)
773 struct access *access;
774 HOST_WIDE_INT offset, size, max_size;
776 bool ptr, unscalarizable_region = false;
778 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
780 if (sra_mode == SRA_MODE_EARLY_IPA
781 && TREE_CODE (base) == MEM_REF)
783 base = get_ssa_base_param (TREE_OPERAND (base, 0));
791 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
794 if (sra_mode == SRA_MODE_EARLY_IPA)
796 if (size < 0 || size != max_size)
798 disqualify_candidate (base, "Encountered a variable sized access.");
801 if (TREE_CODE (expr) == COMPONENT_REF
802 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
804 disqualify_candidate (base, "Encountered a bit-field access.");
807 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
810 mark_parm_dereference (base, offset + size, stmt);
814 if (size != max_size)
817 unscalarizable_region = true;
821 disqualify_candidate (base, "Encountered an unconstrained access.");
826 access = create_access_1 (base, offset, size);
828 access->type = TREE_TYPE (expr);
829 access->write = write;
830 access->grp_unscalarizable_region = unscalarizable_region;
833 if (TREE_CODE (expr) == COMPONENT_REF
834 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
835 access->non_addressable = 1;
841 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
842 register types or (recursively) records with only these two kinds of fields.
843 It also returns false if any of these records contains a bit-field. */
846 type_consists_of_records_p (tree type)
850 if (TREE_CODE (type) != RECORD_TYPE)
853 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
854 if (TREE_CODE (fld) == FIELD_DECL)
856 tree ft = TREE_TYPE (fld);
858 if (DECL_BIT_FIELD (fld))
861 if (!is_gimple_reg_type (ft)
862 && !type_consists_of_records_p (ft))
869 /* Create total_scalarization accesses for all scalar type fields in DECL that
870 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
871 must be the top-most VAR_DECL representing the variable, OFFSET must be the
872 offset of DECL within BASE. REF must be the memory reference expression for
876 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
879 tree fld, decl_type = TREE_TYPE (decl);
881 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
882 if (TREE_CODE (fld) == FIELD_DECL)
884 HOST_WIDE_INT pos = offset + int_bit_position (fld);
885 tree ft = TREE_TYPE (fld);
886 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
889 if (is_gimple_reg_type (ft))
891 struct access *access;
894 size = tree_low_cst (DECL_SIZE (fld), 1);
895 access = create_access_1 (base, pos, size);
898 access->total_scalarization = 1;
899 /* Accesses for intraprocedural SRA can have their stmt NULL. */
902 completely_scalarize_record (base, fld, pos, nref);
907 /* Search the given tree for a declaration by skipping handled components and
908 exclude it from the candidates. */
911 disqualify_base_of_expr (tree t, const char *reason)
913 t = get_base_address (t);
914 if (sra_mode == SRA_MODE_EARLY_IPA
915 && TREE_CODE (t) == MEM_REF)
916 t = get_ssa_base_param (TREE_OPERAND (t, 0));
919 disqualify_candidate (t, reason);
922 /* Scan expression EXPR and create access structures for all accesses to
923 candidates for scalarization. Return the created access or NULL if none is
926 static struct access *
927 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
929 struct access *ret = NULL;
932 if (TREE_CODE (expr) == BIT_FIELD_REF
933 || TREE_CODE (expr) == IMAGPART_EXPR
934 || TREE_CODE (expr) == REALPART_EXPR)
936 expr = TREE_OPERAND (expr, 0);
942 /* We need to dive through V_C_Es in order to get the size of its parameter
943 and not the result type. Ada produces such statements. We are also
944 capable of handling the topmost V_C_E but not any of those buried in other
945 handled components. */
946 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
947 expr = TREE_OPERAND (expr, 0);
949 if (contains_view_convert_expr_p (expr))
951 disqualify_base_of_expr (expr, "V_C_E under a different handled "
956 switch (TREE_CODE (expr))
959 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
960 && sra_mode != SRA_MODE_EARLY_IPA)
968 case ARRAY_RANGE_REF:
969 ret = create_access (expr, stmt, write);
976 if (write && partial_ref && ret)
977 ret->grp_partial_lhs = 1;
982 /* Scan expression EXPR and create access structures for all accesses to
983 candidates for scalarization. Return true if any access has been inserted.
984 STMT must be the statement from which the expression is taken, WRITE must be
985 true if the expression is a store and false otherwise. */
988 build_access_from_expr (tree expr, gimple stmt, bool write)
990 struct access *access;
992 access = build_access_from_expr_1 (expr, stmt, write);
995 /* This means the aggregate is accesses as a whole in a way other than an
996 assign statement and thus cannot be removed even if we had a scalar
997 replacement for everything. */
998 if (cannot_scalarize_away_bitmap)
999 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1005 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1006 modes in which it matters, return true iff they have been disqualified. RHS
1007 may be NULL, in that case ignore it. If we scalarize an aggregate in
1008 intra-SRA we may need to add statements after each statement. This is not
1009 possible if a statement unconditionally has to end the basic block. */
1011 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1013 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1014 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1016 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1018 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1024 /* Scan expressions occuring in STMT, create access structures for all accesses
1025 to candidates for scalarization and remove those candidates which occur in
1026 statements or expressions that prevent them from being split apart. Return
1027 true if any access has been inserted. */
1030 build_accesses_from_assign (gimple stmt)
1033 struct access *lacc, *racc;
1035 if (!gimple_assign_single_p (stmt))
1038 lhs = gimple_assign_lhs (stmt);
1039 rhs = gimple_assign_rhs1 (stmt);
1041 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1044 racc = build_access_from_expr_1 (rhs, stmt, false);
1045 lacc = build_access_from_expr_1 (lhs, stmt, true);
1048 lacc->grp_assignment_write = 1;
1052 racc->grp_assignment_read = 1;
1053 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1054 && !is_gimple_reg_type (racc->type))
1055 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1059 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1060 && !lacc->grp_unscalarizable_region
1061 && !racc->grp_unscalarizable_region
1062 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1063 /* FIXME: Turn the following line into an assert after PR 40058 is
1065 && lacc->size == racc->size
1066 && useless_type_conversion_p (lacc->type, racc->type))
1068 struct assign_link *link;
1070 link = (struct assign_link *) pool_alloc (link_pool);
1071 memset (link, 0, sizeof (struct assign_link));
1076 add_link_to_rhs (racc, link);
1079 return lacc || racc;
1082 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1083 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1086 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1087 void *data ATTRIBUTE_UNUSED)
1089 op = get_base_address (op);
1092 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1097 /* Return true iff callsite CALL has at least as many actual arguments as there
1098 are formal parameters of the function currently processed by IPA-SRA. */
1101 callsite_has_enough_arguments_p (gimple call)
1103 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1106 /* Scan function and look for interesting expressions and create access
1107 structures for them. Return true iff any access is created. */
1110 scan_function (void)
1117 gimple_stmt_iterator gsi;
1118 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1120 gimple stmt = gsi_stmt (gsi);
1124 if (final_bbs && stmt_can_throw_external (stmt))
1125 bitmap_set_bit (final_bbs, bb->index);
1126 switch (gimple_code (stmt))
1129 t = gimple_return_retval (stmt);
1131 ret |= build_access_from_expr (t, stmt, false);
1133 bitmap_set_bit (final_bbs, bb->index);
1137 ret |= build_accesses_from_assign (stmt);
1141 for (i = 0; i < gimple_call_num_args (stmt); i++)
1142 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1145 if (sra_mode == SRA_MODE_EARLY_IPA)
1147 tree dest = gimple_call_fndecl (stmt);
1148 int flags = gimple_call_flags (stmt);
1152 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1153 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1154 encountered_apply_args = true;
1155 if (cgraph_get_node (dest)
1156 == cgraph_get_node (current_function_decl))
1158 encountered_recursive_call = true;
1159 if (!callsite_has_enough_arguments_p (stmt))
1160 encountered_unchangable_recursive_call = true;
1165 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1166 bitmap_set_bit (final_bbs, bb->index);
1169 t = gimple_call_lhs (stmt);
1170 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1171 ret |= build_access_from_expr (t, stmt, true);
1175 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1178 bitmap_set_bit (final_bbs, bb->index);
1180 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1182 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1183 ret |= build_access_from_expr (t, stmt, false);
1185 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1187 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1188 ret |= build_access_from_expr (t, stmt, true);
1201 /* Helper of QSORT function. There are pointers to accesses in the array. An
1202 access is considered smaller than another if it has smaller offset or if the
1203 offsets are the same but is size is bigger. */
1206 compare_access_positions (const void *a, const void *b)
1208 const access_p *fp1 = (const access_p *) a;
1209 const access_p *fp2 = (const access_p *) b;
1210 const access_p f1 = *fp1;
1211 const access_p f2 = *fp2;
1213 if (f1->offset != f2->offset)
1214 return f1->offset < f2->offset ? -1 : 1;
1216 if (f1->size == f2->size)
1218 if (f1->type == f2->type)
1220 /* Put any non-aggregate type before any aggregate type. */
1221 else if (!is_gimple_reg_type (f1->type)
1222 && is_gimple_reg_type (f2->type))
1224 else if (is_gimple_reg_type (f1->type)
1225 && !is_gimple_reg_type (f2->type))
1227 /* Put any complex or vector type before any other scalar type. */
1228 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1229 && TREE_CODE (f1->type) != VECTOR_TYPE
1230 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1231 || TREE_CODE (f2->type) == VECTOR_TYPE))
1233 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1234 || TREE_CODE (f1->type) == VECTOR_TYPE)
1235 && TREE_CODE (f2->type) != COMPLEX_TYPE
1236 && TREE_CODE (f2->type) != VECTOR_TYPE)
1238 /* Put the integral type with the bigger precision first. */
1239 else if (INTEGRAL_TYPE_P (f1->type)
1240 && INTEGRAL_TYPE_P (f2->type))
1241 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1242 /* Put any integral type with non-full precision last. */
1243 else if (INTEGRAL_TYPE_P (f1->type)
1244 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1245 != TYPE_PRECISION (f1->type)))
1247 else if (INTEGRAL_TYPE_P (f2->type)
1248 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1249 != TYPE_PRECISION (f2->type)))
1251 /* Stabilize the sort. */
1252 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1255 /* We want the bigger accesses first, thus the opposite operator in the next
1257 return f1->size > f2->size ? -1 : 1;
1261 /* Append a name of the declaration to the name obstack. A helper function for
1265 make_fancy_decl_name (tree decl)
1269 tree name = DECL_NAME (decl);
1271 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1272 IDENTIFIER_LENGTH (name));
1275 sprintf (buffer, "D%u", DECL_UID (decl));
1276 obstack_grow (&name_obstack, buffer, strlen (buffer));
1280 /* Helper for make_fancy_name. */
1283 make_fancy_name_1 (tree expr)
1290 make_fancy_decl_name (expr);
1294 switch (TREE_CODE (expr))
1297 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1298 obstack_1grow (&name_obstack, '$');
1299 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1303 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1304 obstack_1grow (&name_obstack, '$');
1305 /* Arrays with only one element may not have a constant as their
1307 index = TREE_OPERAND (expr, 1);
1308 if (TREE_CODE (index) != INTEGER_CST)
1310 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1311 obstack_grow (&name_obstack, buffer, strlen (buffer));
1315 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1319 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1320 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1322 obstack_1grow (&name_obstack, '$');
1323 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1324 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1325 obstack_grow (&name_obstack, buffer, strlen (buffer));
1332 gcc_unreachable (); /* we treat these as scalars. */
1339 /* Create a human readable name for replacement variable of ACCESS. */
1342 make_fancy_name (tree expr)
1344 make_fancy_name_1 (expr);
1345 obstack_1grow (&name_obstack, '\0');
1346 return XOBFINISH (&name_obstack, char *);
1349 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1350 EXP_TYPE at the given OFFSET. If BASE is something for which
1351 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1352 to insert new statements either before or below the current one as specified
1353 by INSERT_AFTER. This function is not capable of handling bitfields. */
1356 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1357 tree exp_type, gimple_stmt_iterator *gsi,
1360 tree prev_base = base;
1362 HOST_WIDE_INT base_offset;
1364 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1366 base = get_addr_base_and_unit_offset (base, &base_offset);
1368 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1369 offset such as array[var_index]. */
1375 gcc_checking_assert (gsi);
1376 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1377 add_referenced_var (tmp);
1378 tmp = make_ssa_name (tmp, NULL);
1379 addr = build_fold_addr_expr (unshare_expr (prev_base));
1380 STRIP_USELESS_TYPE_CONVERSION (addr);
1381 stmt = gimple_build_assign (tmp, addr);
1382 gimple_set_location (stmt, loc);
1383 SSA_NAME_DEF_STMT (tmp) = stmt;
1385 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1387 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1390 off = build_int_cst (reference_alias_ptr_type (prev_base),
1391 offset / BITS_PER_UNIT);
1394 else if (TREE_CODE (base) == MEM_REF)
1396 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1397 base_offset + offset / BITS_PER_UNIT);
1398 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1399 base = unshare_expr (TREE_OPERAND (base, 0));
1403 off = build_int_cst (reference_alias_ptr_type (base),
1404 base_offset + offset / BITS_PER_UNIT);
1405 base = build_fold_addr_expr (unshare_expr (base));
1408 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1411 /* Construct a memory reference to a part of an aggregate BASE at the given
1412 OFFSET and of the same type as MODEL. In case this is a reference to a
1413 component, the function will replicate the last COMPONENT_REF of model's
1414 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1415 build_ref_for_offset. */
1418 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1419 struct access *model, gimple_stmt_iterator *gsi,
1422 if (TREE_CODE (model->expr) == COMPONENT_REF)
1424 tree t, exp_type, fld = TREE_OPERAND (model->expr, 1);
1425 tree cr_offset = component_ref_field_offset (model->expr);
1427 gcc_assert (cr_offset && host_integerp (cr_offset, 1));
1428 offset -= TREE_INT_CST_LOW (cr_offset) * BITS_PER_UNIT;
1429 offset -= TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (fld));
1430 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1431 t = build_ref_for_offset (loc, base, offset, exp_type, gsi, insert_after);
1432 return fold_build3_loc (loc, COMPONENT_REF, model->type, t, fld,
1433 TREE_OPERAND (model->expr, 2));
1436 return build_ref_for_offset (loc, base, offset, model->type,
1440 /* Construct a memory reference consisting of component_refs and array_refs to
1441 a part of an aggregate *RES (which is of type TYPE). The requested part
1442 should have type EXP_TYPE at be the given OFFSET. This function might not
1443 succeed, it returns true when it does and only then *RES points to something
1444 meaningful. This function should be used only to build expressions that we
1445 might need to present to user (e.g. in warnings). In all other situations,
1446 build_ref_for_model or build_ref_for_offset should be used instead. */
1449 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1455 tree tr_size, index, minidx;
1456 HOST_WIDE_INT el_size;
1458 if (offset == 0 && exp_type
1459 && types_compatible_p (exp_type, type))
1462 switch (TREE_CODE (type))
1465 case QUAL_UNION_TYPE:
1467 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1469 HOST_WIDE_INT pos, size;
1470 tree expr, *expr_ptr;
1472 if (TREE_CODE (fld) != FIELD_DECL)
1475 pos = int_bit_position (fld);
1476 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1477 tr_size = DECL_SIZE (fld);
1478 if (!tr_size || !host_integerp (tr_size, 1))
1480 size = tree_low_cst (tr_size, 1);
1486 else if (pos > offset || (pos + size) <= offset)
1489 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1492 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1493 offset - pos, exp_type))
1502 tr_size = TYPE_SIZE (TREE_TYPE (type));
1503 if (!tr_size || !host_integerp (tr_size, 1))
1505 el_size = tree_low_cst (tr_size, 1);
1507 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1508 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1510 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1511 if (!integer_zerop (minidx))
1512 index = int_const_binop (PLUS_EXPR, index, minidx);
1513 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1514 NULL_TREE, NULL_TREE);
1515 offset = offset % el_size;
1516 type = TREE_TYPE (type);
1531 /* Return true iff TYPE is stdarg va_list type. */
1534 is_va_list_type (tree type)
1536 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1539 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1540 those with type which is suitable for scalarization. */
1543 find_var_candidates (void)
1546 referenced_var_iterator rvi;
1549 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
1551 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1553 type = TREE_TYPE (var);
1555 if (!AGGREGATE_TYPE_P (type)
1556 || needs_to_live_in_memory (var)
1557 || TREE_THIS_VOLATILE (var)
1558 || !COMPLETE_TYPE_P (type)
1559 || !host_integerp (TYPE_SIZE (type), 1)
1560 || tree_low_cst (TYPE_SIZE (type), 1) == 0
1561 || type_internals_preclude_sra_p (type)
1562 /* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1563 we also want to schedule it rather late. Thus we ignore it in
1565 || (sra_mode == SRA_MODE_EARLY_INTRA
1566 && is_va_list_type (type)))
1569 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1571 if (dump_file && (dump_flags & TDF_DETAILS))
1573 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1574 print_generic_expr (dump_file, var, 0);
1575 fprintf (dump_file, "\n");
1583 /* Sort all accesses for the given variable, check for partial overlaps and
1584 return NULL if there are any. If there are none, pick a representative for
1585 each combination of offset and size and create a linked list out of them.
1586 Return the pointer to the first representative and make sure it is the first
1587 one in the vector of accesses. */
1589 static struct access *
1590 sort_and_splice_var_accesses (tree var)
1592 int i, j, access_count;
1593 struct access *res, **prev_acc_ptr = &res;
1594 VEC (access_p, heap) *access_vec;
1596 HOST_WIDE_INT low = -1, high = 0;
1598 access_vec = get_base_access_vector (var);
1601 access_count = VEC_length (access_p, access_vec);
1603 /* Sort by <OFFSET, SIZE>. */
1604 VEC_qsort (access_p, access_vec, compare_access_positions);
1607 while (i < access_count)
1609 struct access *access = VEC_index (access_p, access_vec, i);
1610 bool grp_write = access->write;
1611 bool grp_read = !access->write;
1612 bool grp_scalar_write = access->write
1613 && is_gimple_reg_type (access->type);
1614 bool grp_scalar_read = !access->write
1615 && is_gimple_reg_type (access->type);
1616 bool grp_assignment_read = access->grp_assignment_read;
1617 bool grp_assignment_write = access->grp_assignment_write;
1618 bool multiple_scalar_reads = false;
1619 bool total_scalarization = access->total_scalarization;
1620 bool grp_partial_lhs = access->grp_partial_lhs;
1621 bool first_scalar = is_gimple_reg_type (access->type);
1622 bool unscalarizable_region = access->grp_unscalarizable_region;
1624 if (first || access->offset >= high)
1627 low = access->offset;
1628 high = access->offset + access->size;
1630 else if (access->offset > low && access->offset + access->size > high)
1633 gcc_assert (access->offset >= low
1634 && access->offset + access->size <= high);
1637 while (j < access_count)
1639 struct access *ac2 = VEC_index (access_p, access_vec, j);
1640 if (ac2->offset != access->offset || ac2->size != access->size)
1645 grp_scalar_write = (grp_scalar_write
1646 || is_gimple_reg_type (ac2->type));
1651 if (is_gimple_reg_type (ac2->type))
1653 if (grp_scalar_read)
1654 multiple_scalar_reads = true;
1656 grp_scalar_read = true;
1659 grp_assignment_read |= ac2->grp_assignment_read;
1660 grp_assignment_write |= ac2->grp_assignment_write;
1661 grp_partial_lhs |= ac2->grp_partial_lhs;
1662 unscalarizable_region |= ac2->grp_unscalarizable_region;
1663 total_scalarization |= ac2->total_scalarization;
1664 relink_to_new_repr (access, ac2);
1666 /* If there are both aggregate-type and scalar-type accesses with
1667 this combination of size and offset, the comparison function
1668 should have put the scalars first. */
1669 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1670 ac2->group_representative = access;
1676 access->group_representative = access;
1677 access->grp_write = grp_write;
1678 access->grp_read = grp_read;
1679 access->grp_scalar_read = grp_scalar_read;
1680 access->grp_scalar_write = grp_scalar_write;
1681 access->grp_assignment_read = grp_assignment_read;
1682 access->grp_assignment_write = grp_assignment_write;
1683 access->grp_hint = multiple_scalar_reads || total_scalarization;
1684 access->grp_partial_lhs = grp_partial_lhs;
1685 access->grp_unscalarizable_region = unscalarizable_region;
1686 if (access->first_link)
1687 add_access_to_work_queue (access);
1689 *prev_acc_ptr = access;
1690 prev_acc_ptr = &access->next_grp;
1693 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1697 /* Create a variable for the given ACCESS which determines the type, name and a
1698 few other properties. Return the variable declaration and store it also to
1699 ACCESS->replacement. */
1702 create_access_replacement (struct access *access, bool rename)
1706 repl = create_tmp_var (access->type, "SR");
1708 add_referenced_var (repl);
1710 mark_sym_for_renaming (repl);
1712 if (!access->grp_partial_lhs
1713 && (TREE_CODE (access->type) == COMPLEX_TYPE
1714 || TREE_CODE (access->type) == VECTOR_TYPE))
1715 DECL_GIMPLE_REG_P (repl) = 1;
1717 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1718 DECL_ARTIFICIAL (repl) = 1;
1719 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1721 if (DECL_NAME (access->base)
1722 && !DECL_IGNORED_P (access->base)
1723 && !DECL_ARTIFICIAL (access->base))
1725 char *pretty_name = make_fancy_name (access->expr);
1726 tree debug_expr = unshare_expr (access->expr), d;
1728 DECL_NAME (repl) = get_identifier (pretty_name);
1729 obstack_free (&name_obstack, pretty_name);
1731 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1732 as DECL_DEBUG_EXPR isn't considered when looking for still
1733 used SSA_NAMEs and thus they could be freed. All debug info
1734 generation cares is whether something is constant or variable
1735 and that get_ref_base_and_extent works properly on the
1737 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1738 switch (TREE_CODE (d))
1741 case ARRAY_RANGE_REF:
1742 if (TREE_OPERAND (d, 1)
1743 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1744 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1745 if (TREE_OPERAND (d, 3)
1746 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1747 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1750 if (TREE_OPERAND (d, 2)
1751 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1752 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1757 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1758 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1759 if (access->grp_no_warning)
1760 TREE_NO_WARNING (repl) = 1;
1762 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1765 TREE_NO_WARNING (repl) = 1;
1769 fprintf (dump_file, "Created a replacement for ");
1770 print_generic_expr (dump_file, access->base, 0);
1771 fprintf (dump_file, " offset: %u, size: %u: ",
1772 (unsigned) access->offset, (unsigned) access->size);
1773 print_generic_expr (dump_file, repl, 0);
1774 fprintf (dump_file, "\n");
1776 sra_stats.replacements++;
1781 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1784 get_access_replacement (struct access *access)
1786 gcc_assert (access->grp_to_be_replaced);
1788 if (!access->replacement_decl)
1789 access->replacement_decl = create_access_replacement (access, true);
1790 return access->replacement_decl;
1793 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1794 not mark it for renaming. */
1797 get_unrenamed_access_replacement (struct access *access)
1799 gcc_assert (!access->grp_to_be_replaced);
1801 if (!access->replacement_decl)
1802 access->replacement_decl = create_access_replacement (access, false);
1803 return access->replacement_decl;
1807 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1808 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1809 to it is not "within" the root. Return false iff some accesses partially
1813 build_access_subtree (struct access **access)
1815 struct access *root = *access, *last_child = NULL;
1816 HOST_WIDE_INT limit = root->offset + root->size;
1818 *access = (*access)->next_grp;
1819 while (*access && (*access)->offset + (*access)->size <= limit)
1822 root->first_child = *access;
1824 last_child->next_sibling = *access;
1825 last_child = *access;
1827 if (!build_access_subtree (access))
1831 if (*access && (*access)->offset < limit)
1837 /* Build a tree of access representatives, ACCESS is the pointer to the first
1838 one, others are linked in a list by the next_grp field. Return false iff
1839 some accesses partially overlap. */
1842 build_access_trees (struct access *access)
1846 struct access *root = access;
1848 if (!build_access_subtree (&access))
1850 root->next_grp = access;
1855 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1859 expr_with_var_bounded_array_refs_p (tree expr)
1861 while (handled_component_p (expr))
1863 if (TREE_CODE (expr) == ARRAY_REF
1864 && !host_integerp (array_ref_low_bound (expr), 0))
1866 expr = TREE_OPERAND (expr, 0);
1871 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1872 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1873 sorts of access flags appropriately along the way, notably always set
1874 grp_read and grp_assign_read according to MARK_READ and grp_write when
1877 Creating a replacement for a scalar access is considered beneficial if its
1878 grp_hint is set (this means we are either attempting total scalarization or
1879 there is more than one direct read access) or according to the following
1882 Access written to through a scalar type (once or more times)
1884 | Written to in an assignment statement
1886 | | Access read as scalar _once_
1888 | | | Read in an assignment statement
1890 | | | | Scalarize Comment
1891 -----------------------------------------------------------------------------
1892 0 0 0 0 No access for the scalar
1893 0 0 0 1 No access for the scalar
1894 0 0 1 0 No Single read - won't help
1895 0 0 1 1 No The same case
1896 0 1 0 0 No access for the scalar
1897 0 1 0 1 No access for the scalar
1898 0 1 1 0 Yes s = *g; return s.i;
1899 0 1 1 1 Yes The same case as above
1900 1 0 0 0 No Won't help
1901 1 0 0 1 Yes s.i = 1; *g = s;
1902 1 0 1 0 Yes s.i = 5; g = s.i;
1903 1 0 1 1 Yes The same case as above
1904 1 1 0 0 No Won't help.
1905 1 1 0 1 Yes s.i = 1; *g = s;
1906 1 1 1 0 Yes s = *g; return s.i;
1907 1 1 1 1 Yes Any of the above yeses */
1910 analyze_access_subtree (struct access *root, struct access *parent,
1911 bool allow_replacements)
1913 struct access *child;
1914 HOST_WIDE_INT limit = root->offset + root->size;
1915 HOST_WIDE_INT covered_to = root->offset;
1916 bool scalar = is_gimple_reg_type (root->type);
1917 bool hole = false, sth_created = false;
1921 if (parent->grp_read)
1923 if (parent->grp_assignment_read)
1924 root->grp_assignment_read = 1;
1925 if (parent->grp_write)
1926 root->grp_write = 1;
1927 if (parent->grp_assignment_write)
1928 root->grp_assignment_write = 1;
1931 if (root->grp_unscalarizable_region)
1932 allow_replacements = false;
1934 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
1935 allow_replacements = false;
1937 for (child = root->first_child; child; child = child->next_sibling)
1939 if (!hole && child->offset < covered_to)
1942 covered_to += child->size;
1944 sth_created |= analyze_access_subtree (child, root,
1945 allow_replacements && !scalar);
1947 root->grp_unscalarized_data |= child->grp_unscalarized_data;
1948 hole |= !child->grp_covered;
1951 if (allow_replacements && scalar && !root->first_child
1953 || ((root->grp_scalar_read || root->grp_assignment_read)
1954 && (root->grp_scalar_write || root->grp_assignment_write))))
1956 if (dump_file && (dump_flags & TDF_DETAILS))
1958 fprintf (dump_file, "Marking ");
1959 print_generic_expr (dump_file, root->base, 0);
1960 fprintf (dump_file, " offset: %u, size: %u: ",
1961 (unsigned) root->offset, (unsigned) root->size);
1962 fprintf (dump_file, " to be replaced.\n");
1965 root->grp_to_be_replaced = 1;
1969 else if (covered_to < limit)
1972 if (sth_created && !hole)
1974 root->grp_covered = 1;
1977 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
1978 root->grp_unscalarized_data = 1; /* not covered and written to */
1984 /* Analyze all access trees linked by next_grp by the means of
1985 analyze_access_subtree. */
1987 analyze_access_trees (struct access *access)
1993 if (analyze_access_subtree (access, NULL, true))
1995 access = access->next_grp;
2001 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2002 SIZE would conflict with an already existing one. If exactly such a child
2003 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2006 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2007 HOST_WIDE_INT size, struct access **exact_match)
2009 struct access *child;
2011 for (child = lacc->first_child; child; child = child->next_sibling)
2013 if (child->offset == norm_offset && child->size == size)
2015 *exact_match = child;
2019 if (child->offset < norm_offset + size
2020 && child->offset + child->size > norm_offset)
2027 /* Create a new child access of PARENT, with all properties just like MODEL
2028 except for its offset and with its grp_write false and grp_read true.
2029 Return the new access or NULL if it cannot be created. Note that this access
2030 is created long after all splicing and sorting, it's not located in any
2031 access vector and is automatically a representative of its group. */
2033 static struct access *
2034 create_artificial_child_access (struct access *parent, struct access *model,
2035 HOST_WIDE_INT new_offset)
2037 struct access *access;
2038 struct access **child;
2039 tree expr = parent->base;
2041 gcc_assert (!model->grp_unscalarizable_region);
2043 access = (struct access *) pool_alloc (access_pool);
2044 memset (access, 0, sizeof (struct access));
2045 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2048 access->grp_no_warning = true;
2049 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2050 new_offset, model, NULL, false);
2053 access->base = parent->base;
2054 access->expr = expr;
2055 access->offset = new_offset;
2056 access->size = model->size;
2057 access->type = model->type;
2058 access->grp_write = true;
2059 access->grp_read = false;
2061 child = &parent->first_child;
2062 while (*child && (*child)->offset < new_offset)
2063 child = &(*child)->next_sibling;
2065 access->next_sibling = *child;
2072 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2073 true if any new subaccess was created. Additionally, if RACC is a scalar
2074 access but LACC is not, change the type of the latter, if possible. */
2077 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2079 struct access *rchild;
2080 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2083 if (is_gimple_reg_type (lacc->type)
2084 || lacc->grp_unscalarizable_region
2085 || racc->grp_unscalarizable_region)
2088 if (!lacc->first_child && !racc->first_child
2089 && is_gimple_reg_type (racc->type))
2091 tree t = lacc->base;
2093 lacc->type = racc->type;
2094 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t), lacc->offset,
2099 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2100 lacc->base, lacc->offset,
2102 lacc->grp_no_warning = true;
2107 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2109 struct access *new_acc = NULL;
2110 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2112 if (rchild->grp_unscalarizable_region)
2115 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2120 rchild->grp_hint = 1;
2121 new_acc->grp_hint |= new_acc->grp_read;
2122 if (rchild->first_child)
2123 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2128 rchild->grp_hint = 1;
2129 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2133 if (racc->first_child)
2134 propagate_subaccesses_across_link (new_acc, rchild);
2141 /* Propagate all subaccesses across assignment links. */
2144 propagate_all_subaccesses (void)
2146 while (work_queue_head)
2148 struct access *racc = pop_access_from_work_queue ();
2149 struct assign_link *link;
2151 gcc_assert (racc->first_link);
2153 for (link = racc->first_link; link; link = link->next)
2155 struct access *lacc = link->lacc;
2157 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2159 lacc = lacc->group_representative;
2160 if (propagate_subaccesses_across_link (lacc, racc)
2161 && lacc->first_link)
2162 add_access_to_work_queue (lacc);
2167 /* Go through all accesses collected throughout the (intraprocedural) analysis
2168 stage, exclude overlapping ones, identify representatives and build trees
2169 out of them, making decisions about scalarization on the way. Return true
2170 iff there are any to-be-scalarized variables after this stage. */
2173 analyze_all_variable_accesses (void)
2176 bitmap tmp = BITMAP_ALLOC (NULL);
2178 unsigned i, max_total_scalarization_size;
2180 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2181 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2183 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2184 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2185 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2187 tree var = referenced_var (i);
2189 if (TREE_CODE (var) == VAR_DECL
2190 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2191 <= max_total_scalarization_size)
2192 && type_consists_of_records_p (TREE_TYPE (var)))
2194 completely_scalarize_record (var, var, 0, var);
2195 if (dump_file && (dump_flags & TDF_DETAILS))
2197 fprintf (dump_file, "Will attempt to totally scalarize ");
2198 print_generic_expr (dump_file, var, 0);
2199 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2204 bitmap_copy (tmp, candidate_bitmap);
2205 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2207 tree var = referenced_var (i);
2208 struct access *access;
2210 access = sort_and_splice_var_accesses (var);
2211 if (!access || !build_access_trees (access))
2212 disqualify_candidate (var,
2213 "No or inhibitingly overlapping accesses.");
2216 propagate_all_subaccesses ();
2218 bitmap_copy (tmp, candidate_bitmap);
2219 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2221 tree var = referenced_var (i);
2222 struct access *access = get_first_repr_for_decl (var);
2224 if (analyze_access_trees (access))
2227 if (dump_file && (dump_flags & TDF_DETAILS))
2229 fprintf (dump_file, "\nAccess trees for ");
2230 print_generic_expr (dump_file, var, 0);
2231 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2232 dump_access_tree (dump_file, access);
2233 fprintf (dump_file, "\n");
2237 disqualify_candidate (var, "No scalar replacements to be created.");
2244 statistics_counter_event (cfun, "Scalarized aggregates", res);
2251 /* Generate statements copying scalar replacements of accesses within a subtree
2252 into or out of AGG. ACCESS, all its children, siblings and their children
2253 are to be processed. AGG is an aggregate type expression (can be a
2254 declaration but does not have to be, it can for example also be a mem_ref or
2255 a series of handled components). TOP_OFFSET is the offset of the processed
2256 subtree which has to be subtracted from offsets of individual accesses to
2257 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2258 replacements in the interval <start_offset, start_offset + chunk_size>,
2259 otherwise copy all. GSI is a statement iterator used to place the new
2260 statements. WRITE should be true when the statements should write from AGG
2261 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2262 statements will be added after the current statement in GSI, they will be
2263 added before the statement otherwise. */
2266 generate_subtree_copies (struct access *access, tree agg,
2267 HOST_WIDE_INT top_offset,
2268 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2269 gimple_stmt_iterator *gsi, bool write,
2270 bool insert_after, location_t loc)
2274 if (chunk_size && access->offset >= start_offset + chunk_size)
2277 if (access->grp_to_be_replaced
2279 || access->offset + access->size > start_offset))
2281 tree expr, repl = get_access_replacement (access);
2284 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2285 access, gsi, insert_after);
2289 if (access->grp_partial_lhs)
2290 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2292 insert_after ? GSI_NEW_STMT
2294 stmt = gimple_build_assign (repl, expr);
2298 TREE_NO_WARNING (repl) = 1;
2299 if (access->grp_partial_lhs)
2300 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2302 insert_after ? GSI_NEW_STMT
2304 stmt = gimple_build_assign (expr, repl);
2306 gimple_set_location (stmt, loc);
2309 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2311 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2313 sra_stats.subtree_copies++;
2316 if (access->first_child)
2317 generate_subtree_copies (access->first_child, agg, top_offset,
2318 start_offset, chunk_size, gsi,
2319 write, insert_after, loc);
2321 access = access->next_sibling;
2326 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2327 the root of the subtree to be processed. GSI is the statement iterator used
2328 for inserting statements which are added after the current statement if
2329 INSERT_AFTER is true or before it otherwise. */
2332 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2333 bool insert_after, location_t loc)
2336 struct access *child;
2338 if (access->grp_to_be_replaced)
2342 stmt = gimple_build_assign (get_access_replacement (access),
2343 build_zero_cst (access->type));
2345 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2347 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2349 gimple_set_location (stmt, loc);
2352 for (child = access->first_child; child; child = child->next_sibling)
2353 init_subtree_with_zero (child, gsi, insert_after, loc);
2356 /* Search for an access representative for the given expression EXPR and
2357 return it or NULL if it cannot be found. */
2359 static struct access *
2360 get_access_for_expr (tree expr)
2362 HOST_WIDE_INT offset, size, max_size;
2365 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2366 a different size than the size of its argument and we need the latter
2368 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2369 expr = TREE_OPERAND (expr, 0);
2371 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2372 if (max_size == -1 || !DECL_P (base))
2375 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2378 return get_var_base_offset_size_access (base, offset, max_size);
2381 /* Replace the expression EXPR with a scalar replacement if there is one and
2382 generate other statements to do type conversion or subtree copying if
2383 necessary. GSI is used to place newly created statements, WRITE is true if
2384 the expression is being written to (it is on a LHS of a statement or output
2385 in an assembly statement). */
2388 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2391 struct access *access;
2394 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2397 expr = &TREE_OPERAND (*expr, 0);
2402 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2403 expr = &TREE_OPERAND (*expr, 0);
2404 access = get_access_for_expr (*expr);
2407 type = TREE_TYPE (*expr);
2409 loc = gimple_location (gsi_stmt (*gsi));
2410 if (access->grp_to_be_replaced)
2412 tree repl = get_access_replacement (access);
2413 /* If we replace a non-register typed access simply use the original
2414 access expression to extract the scalar component afterwards.
2415 This happens if scalarizing a function return value or parameter
2416 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2417 gcc.c-torture/compile/20011217-1.c.
2419 We also want to use this when accessing a complex or vector which can
2420 be accessed as a different type too, potentially creating a need for
2421 type conversion (see PR42196) and when scalarized unions are involved
2422 in assembler statements (see PR42398). */
2423 if (!useless_type_conversion_p (type, access->type))
2427 ref = build_ref_for_model (loc, access->base, access->offset, access,
2434 if (access->grp_partial_lhs)
2435 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2436 false, GSI_NEW_STMT);
2437 stmt = gimple_build_assign (repl, ref);
2438 gimple_set_location (stmt, loc);
2439 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2445 if (access->grp_partial_lhs)
2446 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2447 true, GSI_SAME_STMT);
2448 stmt = gimple_build_assign (ref, repl);
2449 gimple_set_location (stmt, loc);
2450 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2458 if (access->first_child)
2460 HOST_WIDE_INT start_offset, chunk_size;
2462 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2463 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2465 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2466 start_offset = access->offset
2467 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2470 start_offset = chunk_size = 0;
2472 generate_subtree_copies (access->first_child, access->base, 0,
2473 start_offset, chunk_size, gsi, write, write,
2479 /* Where scalar replacements of the RHS have been written to when a replacement
2480 of a LHS of an assigments cannot be direclty loaded from a replacement of
2482 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2483 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2484 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2486 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2487 base aggregate if there are unscalarized data or directly to LHS of the
2488 statement that is pointed to by GSI otherwise. */
2490 static enum unscalarized_data_handling
2491 handle_unscalarized_data_in_subtree (struct access *top_racc,
2492 gimple_stmt_iterator *gsi)
2494 if (top_racc->grp_unscalarized_data)
2496 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2498 gimple_location (gsi_stmt (*gsi)));
2499 return SRA_UDH_RIGHT;
2503 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2504 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2505 0, 0, gsi, false, false,
2506 gimple_location (gsi_stmt (*gsi)));
2507 return SRA_UDH_LEFT;
2512 /* Try to generate statements to load all sub-replacements in an access subtree
2513 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2514 If that is not possible, refresh the TOP_RACC base aggregate and load the
2515 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2516 copied. NEW_GSI is stmt iterator used for statement insertions after the
2517 original assignment, OLD_GSI is used to insert statements before the
2518 assignment. *REFRESHED keeps the information whether we have needed to
2519 refresh replacements of the LHS and from which side of the assignments this
2523 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2524 HOST_WIDE_INT left_offset,
2525 gimple_stmt_iterator *old_gsi,
2526 gimple_stmt_iterator *new_gsi,
2527 enum unscalarized_data_handling *refreshed)
2529 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2530 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2532 if (lacc->grp_to_be_replaced)
2534 struct access *racc;
2535 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2539 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2540 if (racc && racc->grp_to_be_replaced)
2542 rhs = get_access_replacement (racc);
2543 if (!useless_type_conversion_p (lacc->type, racc->type))
2544 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2548 /* No suitable access on the right hand side, need to load from
2549 the aggregate. See if we have to update it first... */
2550 if (*refreshed == SRA_UDH_NONE)
2551 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2554 if (*refreshed == SRA_UDH_LEFT)
2555 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2558 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2562 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2563 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2564 gimple_set_location (stmt, loc);
2566 sra_stats.subreplacements++;
2568 else if (*refreshed == SRA_UDH_NONE
2569 && lacc->grp_read && !lacc->grp_covered)
2570 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2573 if (lacc->first_child)
2574 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2575 old_gsi, new_gsi, refreshed);
2579 /* Result code for SRA assignment modification. */
2580 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2581 SRA_AM_MODIFIED, /* stmt changed but not
2583 SRA_AM_REMOVED }; /* stmt eliminated */
2585 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2586 to the assignment and GSI is the statement iterator pointing at it. Returns
2587 the same values as sra_modify_assign. */
2589 static enum assignment_mod_result
2590 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2592 tree lhs = gimple_assign_lhs (*stmt);
2596 acc = get_access_for_expr (lhs);
2600 loc = gimple_location (*stmt);
2601 if (VEC_length (constructor_elt,
2602 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2604 /* I have never seen this code path trigger but if it can happen the
2605 following should handle it gracefully. */
2606 if (access_has_children_p (acc))
2607 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2609 return SRA_AM_MODIFIED;
2612 if (acc->grp_covered)
2614 init_subtree_with_zero (acc, gsi, false, loc);
2615 unlink_stmt_vdef (*stmt);
2616 gsi_remove (gsi, true);
2617 return SRA_AM_REMOVED;
2621 init_subtree_with_zero (acc, gsi, true, loc);
2622 return SRA_AM_MODIFIED;
2626 /* Create and return a new suitable default definition SSA_NAME for RACC which
2627 is an access describing an uninitialized part of an aggregate that is being
2631 get_repl_default_def_ssa_name (struct access *racc)
2635 decl = get_unrenamed_access_replacement (racc);
2637 repl = gimple_default_def (cfun, decl);
2640 repl = make_ssa_name (decl, gimple_build_nop ());
2641 set_default_def (decl, repl);
2647 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2651 contains_bitfld_comp_ref_p (const_tree ref)
2653 while (handled_component_p (ref))
2655 if (TREE_CODE (ref) == COMPONENT_REF
2656 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2658 ref = TREE_OPERAND (ref, 0);
2664 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2665 bit-field field declaration somewhere in it. */
2668 contains_vce_or_bfcref_p (const_tree ref)
2670 while (handled_component_p (ref))
2672 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2673 || (TREE_CODE (ref) == COMPONENT_REF
2674 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2676 ref = TREE_OPERAND (ref, 0);
2682 /* Examine both sides of the assignment statement pointed to by STMT, replace
2683 them with a scalare replacement if there is one and generate copying of
2684 replacements if scalarized aggregates have been used in the assignment. GSI
2685 is used to hold generated statements for type conversions and subtree
2688 static enum assignment_mod_result
2689 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2691 struct access *lacc, *racc;
2693 bool modify_this_stmt = false;
2694 bool force_gimple_rhs = false;
2696 gimple_stmt_iterator orig_gsi = *gsi;
2698 if (!gimple_assign_single_p (*stmt))
2700 lhs = gimple_assign_lhs (*stmt);
2701 rhs = gimple_assign_rhs1 (*stmt);
2703 if (TREE_CODE (rhs) == CONSTRUCTOR)
2704 return sra_modify_constructor_assign (stmt, gsi);
2706 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2707 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2708 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2710 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2712 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2714 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2717 lacc = get_access_for_expr (lhs);
2718 racc = get_access_for_expr (rhs);
2722 loc = gimple_location (*stmt);
2723 if (lacc && lacc->grp_to_be_replaced)
2725 lhs = get_access_replacement (lacc);
2726 gimple_assign_set_lhs (*stmt, lhs);
2727 modify_this_stmt = true;
2728 if (lacc->grp_partial_lhs)
2729 force_gimple_rhs = true;
2733 if (racc && racc->grp_to_be_replaced)
2735 rhs = get_access_replacement (racc);
2736 modify_this_stmt = true;
2737 if (racc->grp_partial_lhs)
2738 force_gimple_rhs = true;
2742 if (modify_this_stmt)
2744 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2746 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2747 ??? This should move to fold_stmt which we simply should
2748 call after building a VIEW_CONVERT_EXPR here. */
2749 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2750 && !contains_bitfld_comp_ref_p (lhs)
2751 && !access_has_children_p (lacc))
2753 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
2754 gimple_assign_set_lhs (*stmt, lhs);
2756 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2757 && !contains_vce_or_bfcref_p (rhs)
2758 && !access_has_children_p (racc))
2759 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
2761 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2763 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
2765 if (is_gimple_reg_type (TREE_TYPE (lhs))
2766 && TREE_CODE (lhs) != SSA_NAME)
2767 force_gimple_rhs = true;
2772 /* From this point on, the function deals with assignments in between
2773 aggregates when at least one has scalar reductions of some of its
2774 components. There are three possible scenarios: Both the LHS and RHS have
2775 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2777 In the first case, we would like to load the LHS components from RHS
2778 components whenever possible. If that is not possible, we would like to
2779 read it directly from the RHS (after updating it by storing in it its own
2780 components). If there are some necessary unscalarized data in the LHS,
2781 those will be loaded by the original assignment too. If neither of these
2782 cases happen, the original statement can be removed. Most of this is done
2783 by load_assign_lhs_subreplacements.
2785 In the second case, we would like to store all RHS scalarized components
2786 directly into LHS and if they cover the aggregate completely, remove the
2787 statement too. In the third case, we want the LHS components to be loaded
2788 directly from the RHS (DSE will remove the original statement if it
2791 This is a bit complex but manageable when types match and when unions do
2792 not cause confusion in a way that we cannot really load a component of LHS
2793 from the RHS or vice versa (the access representing this level can have
2794 subaccesses that are accessible only through a different union field at a
2795 higher level - different from the one used in the examined expression).
2798 Therefore, I specially handle a fourth case, happening when there is a
2799 specific type cast or it is impossible to locate a scalarized subaccess on
2800 the other side of the expression. If that happens, I simply "refresh" the
2801 RHS by storing in it is scalarized components leave the original statement
2802 there to do the copying and then load the scalar replacements of the LHS.
2803 This is what the first branch does. */
2805 if (gimple_has_volatile_ops (*stmt)
2806 || contains_vce_or_bfcref_p (rhs)
2807 || contains_vce_or_bfcref_p (lhs))
2809 if (access_has_children_p (racc))
2810 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2811 gsi, false, false, loc);
2812 if (access_has_children_p (lacc))
2813 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2814 gsi, true, true, loc);
2815 sra_stats.separate_lhs_rhs_handling++;
2819 if (access_has_children_p (lacc) && access_has_children_p (racc))
2821 gimple_stmt_iterator orig_gsi = *gsi;
2822 enum unscalarized_data_handling refreshed;
2824 if (lacc->grp_read && !lacc->grp_covered)
2825 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
2827 refreshed = SRA_UDH_NONE;
2829 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
2830 &orig_gsi, gsi, &refreshed);
2831 if (refreshed != SRA_UDH_RIGHT)
2834 unlink_stmt_vdef (*stmt);
2835 gsi_remove (&orig_gsi, true);
2836 sra_stats.deleted++;
2837 return SRA_AM_REMOVED;
2844 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2848 fprintf (dump_file, "Removing load: ");
2849 print_gimple_stmt (dump_file, *stmt, 0, 0);
2852 if (TREE_CODE (lhs) == SSA_NAME)
2854 rhs = get_repl_default_def_ssa_name (racc);
2855 if (!useless_type_conversion_p (TREE_TYPE (lhs),
2857 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR,
2858 TREE_TYPE (lhs), rhs);
2862 if (racc->first_child)
2863 generate_subtree_copies (racc->first_child, lhs,
2864 racc->offset, 0, 0, gsi,
2867 gcc_assert (*stmt == gsi_stmt (*gsi));
2868 unlink_stmt_vdef (*stmt);
2869 gsi_remove (gsi, true);
2870 sra_stats.deleted++;
2871 return SRA_AM_REMOVED;
2874 else if (racc->first_child)
2875 generate_subtree_copies (racc->first_child, lhs, racc->offset,
2876 0, 0, gsi, false, true, loc);
2878 if (access_has_children_p (lacc))
2879 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2880 0, 0, gsi, true, true, loc);
2884 /* This gimplification must be done after generate_subtree_copies, lest we
2885 insert the subtree copies in the middle of the gimplified sequence. */
2886 if (force_gimple_rhs)
2887 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2888 true, GSI_SAME_STMT);
2889 if (gimple_assign_rhs1 (*stmt) != rhs)
2891 modify_this_stmt = true;
2892 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2893 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2896 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2899 /* Traverse the function body and all modifications as decided in
2900 analyze_all_variable_accesses. Return true iff the CFG has been
2904 sra_modify_function_body (void)
2906 bool cfg_changed = false;
2911 gimple_stmt_iterator gsi = gsi_start_bb (bb);
2912 while (!gsi_end_p (gsi))
2914 gimple stmt = gsi_stmt (gsi);
2915 enum assignment_mod_result assign_result;
2916 bool modified = false, deleted = false;
2920 switch (gimple_code (stmt))
2923 t = gimple_return_retval_ptr (stmt);
2924 if (*t != NULL_TREE)
2925 modified |= sra_modify_expr (t, &gsi, false);
2929 assign_result = sra_modify_assign (&stmt, &gsi);
2930 modified |= assign_result == SRA_AM_MODIFIED;
2931 deleted = assign_result == SRA_AM_REMOVED;
2935 /* Operands must be processed before the lhs. */
2936 for (i = 0; i < gimple_call_num_args (stmt); i++)
2938 t = gimple_call_arg_ptr (stmt, i);
2939 modified |= sra_modify_expr (t, &gsi, false);
2942 if (gimple_call_lhs (stmt))
2944 t = gimple_call_lhs_ptr (stmt);
2945 modified |= sra_modify_expr (t, &gsi, true);
2950 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
2952 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
2953 modified |= sra_modify_expr (t, &gsi, false);
2955 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
2957 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
2958 modified |= sra_modify_expr (t, &gsi, true);
2969 if (maybe_clean_eh_stmt (stmt)
2970 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
2981 /* Generate statements initializing scalar replacements of parts of function
2985 initialize_parameter_reductions (void)
2987 gimple_stmt_iterator gsi;
2988 gimple_seq seq = NULL;
2991 for (parm = DECL_ARGUMENTS (current_function_decl);
2993 parm = DECL_CHAIN (parm))
2995 VEC (access_p, heap) *access_vec;
2996 struct access *access;
2998 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3000 access_vec = get_base_access_vector (parm);
3006 seq = gimple_seq_alloc ();
3007 gsi = gsi_start (seq);
3010 for (access = VEC_index (access_p, access_vec, 0);
3012 access = access->next_grp)
3013 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3014 EXPR_LOCATION (parm));
3018 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3021 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3022 it reveals there are components of some aggregates to be scalarized, it runs
3023 the required transformations. */
3025 perform_intra_sra (void)
3030 if (!find_var_candidates ())
3033 if (!scan_function ())
3036 if (!analyze_all_variable_accesses ())
3039 if (sra_modify_function_body ())
3040 ret = TODO_update_ssa | TODO_cleanup_cfg;
3042 ret = TODO_update_ssa;
3043 initialize_parameter_reductions ();
3045 statistics_counter_event (cfun, "Scalar replacements created",
3046 sra_stats.replacements);
3047 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3048 statistics_counter_event (cfun, "Subtree copy stmts",
3049 sra_stats.subtree_copies);
3050 statistics_counter_event (cfun, "Subreplacement stmts",
3051 sra_stats.subreplacements);
3052 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3053 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3054 sra_stats.separate_lhs_rhs_handling);
3057 sra_deinitialize ();
3061 /* Perform early intraprocedural SRA. */
3063 early_intra_sra (void)
3065 sra_mode = SRA_MODE_EARLY_INTRA;
3066 return perform_intra_sra ();
3069 /* Perform "late" intraprocedural SRA. */
3071 late_intra_sra (void)
3073 sra_mode = SRA_MODE_INTRA;
3074 return perform_intra_sra ();
3079 gate_intra_sra (void)
3081 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3085 struct gimple_opt_pass pass_sra_early =
3090 gate_intra_sra, /* gate */
3091 early_intra_sra, /* execute */
3094 0, /* static_pass_number */
3095 TV_TREE_SRA, /* tv_id */
3096 PROP_cfg | PROP_ssa, /* properties_required */
3097 0, /* properties_provided */
3098 0, /* properties_destroyed */
3099 0, /* todo_flags_start */
3102 | TODO_verify_ssa /* todo_flags_finish */
3106 struct gimple_opt_pass pass_sra =
3111 gate_intra_sra, /* gate */
3112 late_intra_sra, /* execute */
3115 0, /* static_pass_number */
3116 TV_TREE_SRA, /* tv_id */
3117 PROP_cfg | PROP_ssa, /* properties_required */
3118 0, /* properties_provided */
3119 0, /* properties_destroyed */
3120 TODO_update_address_taken, /* todo_flags_start */
3123 | TODO_verify_ssa /* todo_flags_finish */
3128 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3132 is_unused_scalar_param (tree parm)
3135 return (is_gimple_reg (parm)
3136 && (!(name = gimple_default_def (cfun, parm))
3137 || has_zero_uses (name)));
3140 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3141 examine whether there are any direct or otherwise infeasible ones. If so,
3142 return true, otherwise return false. PARM must be a gimple register with a
3143 non-NULL default definition. */
3146 ptr_parm_has_direct_uses (tree parm)
3148 imm_use_iterator ui;
3150 tree name = gimple_default_def (cfun, parm);
3153 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3156 use_operand_p use_p;
3158 if (is_gimple_debug (stmt))
3161 /* Valid uses include dereferences on the lhs and the rhs. */
3162 if (gimple_has_lhs (stmt))
3164 tree lhs = gimple_get_lhs (stmt);
3165 while (handled_component_p (lhs))
3166 lhs = TREE_OPERAND (lhs, 0);
3167 if (TREE_CODE (lhs) == MEM_REF
3168 && TREE_OPERAND (lhs, 0) == name
3169 && integer_zerop (TREE_OPERAND (lhs, 1))
3170 && types_compatible_p (TREE_TYPE (lhs),
3171 TREE_TYPE (TREE_TYPE (name))))
3174 if (gimple_assign_single_p (stmt))
3176 tree rhs = gimple_assign_rhs1 (stmt);
3177 while (handled_component_p (rhs))
3178 rhs = TREE_OPERAND (rhs, 0);
3179 if (TREE_CODE (rhs) == MEM_REF
3180 && TREE_OPERAND (rhs, 0) == name
3181 && integer_zerop (TREE_OPERAND (rhs, 1))
3182 && types_compatible_p (TREE_TYPE (rhs),
3183 TREE_TYPE (TREE_TYPE (name))))
3186 else if (is_gimple_call (stmt))
3189 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3191 tree arg = gimple_call_arg (stmt, i);
3192 while (handled_component_p (arg))
3193 arg = TREE_OPERAND (arg, 0);
3194 if (TREE_CODE (arg) == MEM_REF
3195 && TREE_OPERAND (arg, 0) == name
3196 && integer_zerop (TREE_OPERAND (arg, 1))
3197 && types_compatible_p (TREE_TYPE (arg),
3198 TREE_TYPE (TREE_TYPE (name))))
3203 /* If the number of valid uses does not match the number of
3204 uses in this stmt there is an unhandled use. */
3205 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3212 BREAK_FROM_IMM_USE_STMT (ui);
3218 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3219 them in candidate_bitmap. Note that these do not necessarily include
3220 parameter which are unused and thus can be removed. Return true iff any
3221 such candidate has been found. */
3224 find_param_candidates (void)
3230 for (parm = DECL_ARGUMENTS (current_function_decl);
3232 parm = DECL_CHAIN (parm))
3234 tree type = TREE_TYPE (parm);
3238 if (TREE_THIS_VOLATILE (parm)
3239 || TREE_ADDRESSABLE (parm)
3240 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3243 if (is_unused_scalar_param (parm))
3249 if (POINTER_TYPE_P (type))
3251 type = TREE_TYPE (type);
3253 if (TREE_CODE (type) == FUNCTION_TYPE
3254 || TYPE_VOLATILE (type)
3255 || (TREE_CODE (type) == ARRAY_TYPE
3256 && TYPE_NONALIASED_COMPONENT (type))
3257 || !is_gimple_reg (parm)
3258 || is_va_list_type (type)
3259 || ptr_parm_has_direct_uses (parm))
3262 else if (!AGGREGATE_TYPE_P (type))
3265 if (!COMPLETE_TYPE_P (type)
3266 || !host_integerp (TYPE_SIZE (type), 1)
3267 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3268 || (AGGREGATE_TYPE_P (type)
3269 && type_internals_preclude_sra_p (type)))
3272 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3274 if (dump_file && (dump_flags & TDF_DETAILS))
3276 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3277 print_generic_expr (dump_file, parm, 0);
3278 fprintf (dump_file, "\n");
3282 func_param_count = count;
3286 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3290 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3293 struct access *repr = (struct access *) data;
3295 repr->grp_maybe_modified = 1;
3299 /* Analyze what representatives (in linked lists accessible from
3300 REPRESENTATIVES) can be modified by side effects of statements in the
3301 current function. */
3304 analyze_modified_params (VEC (access_p, heap) *representatives)
3308 for (i = 0; i < func_param_count; i++)
3310 struct access *repr;
3312 for (repr = VEC_index (access_p, representatives, i);
3314 repr = repr->next_grp)
3316 struct access *access;
3320 if (no_accesses_p (repr))
3322 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3323 || repr->grp_maybe_modified)
3326 ao_ref_init (&ar, repr->expr);
3327 visited = BITMAP_ALLOC (NULL);
3328 for (access = repr; access; access = access->next_sibling)
3330 /* All accesses are read ones, otherwise grp_maybe_modified would
3331 be trivially set. */
3332 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3333 mark_maybe_modified, repr, &visited);
3334 if (repr->grp_maybe_modified)
3337 BITMAP_FREE (visited);
3342 /* Propagate distances in bb_dereferences in the opposite direction than the
3343 control flow edges, in each step storing the maximum of the current value
3344 and the minimum of all successors. These steps are repeated until the table
3345 stabilizes. Note that BBs which might terminate the functions (according to
3346 final_bbs bitmap) never updated in this way. */
3349 propagate_dereference_distances (void)
3351 VEC (basic_block, heap) *queue;
3354 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3355 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3358 VEC_quick_push (basic_block, queue, bb);
3362 while (!VEC_empty (basic_block, queue))
3366 bool change = false;
3369 bb = VEC_pop (basic_block, queue);
3372 if (bitmap_bit_p (final_bbs, bb->index))
3375 for (i = 0; i < func_param_count; i++)
3377 int idx = bb->index * func_param_count + i;
3379 HOST_WIDE_INT inh = 0;
3381 FOR_EACH_EDGE (e, ei, bb->succs)
3383 int succ_idx = e->dest->index * func_param_count + i;
3385 if (e->src == EXIT_BLOCK_PTR)
3391 inh = bb_dereferences [succ_idx];
3393 else if (bb_dereferences [succ_idx] < inh)
3394 inh = bb_dereferences [succ_idx];
3397 if (!first && bb_dereferences[idx] < inh)
3399 bb_dereferences[idx] = inh;
3404 if (change && !bitmap_bit_p (final_bbs, bb->index))
3405 FOR_EACH_EDGE (e, ei, bb->preds)
3410 e->src->aux = e->src;
3411 VEC_quick_push (basic_block, queue, e->src);
3415 VEC_free (basic_block, heap, queue);
3418 /* Dump a dereferences TABLE with heading STR to file F. */
3421 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3425 fprintf (dump_file, str);
3426 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3428 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3429 if (bb != EXIT_BLOCK_PTR)
3432 for (i = 0; i < func_param_count; i++)
3434 int idx = bb->index * func_param_count + i;
3435 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3440 fprintf (dump_file, "\n");
3443 /* Determine what (parts of) parameters passed by reference that are not
3444 assigned to are not certainly dereferenced in this function and thus the
3445 dereferencing cannot be safely moved to the caller without potentially
3446 introducing a segfault. Mark such REPRESENTATIVES as
3447 grp_not_necessarilly_dereferenced.
3449 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3450 part is calculated rather than simple booleans are calculated for each
3451 pointer parameter to handle cases when only a fraction of the whole
3452 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3455 The maximum dereference distances for each pointer parameter and BB are
3456 already stored in bb_dereference. This routine simply propagates these
3457 values upwards by propagate_dereference_distances and then compares the
3458 distances of individual parameters in the ENTRY BB to the equivalent
3459 distances of each representative of a (fraction of a) parameter. */
3462 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3466 if (dump_file && (dump_flags & TDF_DETAILS))
3467 dump_dereferences_table (dump_file,
3468 "Dereference table before propagation:\n",
3471 propagate_dereference_distances ();
3473 if (dump_file && (dump_flags & TDF_DETAILS))
3474 dump_dereferences_table (dump_file,
3475 "Dereference table after propagation:\n",
3478 for (i = 0; i < func_param_count; i++)
3480 struct access *repr = VEC_index (access_p, representatives, i);
3481 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3483 if (!repr || no_accesses_p (repr))
3488 if ((repr->offset + repr->size) > bb_dereferences[idx])
3489 repr->grp_not_necessarilly_dereferenced = 1;
3490 repr = repr->next_grp;
3496 /* Return the representative access for the parameter declaration PARM if it is
3497 a scalar passed by reference which is not written to and the pointer value
3498 is not used directly. Thus, if it is legal to dereference it in the caller
3499 and we can rule out modifications through aliases, such parameter should be
3500 turned into one passed by value. Return NULL otherwise. */
3502 static struct access *
3503 unmodified_by_ref_scalar_representative (tree parm)
3505 int i, access_count;
3506 struct access *repr;
3507 VEC (access_p, heap) *access_vec;
3509 access_vec = get_base_access_vector (parm);
3510 gcc_assert (access_vec);
3511 repr = VEC_index (access_p, access_vec, 0);
3514 repr->group_representative = repr;
3516 access_count = VEC_length (access_p, access_vec);
3517 for (i = 1; i < access_count; i++)
3519 struct access *access = VEC_index (access_p, access_vec, i);
3522 access->group_representative = repr;
3523 access->next_sibling = repr->next_sibling;
3524 repr->next_sibling = access;
3528 repr->grp_scalar_ptr = 1;
3532 /* Return true iff this access precludes IPA-SRA of the parameter it is
3536 access_precludes_ipa_sra_p (struct access *access)
3538 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3539 is incompatible assign in a call statement (and possibly even in asm
3540 statements). This can be relaxed by using a new temporary but only for
3541 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3542 intraprocedural SRA we deal with this by keeping the old aggregate around,
3543 something we cannot do in IPA-SRA.) */
3545 && (is_gimple_call (access->stmt)
3546 || gimple_code (access->stmt) == GIMPLE_ASM))
3553 /* Sort collected accesses for parameter PARM, identify representatives for
3554 each accessed region and link them together. Return NULL if there are
3555 different but overlapping accesses, return the special ptr value meaning
3556 there are no accesses for this parameter if that is the case and return the
3557 first representative otherwise. Set *RO_GRP if there is a group of accesses
3558 with only read (i.e. no write) accesses. */
3560 static struct access *
3561 splice_param_accesses (tree parm, bool *ro_grp)
3563 int i, j, access_count, group_count;
3564 int agg_size, total_size = 0;
3565 struct access *access, *res, **prev_acc_ptr = &res;
3566 VEC (access_p, heap) *access_vec;
3568 access_vec = get_base_access_vector (parm);
3570 return &no_accesses_representant;
3571 access_count = VEC_length (access_p, access_vec);
3573 VEC_qsort (access_p, access_vec, compare_access_positions);
3578 while (i < access_count)
3582 access = VEC_index (access_p, access_vec, i);
3583 modification = access->write;
3584 if (access_precludes_ipa_sra_p (access))
3586 a1_alias_type = reference_alias_ptr_type (access->expr);
3588 /* Access is about to become group representative unless we find some
3589 nasty overlap which would preclude us from breaking this parameter
3593 while (j < access_count)
3595 struct access *ac2 = VEC_index (access_p, access_vec, j);
3596 if (ac2->offset != access->offset)
3598 /* All or nothing law for parameters. */
3599 if (access->offset + access->size > ac2->offset)
3604 else if (ac2->size != access->size)
3607 if (access_precludes_ipa_sra_p (ac2)
3608 || (ac2->type != access->type
3609 && (TREE_ADDRESSABLE (ac2->type)
3610 || TREE_ADDRESSABLE (access->type)))
3611 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3614 modification |= ac2->write;
3615 ac2->group_representative = access;
3616 ac2->next_sibling = access->next_sibling;
3617 access->next_sibling = ac2;
3622 access->grp_maybe_modified = modification;
3625 *prev_acc_ptr = access;
3626 prev_acc_ptr = &access->next_grp;
3627 total_size += access->size;
3631 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3632 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3634 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3635 if (total_size >= agg_size)
3638 gcc_assert (group_count > 0);
3642 /* Decide whether parameters with representative accesses given by REPR should
3643 be reduced into components. */
3646 decide_one_param_reduction (struct access *repr)
3648 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3653 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3654 gcc_assert (cur_parm_size > 0);
3656 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3659 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3664 agg_size = cur_parm_size;
3670 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3671 print_generic_expr (dump_file, parm, 0);
3672 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3673 for (acc = repr; acc; acc = acc->next_grp)
3674 dump_access (dump_file, acc, true);
3678 new_param_count = 0;
3680 for (; repr; repr = repr->next_grp)
3682 gcc_assert (parm == repr->base);
3684 /* Taking the address of a non-addressable field is verboten. */
3685 if (by_ref && repr->non_addressable)
3688 if (!by_ref || (!repr->grp_maybe_modified
3689 && !repr->grp_not_necessarilly_dereferenced))
3690 total_size += repr->size;
3692 total_size += cur_parm_size;
3697 gcc_assert (new_param_count > 0);
3699 if (optimize_function_for_size_p (cfun))
3700 parm_size_limit = cur_parm_size;
3702 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3705 if (total_size < agg_size
3706 && total_size <= parm_size_limit)
3709 fprintf (dump_file, " ....will be split into %i components\n",
3711 return new_param_count;
3717 /* The order of the following enums is important, we need to do extra work for
3718 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3719 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3720 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3722 /* Identify representatives of all accesses to all candidate parameters for
3723 IPA-SRA. Return result based on what representatives have been found. */
3725 static enum ipa_splicing_result
3726 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3728 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3730 struct access *repr;
3732 *representatives = VEC_alloc (access_p, heap, func_param_count);
3734 for (parm = DECL_ARGUMENTS (current_function_decl);
3736 parm = DECL_CHAIN (parm))
3738 if (is_unused_scalar_param (parm))
3740 VEC_quick_push (access_p, *representatives,
3741 &no_accesses_representant);
3742 if (result == NO_GOOD_ACCESS)
3743 result = UNUSED_PARAMS;
3745 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3746 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3747 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3749 repr = unmodified_by_ref_scalar_representative (parm);
3750 VEC_quick_push (access_p, *representatives, repr);
3752 result = UNMODIF_BY_REF_ACCESSES;
3754 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3756 bool ro_grp = false;
3757 repr = splice_param_accesses (parm, &ro_grp);
3758 VEC_quick_push (access_p, *representatives, repr);
3760 if (repr && !no_accesses_p (repr))
3762 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3765 result = UNMODIF_BY_REF_ACCESSES;
3766 else if (result < MODIF_BY_REF_ACCESSES)
3767 result = MODIF_BY_REF_ACCESSES;
3769 else if (result < BY_VAL_ACCESSES)
3770 result = BY_VAL_ACCESSES;
3772 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3773 result = UNUSED_PARAMS;
3776 VEC_quick_push (access_p, *representatives, NULL);
3779 if (result == NO_GOOD_ACCESS)
3781 VEC_free (access_p, heap, *representatives);
3782 *representatives = NULL;
3783 return NO_GOOD_ACCESS;
3789 /* Return the index of BASE in PARMS. Abort if it is not found. */
3792 get_param_index (tree base, VEC(tree, heap) *parms)
3796 len = VEC_length (tree, parms);
3797 for (i = 0; i < len; i++)
3798 if (VEC_index (tree, parms, i) == base)
3803 /* Convert the decisions made at the representative level into compact
3804 parameter adjustments. REPRESENTATIVES are pointers to first
3805 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3806 final number of adjustments. */
3808 static ipa_parm_adjustment_vec
3809 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3810 int adjustments_count)
3812 VEC (tree, heap) *parms;
3813 ipa_parm_adjustment_vec adjustments;
3817 gcc_assert (adjustments_count > 0);
3818 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3819 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3820 parm = DECL_ARGUMENTS (current_function_decl);
3821 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
3823 struct access *repr = VEC_index (access_p, representatives, i);
3825 if (!repr || no_accesses_p (repr))
3827 struct ipa_parm_adjustment *adj;
3829 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3830 memset (adj, 0, sizeof (*adj));
3831 adj->base_index = get_param_index (parm, parms);
3834 adj->copy_param = 1;
3836 adj->remove_param = 1;
3840 struct ipa_parm_adjustment *adj;
3841 int index = get_param_index (parm, parms);
3843 for (; repr; repr = repr->next_grp)
3845 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3846 memset (adj, 0, sizeof (*adj));
3847 gcc_assert (repr->base == parm);
3848 adj->base_index = index;
3849 adj->base = repr->base;
3850 adj->type = repr->type;
3851 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
3852 adj->offset = repr->offset;
3853 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3854 && (repr->grp_maybe_modified
3855 || repr->grp_not_necessarilly_dereferenced));
3860 VEC_free (tree, heap, parms);
3864 /* Analyze the collected accesses and produce a plan what to do with the
3865 parameters in the form of adjustments, NULL meaning nothing. */
3867 static ipa_parm_adjustment_vec
3868 analyze_all_param_acesses (void)
3870 enum ipa_splicing_result repr_state;
3871 bool proceed = false;
3872 int i, adjustments_count = 0;
3873 VEC (access_p, heap) *representatives;
3874 ipa_parm_adjustment_vec adjustments;
3876 repr_state = splice_all_param_accesses (&representatives);
3877 if (repr_state == NO_GOOD_ACCESS)
3880 /* If there are any parameters passed by reference which are not modified
3881 directly, we need to check whether they can be modified indirectly. */
3882 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3884 analyze_caller_dereference_legality (representatives);
3885 analyze_modified_params (representatives);
3888 for (i = 0; i < func_param_count; i++)
3890 struct access *repr = VEC_index (access_p, representatives, i);
3892 if (repr && !no_accesses_p (repr))
3894 if (repr->grp_scalar_ptr)
3896 adjustments_count++;
3897 if (repr->grp_not_necessarilly_dereferenced
3898 || repr->grp_maybe_modified)
3899 VEC_replace (access_p, representatives, i, NULL);
3903 sra_stats.scalar_by_ref_to_by_val++;
3908 int new_components = decide_one_param_reduction (repr);
3910 if (new_components == 0)
3912 VEC_replace (access_p, representatives, i, NULL);
3913 adjustments_count++;
3917 adjustments_count += new_components;
3918 sra_stats.aggregate_params_reduced++;
3919 sra_stats.param_reductions_created += new_components;
3926 if (no_accesses_p (repr))
3929 sra_stats.deleted_unused_parameters++;
3931 adjustments_count++;
3935 if (!proceed && dump_file)
3936 fprintf (dump_file, "NOT proceeding to change params.\n");
3939 adjustments = turn_representatives_into_adjustments (representatives,
3944 VEC_free (access_p, heap, representatives);
3948 /* If a parameter replacement identified by ADJ does not yet exist in the form
3949 of declaration, create it and record it, otherwise return the previously
3953 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
3956 if (!adj->new_ssa_base)
3958 char *pretty_name = make_fancy_name (adj->base);
3960 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
3961 DECL_NAME (repl) = get_identifier (pretty_name);
3962 obstack_free (&name_obstack, pretty_name);
3965 add_referenced_var (repl);
3966 adj->new_ssa_base = repl;
3969 repl = adj->new_ssa_base;
3973 /* Find the first adjustment for a particular parameter BASE in a vector of
3974 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
3977 static struct ipa_parm_adjustment *
3978 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
3982 len = VEC_length (ipa_parm_adjustment_t, adjustments);
3983 for (i = 0; i < len; i++)
3985 struct ipa_parm_adjustment *adj;
3987 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
3988 if (!adj->copy_param && adj->base == base)
3995 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
3996 removed because its value is not used, replace the SSA_NAME with a one
3997 relating to a created VAR_DECL together all of its uses and return true.
3998 ADJUSTMENTS is a pointer to an adjustments vector. */
4001 replace_removed_params_ssa_names (gimple stmt,
4002 ipa_parm_adjustment_vec adjustments)
4004 struct ipa_parm_adjustment *adj;
4005 tree lhs, decl, repl, name;
4007 if (gimple_code (stmt) == GIMPLE_PHI)
4008 lhs = gimple_phi_result (stmt);
4009 else if (is_gimple_assign (stmt))
4010 lhs = gimple_assign_lhs (stmt);
4011 else if (is_gimple_call (stmt))
4012 lhs = gimple_call_lhs (stmt);
4016 if (TREE_CODE (lhs) != SSA_NAME)
4018 decl = SSA_NAME_VAR (lhs);
4019 if (TREE_CODE (decl) != PARM_DECL)
4022 adj = get_adjustment_for_base (adjustments, decl);
4026 repl = get_replaced_param_substitute (adj);
4027 name = make_ssa_name (repl, stmt);
4031 fprintf (dump_file, "replacing an SSA name of a removed param ");
4032 print_generic_expr (dump_file, lhs, 0);
4033 fprintf (dump_file, " with ");
4034 print_generic_expr (dump_file, name, 0);
4035 fprintf (dump_file, "\n");
4038 if (is_gimple_assign (stmt))
4039 gimple_assign_set_lhs (stmt, name);
4040 else if (is_gimple_call (stmt))
4041 gimple_call_set_lhs (stmt, name);
4043 gimple_phi_set_result (stmt, name);
4045 replace_uses_by (lhs, name);
4046 release_ssa_name (lhs);
4050 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4051 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4052 specifies whether the function should care about type incompatibility the
4053 current and new expressions. If it is false, the function will leave
4054 incompatibility issues to the caller. Return true iff the expression
4058 sra_ipa_modify_expr (tree *expr, bool convert,
4059 ipa_parm_adjustment_vec adjustments)
4062 struct ipa_parm_adjustment *adj, *cand = NULL;
4063 HOST_WIDE_INT offset, size, max_size;
4066 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4068 if (TREE_CODE (*expr) == BIT_FIELD_REF
4069 || TREE_CODE (*expr) == IMAGPART_EXPR
4070 || TREE_CODE (*expr) == REALPART_EXPR)
4072 expr = &TREE_OPERAND (*expr, 0);
4076 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4077 if (!base || size == -1 || max_size == -1)
4080 if (TREE_CODE (base) == MEM_REF)
4082 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4083 base = TREE_OPERAND (base, 0);
4086 base = get_ssa_base_param (base);
4087 if (!base || TREE_CODE (base) != PARM_DECL)
4090 for (i = 0; i < len; i++)
4092 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4094 if (adj->base == base &&
4095 (adj->offset == offset || adj->remove_param))
4101 if (!cand || cand->copy_param || cand->remove_param)
4105 src = build_simple_mem_ref (cand->reduction);
4107 src = cand->reduction;
4109 if (dump_file && (dump_flags & TDF_DETAILS))
4111 fprintf (dump_file, "About to replace expr ");
4112 print_generic_expr (dump_file, *expr, 0);
4113 fprintf (dump_file, " with ");
4114 print_generic_expr (dump_file, src, 0);
4115 fprintf (dump_file, "\n");
4118 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4120 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4128 /* If the statement pointed to by STMT_PTR contains any expressions that need
4129 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4130 potential type incompatibilities (GSI is used to accommodate conversion
4131 statements and must point to the statement). Return true iff the statement
4135 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4136 ipa_parm_adjustment_vec adjustments)
4138 gimple stmt = *stmt_ptr;
4139 tree *lhs_p, *rhs_p;
4142 if (!gimple_assign_single_p (stmt))
4145 rhs_p = gimple_assign_rhs1_ptr (stmt);
4146 lhs_p = gimple_assign_lhs_ptr (stmt);
4148 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4149 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4152 tree new_rhs = NULL_TREE;
4154 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4156 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4158 /* V_C_Es of constructors can cause trouble (PR 42714). */
4159 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4160 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4162 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4165 new_rhs = fold_build1_loc (gimple_location (stmt),
4166 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4169 else if (REFERENCE_CLASS_P (*rhs_p)
4170 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4171 && !is_gimple_reg (*lhs_p))
4172 /* This can happen when an assignment in between two single field
4173 structures is turned into an assignment in between two pointers to
4174 scalars (PR 42237). */
4179 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4180 true, GSI_SAME_STMT);
4182 gimple_assign_set_rhs_from_tree (gsi, tmp);
4191 /* Traverse the function body and all modifications as described in
4192 ADJUSTMENTS. Return true iff the CFG has been changed. */
4195 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4197 bool cfg_changed = false;
4202 gimple_stmt_iterator gsi;
4204 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4205 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4207 gsi = gsi_start_bb (bb);
4208 while (!gsi_end_p (gsi))
4210 gimple stmt = gsi_stmt (gsi);
4211 bool modified = false;
4215 switch (gimple_code (stmt))
4218 t = gimple_return_retval_ptr (stmt);
4219 if (*t != NULL_TREE)
4220 modified |= sra_ipa_modify_expr (t, true, adjustments);
4224 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4225 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4229 /* Operands must be processed before the lhs. */
4230 for (i = 0; i < gimple_call_num_args (stmt); i++)
4232 t = gimple_call_arg_ptr (stmt, i);
4233 modified |= sra_ipa_modify_expr (t, true, adjustments);
4236 if (gimple_call_lhs (stmt))
4238 t = gimple_call_lhs_ptr (stmt);
4239 modified |= sra_ipa_modify_expr (t, false, adjustments);
4240 modified |= replace_removed_params_ssa_names (stmt,
4246 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4248 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4249 modified |= sra_ipa_modify_expr (t, true, adjustments);
4251 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4253 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4254 modified |= sra_ipa_modify_expr (t, false, adjustments);
4265 if (maybe_clean_eh_stmt (stmt)
4266 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4276 /* Call gimple_debug_bind_reset_value on all debug statements describing
4277 gimple register parameters that are being removed or replaced. */
4280 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4284 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4285 for (i = 0; i < len; i++)
4287 struct ipa_parm_adjustment *adj;
4288 imm_use_iterator ui;
4292 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4293 if (adj->copy_param || !is_gimple_reg (adj->base))
4295 name = gimple_default_def (cfun, adj->base);
4298 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4300 /* All other users must have been removed by
4301 ipa_sra_modify_function_body. */
4302 gcc_assert (is_gimple_debug (stmt));
4303 gimple_debug_bind_reset_value (stmt);
4309 /* Return false iff all callers have at least as many actual arguments as there
4310 are formal parameters in the current function. */
4313 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4314 void *data ATTRIBUTE_UNUSED)
4316 struct cgraph_edge *cs;
4317 for (cs = node->callers; cs; cs = cs->next_caller)
4318 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4324 /* Convert all callers of NODE. */
4327 convert_callers_for_node (struct cgraph_node *node,
4330 ipa_parm_adjustment_vec adjustments = (ipa_parm_adjustment_vec)data;
4331 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4332 struct cgraph_edge *cs;
4334 for (cs = node->callers; cs; cs = cs->next_caller)
4336 current_function_decl = cs->caller->decl;
4337 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4340 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4341 cs->caller->uid, cs->callee->uid,
4342 cgraph_node_name (cs->caller),
4343 cgraph_node_name (cs->callee));
4345 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4350 for (cs = node->callers; cs; cs = cs->next_caller)
4351 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4352 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4353 compute_inline_parameters (cs->caller, true);
4354 BITMAP_FREE (recomputed_callers);
4359 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4362 convert_callers (struct cgraph_node *node, tree old_decl,
4363 ipa_parm_adjustment_vec adjustments)
4365 tree old_cur_fndecl = current_function_decl;
4366 basic_block this_block;
4368 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4369 adjustments, false);
4371 current_function_decl = old_cur_fndecl;
4373 if (!encountered_recursive_call)
4376 FOR_EACH_BB (this_block)
4378 gimple_stmt_iterator gsi;
4380 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4382 gimple stmt = gsi_stmt (gsi);
4384 if (gimple_code (stmt) != GIMPLE_CALL)
4386 call_fndecl = gimple_call_fndecl (stmt);
4387 if (call_fndecl == old_decl)
4390 fprintf (dump_file, "Adjusting recursive call");
4391 gimple_call_set_fndecl (stmt, node->decl);
4392 ipa_modify_call_arguments (NULL, stmt, adjustments);
4400 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4401 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4404 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4406 struct cgraph_node *new_node;
4408 VEC (cgraph_edge_p, heap) * redirect_callers = collect_callers_of_node (node);
4410 rebuild_cgraph_edges ();
4412 current_function_decl = NULL_TREE;
4414 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4415 NULL, NULL, "isra");
4416 current_function_decl = new_node->decl;
4417 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4419 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4420 cfg_changed = ipa_sra_modify_function_body (adjustments);
4421 sra_ipa_reset_debug_stmts (adjustments);
4422 convert_callers (new_node, node->decl, adjustments);
4423 cgraph_make_node_local (new_node);
4427 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4428 attributes, return true otherwise. NODE is the cgraph node of the current
4432 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4434 if (!cgraph_node_can_be_local_p (node))
4437 fprintf (dump_file, "Function not local to this compilation unit.\n");
4441 if (!node->local.can_change_signature)
4444 fprintf (dump_file, "Function can not change signature.\n");
4448 if (!tree_versionable_function_p (node->decl))
4451 fprintf (dump_file, "Function is not versionable.\n");
4455 if (DECL_VIRTUAL_P (current_function_decl))
4458 fprintf (dump_file, "Function is a virtual method.\n");
4462 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4463 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4466 fprintf (dump_file, "Function too big to be made truly local.\n");
4474 "Function has no callers in this compilation unit.\n");
4481 fprintf (dump_file, "Function uses stdarg. \n");
4485 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4491 /* Perform early interprocedural SRA. */
4494 ipa_early_sra (void)
4496 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4497 ipa_parm_adjustment_vec adjustments;
4500 if (!ipa_sra_preliminary_function_checks (node))
4504 sra_mode = SRA_MODE_EARLY_IPA;
4506 if (!find_param_candidates ())
4509 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4513 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4517 fprintf (dump_file, "There are callers with insufficient number of "
4522 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4524 * last_basic_block_for_function (cfun));
4525 final_bbs = BITMAP_ALLOC (NULL);
4528 if (encountered_apply_args)
4531 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4535 if (encountered_unchangable_recursive_call)
4538 fprintf (dump_file, "Function calls itself with insufficient "
4539 "number of arguments.\n");
4543 adjustments = analyze_all_param_acesses ();
4547 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4549 if (modify_function (node, adjustments))
4550 ret = TODO_update_ssa | TODO_cleanup_cfg;
4552 ret = TODO_update_ssa;
4553 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4555 statistics_counter_event (cfun, "Unused parameters deleted",
4556 sra_stats.deleted_unused_parameters);
4557 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4558 sra_stats.scalar_by_ref_to_by_val);
4559 statistics_counter_event (cfun, "Aggregate parameters broken up",
4560 sra_stats.aggregate_params_reduced);
4561 statistics_counter_event (cfun, "Aggregate parameter components created",
4562 sra_stats.param_reductions_created);
4565 BITMAP_FREE (final_bbs);
4566 free (bb_dereferences);
4568 sra_deinitialize ();
4572 /* Return if early ipa sra shall be performed. */
4574 ipa_early_sra_gate (void)
4576 return flag_ipa_sra && dbg_cnt (eipa_sra);
4579 struct gimple_opt_pass pass_early_ipa_sra =
4583 "eipa_sra", /* name */
4584 ipa_early_sra_gate, /* gate */
4585 ipa_early_sra, /* execute */
4588 0, /* static_pass_number */
4589 TV_IPA_SRA, /* tv_id */
4590 0, /* properties_required */
4591 0, /* properties_provided */
4592 0, /* properties_destroyed */
4593 0, /* todo_flags_start */
4594 TODO_dump_cgraph /* todo_flags_finish */