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, 2011 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 access to a non-addressable field? */
174 unsigned non_addressable : 1;
176 /* Is this access currently in the work queue? */
177 unsigned grp_queued : 1;
179 /* Does this group contain a write access? This flag is propagated down the
181 unsigned grp_write : 1;
183 /* Does this group contain a read access? This flag is propagated down the
185 unsigned grp_read : 1;
187 /* Does this group contain a read access that comes from an assignment
188 statement? This flag is propagated down the access tree. */
189 unsigned grp_assignment_read : 1;
191 /* Does this group contain a write access that comes from an assignment
192 statement? This flag is propagated down the access tree. */
193 unsigned grp_assignment_write : 1;
195 /* Does this group contain a read access through a scalar type? This flag is
196 not propagated in the access tree in any direction. */
197 unsigned grp_scalar_read : 1;
199 /* Does this group contain a write access through a scalar type? This flag
200 is not propagated in the access tree in any direction. */
201 unsigned grp_scalar_write : 1;
203 /* Is this access an artificial one created to scalarize some record
205 unsigned grp_total_scalarization : 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, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
381 "grp_assignment_write = %d, grp_scalar_read = %d, "
382 "grp_scalar_write = %d, grp_total_scalarization = %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->grp_read, access->grp_write, access->grp_assignment_read,
389 access->grp_assignment_write, access->grp_scalar_read,
390 access->grp_scalar_write, access->grp_total_scalarization,
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, grp_total_scalarization = %d, "
398 "grp_partial_lhs = %d\n",
399 access->write, access->grp_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, const char **msg)
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))
668 *msg = "volatile structure field";
671 if (!DECL_FIELD_OFFSET (fld))
673 *msg = "no structure field offset";
676 if (!DECL_SIZE (fld))
678 *msg = "zero structure field size";
681 if (!host_integerp (DECL_FIELD_OFFSET (fld), 1))
683 *msg = "structure field offset not fixed";
686 if (!host_integerp (DECL_SIZE (fld), 1))
688 *msg = "structure field size not fixed";
691 if (AGGREGATE_TYPE_P (ft)
692 && int_bit_position (fld) % BITS_PER_UNIT != 0)
694 *msg = "structure field is bit field";
698 if (AGGREGATE_TYPE_P (ft) && type_internals_preclude_sra_p (ft, msg))
705 et = TREE_TYPE (type);
707 if (TYPE_VOLATILE (et))
709 *msg = "element type is volatile";
713 if (AGGREGATE_TYPE_P (et) && type_internals_preclude_sra_p (et, msg))
723 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
724 base variable if it is. Return T if it is not an SSA_NAME. */
727 get_ssa_base_param (tree t)
729 if (TREE_CODE (t) == SSA_NAME)
731 if (SSA_NAME_IS_DEFAULT_DEF (t))
732 return SSA_NAME_VAR (t);
739 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
740 belongs to, unless the BB has already been marked as a potentially
744 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
746 basic_block bb = gimple_bb (stmt);
747 int idx, parm_index = 0;
750 if (bitmap_bit_p (final_bbs, bb->index))
753 for (parm = DECL_ARGUMENTS (current_function_decl);
754 parm && parm != base;
755 parm = DECL_CHAIN (parm))
758 gcc_assert (parm_index < func_param_count);
760 idx = bb->index * func_param_count + parm_index;
761 if (bb_dereferences[idx] < dist)
762 bb_dereferences[idx] = dist;
765 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
766 the three fields. Also add it to the vector of accesses corresponding to
767 the base. Finally, return the new access. */
769 static struct access *
770 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
772 VEC (access_p, heap) *vec;
773 struct access *access;
776 access = (struct access *) pool_alloc (access_pool);
777 memset (access, 0, sizeof (struct access));
779 access->offset = offset;
782 slot = pointer_map_contains (base_access_vec, base);
784 vec = (VEC (access_p, heap) *) *slot;
786 vec = VEC_alloc (access_p, heap, 32);
788 VEC_safe_push (access_p, heap, vec, access);
790 *((struct VEC (access_p,heap) **)
791 pointer_map_insert (base_access_vec, base)) = vec;
796 /* Create and insert access for EXPR. Return created access, or NULL if it is
799 static struct access *
800 create_access (tree expr, gimple stmt, bool write)
802 struct access *access;
803 HOST_WIDE_INT offset, size, max_size;
805 bool ptr, unscalarizable_region = false;
807 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
809 if (sra_mode == SRA_MODE_EARLY_IPA
810 && TREE_CODE (base) == MEM_REF)
812 base = get_ssa_base_param (TREE_OPERAND (base, 0));
820 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
823 if (sra_mode == SRA_MODE_EARLY_IPA)
825 if (size < 0 || size != max_size)
827 disqualify_candidate (base, "Encountered a variable sized access.");
830 if (TREE_CODE (expr) == COMPONENT_REF
831 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
833 disqualify_candidate (base, "Encountered a bit-field access.");
836 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
839 mark_parm_dereference (base, offset + size, stmt);
843 if (size != max_size)
846 unscalarizable_region = true;
850 disqualify_candidate (base, "Encountered an unconstrained access.");
855 access = create_access_1 (base, offset, size);
857 access->type = TREE_TYPE (expr);
858 access->write = write;
859 access->grp_unscalarizable_region = unscalarizable_region;
862 if (TREE_CODE (expr) == COMPONENT_REF
863 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
864 access->non_addressable = 1;
870 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
871 register types or (recursively) records with only these two kinds of fields.
872 It also returns false if any of these records contains a bit-field. */
875 type_consists_of_records_p (tree type)
879 if (TREE_CODE (type) != RECORD_TYPE)
882 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
883 if (TREE_CODE (fld) == FIELD_DECL)
885 tree ft = TREE_TYPE (fld);
887 if (DECL_BIT_FIELD (fld))
890 if (!is_gimple_reg_type (ft)
891 && !type_consists_of_records_p (ft))
898 /* Create total_scalarization accesses for all scalar type fields in DECL that
899 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
900 must be the top-most VAR_DECL representing the variable, OFFSET must be the
901 offset of DECL within BASE. REF must be the memory reference expression for
905 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
908 tree fld, decl_type = TREE_TYPE (decl);
910 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
911 if (TREE_CODE (fld) == FIELD_DECL)
913 HOST_WIDE_INT pos = offset + int_bit_position (fld);
914 tree ft = TREE_TYPE (fld);
915 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
918 if (is_gimple_reg_type (ft))
920 struct access *access;
923 size = tree_low_cst (DECL_SIZE (fld), 1);
924 access = create_access_1 (base, pos, size);
927 access->grp_total_scalarization = 1;
928 /* Accesses for intraprocedural SRA can have their stmt NULL. */
931 completely_scalarize_record (base, fld, pos, nref);
935 /* Create total_scalarization accesses for all scalar type fields in VAR and
936 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
937 type_consists_of_records_p. */
940 completely_scalarize_var (tree var)
942 HOST_WIDE_INT size = tree_low_cst (DECL_SIZE (var), 1);
943 struct access *access;
945 access = create_access_1 (var, 0, size);
947 access->type = TREE_TYPE (var);
948 access->grp_total_scalarization = 1;
950 completely_scalarize_record (var, var, 0, var);
953 /* Search the given tree for a declaration by skipping handled components and
954 exclude it from the candidates. */
957 disqualify_base_of_expr (tree t, const char *reason)
959 t = get_base_address (t);
960 if (sra_mode == SRA_MODE_EARLY_IPA
961 && TREE_CODE (t) == MEM_REF)
962 t = get_ssa_base_param (TREE_OPERAND (t, 0));
965 disqualify_candidate (t, reason);
968 /* Scan expression EXPR and create access structures for all accesses to
969 candidates for scalarization. Return the created access or NULL if none is
972 static struct access *
973 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
975 struct access *ret = NULL;
978 if (TREE_CODE (expr) == BIT_FIELD_REF
979 || TREE_CODE (expr) == IMAGPART_EXPR
980 || TREE_CODE (expr) == REALPART_EXPR)
982 expr = TREE_OPERAND (expr, 0);
988 /* We need to dive through V_C_Es in order to get the size of its parameter
989 and not the result type. Ada produces such statements. We are also
990 capable of handling the topmost V_C_E but not any of those buried in other
991 handled components. */
992 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
993 expr = TREE_OPERAND (expr, 0);
995 if (contains_view_convert_expr_p (expr))
997 disqualify_base_of_expr (expr, "V_C_E under a different handled "
1002 switch (TREE_CODE (expr))
1005 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
1006 && sra_mode != SRA_MODE_EARLY_IPA)
1014 case ARRAY_RANGE_REF:
1015 ret = create_access (expr, stmt, write);
1022 if (write && partial_ref && ret)
1023 ret->grp_partial_lhs = 1;
1028 /* Scan expression EXPR and create access structures for all accesses to
1029 candidates for scalarization. Return true if any access has been inserted.
1030 STMT must be the statement from which the expression is taken, WRITE must be
1031 true if the expression is a store and false otherwise. */
1034 build_access_from_expr (tree expr, gimple stmt, bool write)
1036 struct access *access;
1038 access = build_access_from_expr_1 (expr, stmt, write);
1041 /* This means the aggregate is accesses as a whole in a way other than an
1042 assign statement and thus cannot be removed even if we had a scalar
1043 replacement for everything. */
1044 if (cannot_scalarize_away_bitmap)
1045 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1051 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1052 modes in which it matters, return true iff they have been disqualified. RHS
1053 may be NULL, in that case ignore it. If we scalarize an aggregate in
1054 intra-SRA we may need to add statements after each statement. This is not
1055 possible if a statement unconditionally has to end the basic block. */
1057 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1059 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1060 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1062 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1064 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1070 /* Return true if EXP is a memory reference less aligned than ALIGN. This is
1071 invoked only on strict-alignment targets. */
1074 tree_non_aligned_mem_p (tree exp, unsigned int align)
1076 unsigned int exp_align;
1078 if (TREE_CODE (exp) == VIEW_CONVERT_EXPR)
1079 exp = TREE_OPERAND (exp, 0);
1081 if (TREE_CODE (exp) == SSA_NAME || is_gimple_min_invariant (exp))
1084 /* get_object_alignment will fall back to BITS_PER_UNIT if it cannot
1085 compute an explicit alignment. Pretend that dereferenced pointers
1086 are always aligned on strict-alignment targets. */
1087 if (TREE_CODE (exp) == MEM_REF || TREE_CODE (exp) == TARGET_MEM_REF)
1088 exp_align = get_object_or_type_alignment (exp);
1090 exp_align = get_object_alignment (exp);
1092 if (exp_align < align)
1098 /* Return true if EXP is a memory reference less aligned than what the access
1099 ACC would require. This is invoked only on strict-alignment targets. */
1102 tree_non_aligned_mem_for_access_p (tree exp, struct access *acc)
1104 unsigned int acc_align;
1106 /* The alignment of the access is that of its expression. However, it may
1107 have been artificially increased, e.g. by a local alignment promotion,
1108 so we cap it to the alignment of the type of the base, on the grounds
1109 that valid sub-accesses cannot be more aligned than that. */
1110 acc_align = get_object_alignment (acc->expr);
1111 if (acc->base && acc_align > TYPE_ALIGN (TREE_TYPE (acc->base)))
1112 acc_align = TYPE_ALIGN (TREE_TYPE (acc->base));
1114 return tree_non_aligned_mem_p (exp, acc_align);
1117 /* Scan expressions occuring in STMT, create access structures for all accesses
1118 to candidates for scalarization and remove those candidates which occur in
1119 statements or expressions that prevent them from being split apart. Return
1120 true if any access has been inserted. */
1123 build_accesses_from_assign (gimple stmt)
1126 struct access *lacc, *racc;
1128 if (!gimple_assign_single_p (stmt)
1129 /* Scope clobbers don't influence scalarization. */
1130 || gimple_clobber_p (stmt))
1133 lhs = gimple_assign_lhs (stmt);
1134 rhs = gimple_assign_rhs1 (stmt);
1136 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1139 racc = build_access_from_expr_1 (rhs, stmt, false);
1140 lacc = build_access_from_expr_1 (lhs, stmt, true);
1144 lacc->grp_assignment_write = 1;
1145 if (STRICT_ALIGNMENT && tree_non_aligned_mem_for_access_p (rhs, lacc))
1146 lacc->grp_unscalarizable_region = 1;
1151 racc->grp_assignment_read = 1;
1152 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1153 && !is_gimple_reg_type (racc->type))
1154 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1155 if (STRICT_ALIGNMENT && tree_non_aligned_mem_for_access_p (lhs, racc))
1156 racc->grp_unscalarizable_region = 1;
1160 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1161 && !lacc->grp_unscalarizable_region
1162 && !racc->grp_unscalarizable_region
1163 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1164 /* FIXME: Turn the following line into an assert after PR 40058 is
1166 && lacc->size == racc->size
1167 && useless_type_conversion_p (lacc->type, racc->type))
1169 struct assign_link *link;
1171 link = (struct assign_link *) pool_alloc (link_pool);
1172 memset (link, 0, sizeof (struct assign_link));
1177 add_link_to_rhs (racc, link);
1180 return lacc || racc;
1183 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1184 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1187 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1188 void *data ATTRIBUTE_UNUSED)
1190 op = get_base_address (op);
1193 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1198 /* Return true iff callsite CALL has at least as many actual arguments as there
1199 are formal parameters of the function currently processed by IPA-SRA. */
1202 callsite_has_enough_arguments_p (gimple call)
1204 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1207 /* Scan function and look for interesting expressions and create access
1208 structures for them. Return true iff any access is created. */
1211 scan_function (void)
1218 gimple_stmt_iterator gsi;
1219 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1221 gimple stmt = gsi_stmt (gsi);
1225 if (final_bbs && stmt_can_throw_external (stmt))
1226 bitmap_set_bit (final_bbs, bb->index);
1227 switch (gimple_code (stmt))
1230 t = gimple_return_retval (stmt);
1232 ret |= build_access_from_expr (t, stmt, false);
1234 bitmap_set_bit (final_bbs, bb->index);
1238 ret |= build_accesses_from_assign (stmt);
1242 for (i = 0; i < gimple_call_num_args (stmt); i++)
1243 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1246 if (sra_mode == SRA_MODE_EARLY_IPA)
1248 tree dest = gimple_call_fndecl (stmt);
1249 int flags = gimple_call_flags (stmt);
1253 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1254 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1255 encountered_apply_args = true;
1256 if (cgraph_get_node (dest)
1257 == cgraph_get_node (current_function_decl))
1259 encountered_recursive_call = true;
1260 if (!callsite_has_enough_arguments_p (stmt))
1261 encountered_unchangable_recursive_call = true;
1266 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1267 bitmap_set_bit (final_bbs, bb->index);
1270 t = gimple_call_lhs (stmt);
1271 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1272 ret |= build_access_from_expr (t, stmt, true);
1276 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1279 bitmap_set_bit (final_bbs, bb->index);
1281 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1283 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1284 ret |= build_access_from_expr (t, stmt, false);
1286 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1288 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1289 ret |= build_access_from_expr (t, stmt, true);
1302 /* Helper of QSORT function. There are pointers to accesses in the array. An
1303 access is considered smaller than another if it has smaller offset or if the
1304 offsets are the same but is size is bigger. */
1307 compare_access_positions (const void *a, const void *b)
1309 const access_p *fp1 = (const access_p *) a;
1310 const access_p *fp2 = (const access_p *) b;
1311 const access_p f1 = *fp1;
1312 const access_p f2 = *fp2;
1314 if (f1->offset != f2->offset)
1315 return f1->offset < f2->offset ? -1 : 1;
1317 if (f1->size == f2->size)
1319 if (f1->type == f2->type)
1321 /* Put any non-aggregate type before any aggregate type. */
1322 else if (!is_gimple_reg_type (f1->type)
1323 && is_gimple_reg_type (f2->type))
1325 else if (is_gimple_reg_type (f1->type)
1326 && !is_gimple_reg_type (f2->type))
1328 /* Put any complex or vector type before any other scalar type. */
1329 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1330 && TREE_CODE (f1->type) != VECTOR_TYPE
1331 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1332 || TREE_CODE (f2->type) == VECTOR_TYPE))
1334 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1335 || TREE_CODE (f1->type) == VECTOR_TYPE)
1336 && TREE_CODE (f2->type) != COMPLEX_TYPE
1337 && TREE_CODE (f2->type) != VECTOR_TYPE)
1339 /* Put the integral type with the bigger precision first. */
1340 else if (INTEGRAL_TYPE_P (f1->type)
1341 && INTEGRAL_TYPE_P (f2->type))
1342 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1343 /* Put any integral type with non-full precision last. */
1344 else if (INTEGRAL_TYPE_P (f1->type)
1345 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1346 != TYPE_PRECISION (f1->type)))
1348 else if (INTEGRAL_TYPE_P (f2->type)
1349 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1350 != TYPE_PRECISION (f2->type)))
1352 /* Stabilize the sort. */
1353 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1356 /* We want the bigger accesses first, thus the opposite operator in the next
1358 return f1->size > f2->size ? -1 : 1;
1362 /* Append a name of the declaration to the name obstack. A helper function for
1366 make_fancy_decl_name (tree decl)
1370 tree name = DECL_NAME (decl);
1372 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1373 IDENTIFIER_LENGTH (name));
1376 sprintf (buffer, "D%u", DECL_UID (decl));
1377 obstack_grow (&name_obstack, buffer, strlen (buffer));
1381 /* Helper for make_fancy_name. */
1384 make_fancy_name_1 (tree expr)
1391 make_fancy_decl_name (expr);
1395 switch (TREE_CODE (expr))
1398 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1399 obstack_1grow (&name_obstack, '$');
1400 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1404 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1405 obstack_1grow (&name_obstack, '$');
1406 /* Arrays with only one element may not have a constant as their
1408 index = TREE_OPERAND (expr, 1);
1409 if (TREE_CODE (index) != INTEGER_CST)
1411 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1412 obstack_grow (&name_obstack, buffer, strlen (buffer));
1416 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1420 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1421 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1423 obstack_1grow (&name_obstack, '$');
1424 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1425 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1426 obstack_grow (&name_obstack, buffer, strlen (buffer));
1433 gcc_unreachable (); /* we treat these as scalars. */
1440 /* Create a human readable name for replacement variable of ACCESS. */
1443 make_fancy_name (tree expr)
1445 make_fancy_name_1 (expr);
1446 obstack_1grow (&name_obstack, '\0');
1447 return XOBFINISH (&name_obstack, char *);
1450 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1451 EXP_TYPE at the given OFFSET. If BASE is something for which
1452 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1453 to insert new statements either before or below the current one as specified
1454 by INSERT_AFTER. This function is not capable of handling bitfields. */
1457 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1458 tree exp_type, gimple_stmt_iterator *gsi,
1461 tree prev_base = base;
1463 HOST_WIDE_INT base_offset;
1464 unsigned HOST_WIDE_INT misalign;
1467 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1469 base = get_addr_base_and_unit_offset (base, &base_offset);
1471 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1472 offset such as array[var_index]. */
1478 gcc_checking_assert (gsi);
1479 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1480 add_referenced_var (tmp);
1481 tmp = make_ssa_name (tmp, NULL);
1482 addr = build_fold_addr_expr (unshare_expr (prev_base));
1483 STRIP_USELESS_TYPE_CONVERSION (addr);
1484 stmt = gimple_build_assign (tmp, addr);
1485 gimple_set_location (stmt, loc);
1486 SSA_NAME_DEF_STMT (tmp) = stmt;
1488 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1490 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1493 off = build_int_cst (reference_alias_ptr_type (prev_base),
1494 offset / BITS_PER_UNIT);
1497 else if (TREE_CODE (base) == MEM_REF)
1499 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1500 base_offset + offset / BITS_PER_UNIT);
1501 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1502 base = unshare_expr (TREE_OPERAND (base, 0));
1506 off = build_int_cst (reference_alias_ptr_type (base),
1507 base_offset + offset / BITS_PER_UNIT);
1508 base = build_fold_addr_expr (unshare_expr (base));
1511 /* If prev_base were always an originally performed access
1512 we can extract more optimistic alignment information
1513 by looking at the access mode. That would constrain the
1514 alignment of base + base_offset which we would need to
1515 adjust according to offset.
1516 ??? But it is not at all clear that prev_base is an access
1517 that was in the IL that way, so be conservative for now. */
1518 align = get_pointer_alignment_1 (base, &misalign);
1519 misalign += (double_int_sext (tree_to_double_int (off),
1520 TYPE_PRECISION (TREE_TYPE (off))).low
1522 misalign = misalign & (align - 1);
1524 align = (misalign & -misalign);
1525 if (align < TYPE_ALIGN (exp_type))
1526 exp_type = build_aligned_type (exp_type, align);
1528 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1531 DEF_VEC_ALLOC_P_STACK (tree);
1532 #define VEC_tree_stack_alloc(alloc) VEC_stack_alloc (tree, alloc)
1534 /* Construct a memory reference to a part of an aggregate BASE at the given
1535 OFFSET and of the type of MODEL. In case this is a chain of references
1536 to component, the function will replicate the chain of COMPONENT_REFs of
1537 the expression of MODEL to access it. GSI and INSERT_AFTER have the same
1538 meaning as in build_ref_for_offset. */
1541 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1542 struct access *model, gimple_stmt_iterator *gsi,
1545 tree type = model->type, t;
1546 VEC(tree,stack) *cr_stack = NULL;
1548 if (TREE_CODE (model->expr) == COMPONENT_REF)
1550 tree expr = model->expr;
1552 /* Create a stack of the COMPONENT_REFs so later we can walk them in
1553 order from inner to outer. */
1554 cr_stack = VEC_alloc (tree, stack, 6);
1557 tree field = TREE_OPERAND (expr, 1);
1558 tree cr_offset = component_ref_field_offset (expr);
1559 HOST_WIDE_INT bit_pos
1560 = tree_low_cst (cr_offset, 1) * BITS_PER_UNIT
1561 + TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (field));
1563 /* We can be called with a model different from the one associated
1564 with BASE so we need to avoid going up the chain too far. */
1565 if (offset - bit_pos < 0)
1569 VEC_safe_push (tree, stack, cr_stack, expr);
1571 expr = TREE_OPERAND (expr, 0);
1572 type = TREE_TYPE (expr);
1573 } while (TREE_CODE (expr) == COMPONENT_REF);
1576 t = build_ref_for_offset (loc, base, offset, type, gsi, insert_after);
1578 if (TREE_CODE (model->expr) == COMPONENT_REF)
1583 /* Now replicate the chain of COMPONENT_REFs from inner to outer. */
1584 FOR_EACH_VEC_ELT_REVERSE (tree, cr_stack, i, expr)
1586 tree field = TREE_OPERAND (expr, 1);
1587 t = fold_build3_loc (loc, COMPONENT_REF, TREE_TYPE (field), t, field,
1588 TREE_OPERAND (expr, 2));
1591 VEC_free (tree, stack, cr_stack);
1597 /* Construct a memory reference consisting of component_refs and array_refs to
1598 a part of an aggregate *RES (which is of type TYPE). The requested part
1599 should have type EXP_TYPE at be the given OFFSET. This function might not
1600 succeed, it returns true when it does and only then *RES points to something
1601 meaningful. This function should be used only to build expressions that we
1602 might need to present to user (e.g. in warnings). In all other situations,
1603 build_ref_for_model or build_ref_for_offset should be used instead. */
1606 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1612 tree tr_size, index, minidx;
1613 HOST_WIDE_INT el_size;
1615 if (offset == 0 && exp_type
1616 && types_compatible_p (exp_type, type))
1619 switch (TREE_CODE (type))
1622 case QUAL_UNION_TYPE:
1624 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1626 HOST_WIDE_INT pos, size;
1627 tree expr, *expr_ptr;
1629 if (TREE_CODE (fld) != FIELD_DECL)
1632 pos = int_bit_position (fld);
1633 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1634 tr_size = DECL_SIZE (fld);
1635 if (!tr_size || !host_integerp (tr_size, 1))
1637 size = tree_low_cst (tr_size, 1);
1643 else if (pos > offset || (pos + size) <= offset)
1646 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1649 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1650 offset - pos, exp_type))
1659 tr_size = TYPE_SIZE (TREE_TYPE (type));
1660 if (!tr_size || !host_integerp (tr_size, 1))
1662 el_size = tree_low_cst (tr_size, 1);
1664 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1665 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1667 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1668 if (!integer_zerop (minidx))
1669 index = int_const_binop (PLUS_EXPR, index, minidx);
1670 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1671 NULL_TREE, NULL_TREE);
1672 offset = offset % el_size;
1673 type = TREE_TYPE (type);
1688 /* Return true iff TYPE is stdarg va_list type. */
1691 is_va_list_type (tree type)
1693 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1696 /* Print message to dump file why a variable was rejected. */
1699 reject (tree var, const char *msg)
1701 if (dump_file && (dump_flags & TDF_DETAILS))
1703 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1704 print_generic_expr (dump_file, var, 0);
1705 fprintf (dump_file, "\n");
1709 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1710 those with type which is suitable for scalarization. */
1713 find_var_candidates (void)
1716 referenced_var_iterator rvi;
1720 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
1722 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1724 type = TREE_TYPE (var);
1726 if (!AGGREGATE_TYPE_P (type))
1728 reject (var, "not aggregate");
1731 if (needs_to_live_in_memory (var))
1733 reject (var, "needs to live in memory");
1736 if (TREE_THIS_VOLATILE (var))
1738 reject (var, "is volatile");
1741 if (!COMPLETE_TYPE_P (type))
1743 reject (var, "has incomplete type");
1746 if (!host_integerp (TYPE_SIZE (type), 1))
1748 reject (var, "type size not fixed");
1751 if (tree_low_cst (TYPE_SIZE (type), 1) == 0)
1753 reject (var, "type size is zero");
1756 if (type_internals_preclude_sra_p (type, &msg))
1761 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1762 we also want to schedule it rather late. Thus we ignore it in
1764 (sra_mode == SRA_MODE_EARLY_INTRA
1765 && is_va_list_type (type)))
1767 reject (var, "is va_list");
1771 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1773 if (dump_file && (dump_flags & TDF_DETAILS))
1775 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1776 print_generic_expr (dump_file, var, 0);
1777 fprintf (dump_file, "\n");
1785 /* Sort all accesses for the given variable, check for partial overlaps and
1786 return NULL if there are any. If there are none, pick a representative for
1787 each combination of offset and size and create a linked list out of them.
1788 Return the pointer to the first representative and make sure it is the first
1789 one in the vector of accesses. */
1791 static struct access *
1792 sort_and_splice_var_accesses (tree var)
1794 int i, j, access_count;
1795 struct access *res, **prev_acc_ptr = &res;
1796 VEC (access_p, heap) *access_vec;
1798 HOST_WIDE_INT low = -1, high = 0;
1800 access_vec = get_base_access_vector (var);
1803 access_count = VEC_length (access_p, access_vec);
1805 /* Sort by <OFFSET, SIZE>. */
1806 VEC_qsort (access_p, access_vec, compare_access_positions);
1809 while (i < access_count)
1811 struct access *access = VEC_index (access_p, access_vec, i);
1812 bool grp_write = access->write;
1813 bool grp_read = !access->write;
1814 bool grp_scalar_write = access->write
1815 && is_gimple_reg_type (access->type);
1816 bool grp_scalar_read = !access->write
1817 && is_gimple_reg_type (access->type);
1818 bool grp_assignment_read = access->grp_assignment_read;
1819 bool grp_assignment_write = access->grp_assignment_write;
1820 bool multiple_scalar_reads = false;
1821 bool total_scalarization = access->grp_total_scalarization;
1822 bool grp_partial_lhs = access->grp_partial_lhs;
1823 bool first_scalar = is_gimple_reg_type (access->type);
1824 bool unscalarizable_region = access->grp_unscalarizable_region;
1826 if (first || access->offset >= high)
1829 low = access->offset;
1830 high = access->offset + access->size;
1832 else if (access->offset > low && access->offset + access->size > high)
1835 gcc_assert (access->offset >= low
1836 && access->offset + access->size <= high);
1839 while (j < access_count)
1841 struct access *ac2 = VEC_index (access_p, access_vec, j);
1842 if (ac2->offset != access->offset || ac2->size != access->size)
1847 grp_scalar_write = (grp_scalar_write
1848 || is_gimple_reg_type (ac2->type));
1853 if (is_gimple_reg_type (ac2->type))
1855 if (grp_scalar_read)
1856 multiple_scalar_reads = true;
1858 grp_scalar_read = true;
1861 grp_assignment_read |= ac2->grp_assignment_read;
1862 grp_assignment_write |= ac2->grp_assignment_write;
1863 grp_partial_lhs |= ac2->grp_partial_lhs;
1864 unscalarizable_region |= ac2->grp_unscalarizable_region;
1865 total_scalarization |= ac2->grp_total_scalarization;
1866 relink_to_new_repr (access, ac2);
1868 /* If there are both aggregate-type and scalar-type accesses with
1869 this combination of size and offset, the comparison function
1870 should have put the scalars first. */
1871 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1872 ac2->group_representative = access;
1878 access->group_representative = access;
1879 access->grp_write = grp_write;
1880 access->grp_read = grp_read;
1881 access->grp_scalar_read = grp_scalar_read;
1882 access->grp_scalar_write = grp_scalar_write;
1883 access->grp_assignment_read = grp_assignment_read;
1884 access->grp_assignment_write = grp_assignment_write;
1885 access->grp_hint = multiple_scalar_reads || total_scalarization;
1886 access->grp_total_scalarization = total_scalarization;
1887 access->grp_partial_lhs = grp_partial_lhs;
1888 access->grp_unscalarizable_region = unscalarizable_region;
1889 if (access->first_link)
1890 add_access_to_work_queue (access);
1892 *prev_acc_ptr = access;
1893 prev_acc_ptr = &access->next_grp;
1896 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1900 /* Create a variable for the given ACCESS which determines the type, name and a
1901 few other properties. Return the variable declaration and store it also to
1902 ACCESS->replacement. */
1905 create_access_replacement (struct access *access, bool rename)
1909 repl = create_tmp_var (access->type, "SR");
1910 add_referenced_var (repl);
1912 mark_sym_for_renaming (repl);
1914 if (!access->grp_partial_lhs
1915 && (TREE_CODE (access->type) == COMPLEX_TYPE
1916 || TREE_CODE (access->type) == VECTOR_TYPE))
1917 DECL_GIMPLE_REG_P (repl) = 1;
1919 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1920 DECL_ARTIFICIAL (repl) = 1;
1921 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1923 if (DECL_NAME (access->base)
1924 && !DECL_IGNORED_P (access->base)
1925 && !DECL_ARTIFICIAL (access->base))
1927 char *pretty_name = make_fancy_name (access->expr);
1928 tree debug_expr = unshare_expr (access->expr), d;
1930 DECL_NAME (repl) = get_identifier (pretty_name);
1931 obstack_free (&name_obstack, pretty_name);
1933 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1934 as DECL_DEBUG_EXPR isn't considered when looking for still
1935 used SSA_NAMEs and thus they could be freed. All debug info
1936 generation cares is whether something is constant or variable
1937 and that get_ref_base_and_extent works properly on the
1939 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1940 switch (TREE_CODE (d))
1943 case ARRAY_RANGE_REF:
1944 if (TREE_OPERAND (d, 1)
1945 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1946 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1947 if (TREE_OPERAND (d, 3)
1948 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1949 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1952 if (TREE_OPERAND (d, 2)
1953 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1954 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1959 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1960 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1961 if (access->grp_no_warning)
1962 TREE_NO_WARNING (repl) = 1;
1964 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1967 TREE_NO_WARNING (repl) = 1;
1971 fprintf (dump_file, "Created a replacement for ");
1972 print_generic_expr (dump_file, access->base, 0);
1973 fprintf (dump_file, " offset: %u, size: %u: ",
1974 (unsigned) access->offset, (unsigned) access->size);
1975 print_generic_expr (dump_file, repl, 0);
1976 fprintf (dump_file, "\n");
1978 sra_stats.replacements++;
1983 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1986 get_access_replacement (struct access *access)
1988 gcc_assert (access->grp_to_be_replaced);
1990 if (!access->replacement_decl)
1991 access->replacement_decl = create_access_replacement (access, true);
1992 return access->replacement_decl;
1995 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1996 not mark it for renaming. */
1999 get_unrenamed_access_replacement (struct access *access)
2001 gcc_assert (!access->grp_to_be_replaced);
2003 if (!access->replacement_decl)
2004 access->replacement_decl = create_access_replacement (access, false);
2005 return access->replacement_decl;
2009 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
2010 linked list along the way. Stop when *ACCESS is NULL or the access pointed
2011 to it is not "within" the root. Return false iff some accesses partially
2015 build_access_subtree (struct access **access)
2017 struct access *root = *access, *last_child = NULL;
2018 HOST_WIDE_INT limit = root->offset + root->size;
2020 *access = (*access)->next_grp;
2021 while (*access && (*access)->offset + (*access)->size <= limit)
2024 root->first_child = *access;
2026 last_child->next_sibling = *access;
2027 last_child = *access;
2029 if (!build_access_subtree (access))
2033 if (*access && (*access)->offset < limit)
2039 /* Build a tree of access representatives, ACCESS is the pointer to the first
2040 one, others are linked in a list by the next_grp field. Return false iff
2041 some accesses partially overlap. */
2044 build_access_trees (struct access *access)
2048 struct access *root = access;
2050 if (!build_access_subtree (&access))
2052 root->next_grp = access;
2057 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2061 expr_with_var_bounded_array_refs_p (tree expr)
2063 while (handled_component_p (expr))
2065 if (TREE_CODE (expr) == ARRAY_REF
2066 && !host_integerp (array_ref_low_bound (expr), 0))
2068 expr = TREE_OPERAND (expr, 0);
2073 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2074 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2075 sorts of access flags appropriately along the way, notably always set
2076 grp_read and grp_assign_read according to MARK_READ and grp_write when
2079 Creating a replacement for a scalar access is considered beneficial if its
2080 grp_hint is set (this means we are either attempting total scalarization or
2081 there is more than one direct read access) or according to the following
2084 Access written to through a scalar type (once or more times)
2086 | Written to in an assignment statement
2088 | | Access read as scalar _once_
2090 | | | Read in an assignment statement
2092 | | | | Scalarize Comment
2093 -----------------------------------------------------------------------------
2094 0 0 0 0 No access for the scalar
2095 0 0 0 1 No access for the scalar
2096 0 0 1 0 No Single read - won't help
2097 0 0 1 1 No The same case
2098 0 1 0 0 No access for the scalar
2099 0 1 0 1 No access for the scalar
2100 0 1 1 0 Yes s = *g; return s.i;
2101 0 1 1 1 Yes The same case as above
2102 1 0 0 0 No Won't help
2103 1 0 0 1 Yes s.i = 1; *g = s;
2104 1 0 1 0 Yes s.i = 5; g = s.i;
2105 1 0 1 1 Yes The same case as above
2106 1 1 0 0 No Won't help.
2107 1 1 0 1 Yes s.i = 1; *g = s;
2108 1 1 1 0 Yes s = *g; return s.i;
2109 1 1 1 1 Yes Any of the above yeses */
2112 analyze_access_subtree (struct access *root, struct access *parent,
2113 bool allow_replacements)
2115 struct access *child;
2116 HOST_WIDE_INT limit = root->offset + root->size;
2117 HOST_WIDE_INT covered_to = root->offset;
2118 bool scalar = is_gimple_reg_type (root->type);
2119 bool hole = false, sth_created = false;
2123 if (parent->grp_read)
2125 if (parent->grp_assignment_read)
2126 root->grp_assignment_read = 1;
2127 if (parent->grp_write)
2128 root->grp_write = 1;
2129 if (parent->grp_assignment_write)
2130 root->grp_assignment_write = 1;
2131 if (parent->grp_total_scalarization)
2132 root->grp_total_scalarization = 1;
2135 if (root->grp_unscalarizable_region)
2136 allow_replacements = false;
2138 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2139 allow_replacements = false;
2141 for (child = root->first_child; child; child = child->next_sibling)
2143 hole |= covered_to < child->offset;
2144 sth_created |= analyze_access_subtree (child, root,
2145 allow_replacements && !scalar);
2147 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2148 root->grp_total_scalarization &= child->grp_total_scalarization;
2149 if (child->grp_covered)
2150 covered_to += child->size;
2155 if (allow_replacements && scalar && !root->first_child
2157 || ((root->grp_scalar_read || root->grp_assignment_read)
2158 && (root->grp_scalar_write || root->grp_assignment_write))))
2160 bool new_integer_type;
2161 if (TREE_CODE (root->type) == ENUMERAL_TYPE)
2163 tree rt = root->type;
2164 root->type = build_nonstandard_integer_type (TYPE_PRECISION (rt),
2165 TYPE_UNSIGNED (rt));
2166 new_integer_type = true;
2169 new_integer_type = false;
2171 if (dump_file && (dump_flags & TDF_DETAILS))
2173 fprintf (dump_file, "Marking ");
2174 print_generic_expr (dump_file, root->base, 0);
2175 fprintf (dump_file, " offset: %u, size: %u ",
2176 (unsigned) root->offset, (unsigned) root->size);
2177 fprintf (dump_file, " to be replaced%s.\n",
2178 new_integer_type ? " with an integer": "");
2181 root->grp_to_be_replaced = 1;
2187 if (covered_to < limit)
2190 root->grp_total_scalarization = 0;
2194 && (!hole || root->grp_total_scalarization))
2196 root->grp_covered = 1;
2199 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2200 root->grp_unscalarized_data = 1; /* not covered and written to */
2206 /* Analyze all access trees linked by next_grp by the means of
2207 analyze_access_subtree. */
2209 analyze_access_trees (struct access *access)
2215 if (analyze_access_subtree (access, NULL, true))
2217 access = access->next_grp;
2223 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2224 SIZE would conflict with an already existing one. If exactly such a child
2225 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2228 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2229 HOST_WIDE_INT size, struct access **exact_match)
2231 struct access *child;
2233 for (child = lacc->first_child; child; child = child->next_sibling)
2235 if (child->offset == norm_offset && child->size == size)
2237 *exact_match = child;
2241 if (child->offset < norm_offset + size
2242 && child->offset + child->size > norm_offset)
2249 /* Create a new child access of PARENT, with all properties just like MODEL
2250 except for its offset and with its grp_write false and grp_read true.
2251 Return the new access or NULL if it cannot be created. Note that this access
2252 is created long after all splicing and sorting, it's not located in any
2253 access vector and is automatically a representative of its group. */
2255 static struct access *
2256 create_artificial_child_access (struct access *parent, struct access *model,
2257 HOST_WIDE_INT new_offset)
2259 struct access *access;
2260 struct access **child;
2261 tree expr = parent->base;
2263 gcc_assert (!model->grp_unscalarizable_region);
2265 access = (struct access *) pool_alloc (access_pool);
2266 memset (access, 0, sizeof (struct access));
2267 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2270 access->grp_no_warning = true;
2271 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2272 new_offset, model, NULL, false);
2275 access->base = parent->base;
2276 access->expr = expr;
2277 access->offset = new_offset;
2278 access->size = model->size;
2279 access->type = model->type;
2280 access->grp_write = true;
2281 access->grp_read = false;
2283 child = &parent->first_child;
2284 while (*child && (*child)->offset < new_offset)
2285 child = &(*child)->next_sibling;
2287 access->next_sibling = *child;
2294 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2295 true if any new subaccess was created. Additionally, if RACC is a scalar
2296 access but LACC is not, change the type of the latter, if possible. */
2299 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2301 struct access *rchild;
2302 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2305 if (is_gimple_reg_type (lacc->type)
2306 || lacc->grp_unscalarizable_region
2307 || racc->grp_unscalarizable_region)
2310 if (is_gimple_reg_type (racc->type))
2312 if (!lacc->first_child && !racc->first_child)
2314 tree t = lacc->base;
2316 lacc->type = racc->type;
2317 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t),
2318 lacc->offset, racc->type))
2322 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2323 lacc->base, lacc->offset,
2325 lacc->grp_no_warning = true;
2331 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2333 struct access *new_acc = NULL;
2334 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2336 if (rchild->grp_unscalarizable_region)
2339 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2344 rchild->grp_hint = 1;
2345 new_acc->grp_hint |= new_acc->grp_read;
2346 if (rchild->first_child)
2347 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2352 rchild->grp_hint = 1;
2353 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2357 if (racc->first_child)
2358 propagate_subaccesses_across_link (new_acc, rchild);
2365 /* Propagate all subaccesses across assignment links. */
2368 propagate_all_subaccesses (void)
2370 while (work_queue_head)
2372 struct access *racc = pop_access_from_work_queue ();
2373 struct assign_link *link;
2375 gcc_assert (racc->first_link);
2377 for (link = racc->first_link; link; link = link->next)
2379 struct access *lacc = link->lacc;
2381 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2383 lacc = lacc->group_representative;
2384 if (propagate_subaccesses_across_link (lacc, racc)
2385 && lacc->first_link)
2386 add_access_to_work_queue (lacc);
2391 /* Go through all accesses collected throughout the (intraprocedural) analysis
2392 stage, exclude overlapping ones, identify representatives and build trees
2393 out of them, making decisions about scalarization on the way. Return true
2394 iff there are any to-be-scalarized variables after this stage. */
2397 analyze_all_variable_accesses (void)
2400 bitmap tmp = BITMAP_ALLOC (NULL);
2402 unsigned i, max_total_scalarization_size;
2404 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2405 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2407 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2408 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2409 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2411 tree var = referenced_var (i);
2413 if (TREE_CODE (var) == VAR_DECL
2414 && type_consists_of_records_p (TREE_TYPE (var)))
2416 if ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2417 <= max_total_scalarization_size)
2419 completely_scalarize_var (var);
2420 if (dump_file && (dump_flags & TDF_DETAILS))
2422 fprintf (dump_file, "Will attempt to totally scalarize ");
2423 print_generic_expr (dump_file, var, 0);
2424 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2427 else if (dump_file && (dump_flags & TDF_DETAILS))
2429 fprintf (dump_file, "Too big to totally scalarize: ");
2430 print_generic_expr (dump_file, var, 0);
2431 fprintf (dump_file, " (UID: %u)\n", DECL_UID (var));
2436 bitmap_copy (tmp, candidate_bitmap);
2437 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2439 tree var = referenced_var (i);
2440 struct access *access;
2442 access = sort_and_splice_var_accesses (var);
2443 if (!access || !build_access_trees (access))
2444 disqualify_candidate (var,
2445 "No or inhibitingly overlapping accesses.");
2448 propagate_all_subaccesses ();
2450 bitmap_copy (tmp, candidate_bitmap);
2451 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2453 tree var = referenced_var (i);
2454 struct access *access = get_first_repr_for_decl (var);
2456 if (analyze_access_trees (access))
2459 if (dump_file && (dump_flags & TDF_DETAILS))
2461 fprintf (dump_file, "\nAccess trees for ");
2462 print_generic_expr (dump_file, var, 0);
2463 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2464 dump_access_tree (dump_file, access);
2465 fprintf (dump_file, "\n");
2469 disqualify_candidate (var, "No scalar replacements to be created.");
2476 statistics_counter_event (cfun, "Scalarized aggregates", res);
2483 /* Generate statements copying scalar replacements of accesses within a subtree
2484 into or out of AGG. ACCESS, all its children, siblings and their children
2485 are to be processed. AGG is an aggregate type expression (can be a
2486 declaration but does not have to be, it can for example also be a mem_ref or
2487 a series of handled components). TOP_OFFSET is the offset of the processed
2488 subtree which has to be subtracted from offsets of individual accesses to
2489 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2490 replacements in the interval <start_offset, start_offset + chunk_size>,
2491 otherwise copy all. GSI is a statement iterator used to place the new
2492 statements. WRITE should be true when the statements should write from AGG
2493 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2494 statements will be added after the current statement in GSI, they will be
2495 added before the statement otherwise. */
2498 generate_subtree_copies (struct access *access, tree agg,
2499 HOST_WIDE_INT top_offset,
2500 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2501 gimple_stmt_iterator *gsi, bool write,
2502 bool insert_after, location_t loc)
2506 if (chunk_size && access->offset >= start_offset + chunk_size)
2509 if (access->grp_to_be_replaced
2511 || access->offset + access->size > start_offset))
2513 tree expr, repl = get_access_replacement (access);
2516 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2517 access, gsi, insert_after);
2521 if (access->grp_partial_lhs)
2522 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2524 insert_after ? GSI_NEW_STMT
2526 stmt = gimple_build_assign (repl, expr);
2530 TREE_NO_WARNING (repl) = 1;
2531 if (access->grp_partial_lhs)
2532 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2534 insert_after ? GSI_NEW_STMT
2536 stmt = gimple_build_assign (expr, repl);
2538 gimple_set_location (stmt, loc);
2541 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2543 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2545 sra_stats.subtree_copies++;
2548 if (access->first_child)
2549 generate_subtree_copies (access->first_child, agg, top_offset,
2550 start_offset, chunk_size, gsi,
2551 write, insert_after, loc);
2553 access = access->next_sibling;
2558 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2559 the root of the subtree to be processed. GSI is the statement iterator used
2560 for inserting statements which are added after the current statement if
2561 INSERT_AFTER is true or before it otherwise. */
2564 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2565 bool insert_after, location_t loc)
2568 struct access *child;
2570 if (access->grp_to_be_replaced)
2574 stmt = gimple_build_assign (get_access_replacement (access),
2575 build_zero_cst (access->type));
2577 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2579 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2581 gimple_set_location (stmt, loc);
2584 for (child = access->first_child; child; child = child->next_sibling)
2585 init_subtree_with_zero (child, gsi, insert_after, loc);
2588 /* Search for an access representative for the given expression EXPR and
2589 return it or NULL if it cannot be found. */
2591 static struct access *
2592 get_access_for_expr (tree expr)
2594 HOST_WIDE_INT offset, size, max_size;
2597 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2598 a different size than the size of its argument and we need the latter
2600 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2601 expr = TREE_OPERAND (expr, 0);
2603 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2604 if (max_size == -1 || !DECL_P (base))
2607 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2610 return get_var_base_offset_size_access (base, offset, max_size);
2613 /* Replace the expression EXPR with a scalar replacement if there is one and
2614 generate other statements to do type conversion or subtree copying if
2615 necessary. GSI is used to place newly created statements, WRITE is true if
2616 the expression is being written to (it is on a LHS of a statement or output
2617 in an assembly statement). */
2620 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2623 struct access *access;
2626 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2629 expr = &TREE_OPERAND (*expr, 0);
2634 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2635 expr = &TREE_OPERAND (*expr, 0);
2636 access = get_access_for_expr (*expr);
2639 type = TREE_TYPE (*expr);
2641 loc = gimple_location (gsi_stmt (*gsi));
2642 if (access->grp_to_be_replaced)
2644 tree repl = get_access_replacement (access);
2645 /* If we replace a non-register typed access simply use the original
2646 access expression to extract the scalar component afterwards.
2647 This happens if scalarizing a function return value or parameter
2648 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2649 gcc.c-torture/compile/20011217-1.c.
2651 We also want to use this when accessing a complex or vector which can
2652 be accessed as a different type too, potentially creating a need for
2653 type conversion (see PR42196) and when scalarized unions are involved
2654 in assembler statements (see PR42398). */
2655 if (!useless_type_conversion_p (type, access->type))
2659 ref = build_ref_for_model (loc, access->base, access->offset, access,
2666 if (access->grp_partial_lhs)
2667 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2668 false, GSI_NEW_STMT);
2669 stmt = gimple_build_assign (repl, ref);
2670 gimple_set_location (stmt, loc);
2671 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2677 if (access->grp_partial_lhs)
2678 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2679 true, GSI_SAME_STMT);
2680 stmt = gimple_build_assign (ref, repl);
2681 gimple_set_location (stmt, loc);
2682 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2690 if (access->first_child)
2692 HOST_WIDE_INT start_offset, chunk_size;
2694 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2695 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2697 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2698 start_offset = access->offset
2699 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2702 start_offset = chunk_size = 0;
2704 generate_subtree_copies (access->first_child, access->base, 0,
2705 start_offset, chunk_size, gsi, write, write,
2711 /* Where scalar replacements of the RHS have been written to when a replacement
2712 of a LHS of an assigments cannot be direclty loaded from a replacement of
2714 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2715 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2716 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2718 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2719 base aggregate if there are unscalarized data or directly to LHS of the
2720 statement that is pointed to by GSI otherwise. */
2722 static enum unscalarized_data_handling
2723 handle_unscalarized_data_in_subtree (struct access *top_racc,
2724 gimple_stmt_iterator *gsi)
2726 if (top_racc->grp_unscalarized_data)
2728 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2730 gimple_location (gsi_stmt (*gsi)));
2731 return SRA_UDH_RIGHT;
2735 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2736 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2737 0, 0, gsi, false, false,
2738 gimple_location (gsi_stmt (*gsi)));
2739 return SRA_UDH_LEFT;
2744 /* Try to generate statements to load all sub-replacements in an access subtree
2745 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2746 If that is not possible, refresh the TOP_RACC base aggregate and load the
2747 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2748 copied. NEW_GSI is stmt iterator used for statement insertions after the
2749 original assignment, OLD_GSI is used to insert statements before the
2750 assignment. *REFRESHED keeps the information whether we have needed to
2751 refresh replacements of the LHS and from which side of the assignments this
2755 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2756 HOST_WIDE_INT left_offset,
2757 gimple_stmt_iterator *old_gsi,
2758 gimple_stmt_iterator *new_gsi,
2759 enum unscalarized_data_handling *refreshed)
2761 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2762 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2764 if (lacc->grp_to_be_replaced)
2766 struct access *racc;
2767 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2771 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2772 if (racc && racc->grp_to_be_replaced)
2774 rhs = get_access_replacement (racc);
2775 if (!useless_type_conversion_p (lacc->type, racc->type))
2776 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2778 if (racc->grp_partial_lhs && lacc->grp_partial_lhs)
2779 rhs = force_gimple_operand_gsi (old_gsi, rhs, true, NULL_TREE,
2780 true, GSI_SAME_STMT);
2784 /* No suitable access on the right hand side, need to load from
2785 the aggregate. See if we have to update it first... */
2786 if (*refreshed == SRA_UDH_NONE)
2787 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2790 if (*refreshed == SRA_UDH_LEFT)
2791 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2794 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2796 if (lacc->grp_partial_lhs)
2797 rhs = force_gimple_operand_gsi (new_gsi, rhs, true, NULL_TREE,
2798 false, GSI_NEW_STMT);
2801 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2802 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2803 gimple_set_location (stmt, loc);
2805 sra_stats.subreplacements++;
2807 else if (*refreshed == SRA_UDH_NONE
2808 && lacc->grp_read && !lacc->grp_covered)
2809 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2812 if (lacc->first_child)
2813 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2814 old_gsi, new_gsi, refreshed);
2818 /* Result code for SRA assignment modification. */
2819 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2820 SRA_AM_MODIFIED, /* stmt changed but not
2822 SRA_AM_REMOVED }; /* stmt eliminated */
2824 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2825 to the assignment and GSI is the statement iterator pointing at it. Returns
2826 the same values as sra_modify_assign. */
2828 static enum assignment_mod_result
2829 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2831 tree lhs = gimple_assign_lhs (*stmt);
2835 acc = get_access_for_expr (lhs);
2839 if (gimple_clobber_p (*stmt))
2841 /* Remove clobbers of fully scalarized variables, otherwise
2843 if (acc->grp_covered)
2845 unlink_stmt_vdef (*stmt);
2846 gsi_remove (gsi, true);
2847 return SRA_AM_REMOVED;
2853 loc = gimple_location (*stmt);
2854 if (VEC_length (constructor_elt,
2855 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2857 /* I have never seen this code path trigger but if it can happen the
2858 following should handle it gracefully. */
2859 if (access_has_children_p (acc))
2860 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2862 return SRA_AM_MODIFIED;
2865 if (acc->grp_covered)
2867 init_subtree_with_zero (acc, gsi, false, loc);
2868 unlink_stmt_vdef (*stmt);
2869 gsi_remove (gsi, true);
2870 return SRA_AM_REMOVED;
2874 init_subtree_with_zero (acc, gsi, true, loc);
2875 return SRA_AM_MODIFIED;
2879 /* Create and return a new suitable default definition SSA_NAME for RACC which
2880 is an access describing an uninitialized part of an aggregate that is being
2884 get_repl_default_def_ssa_name (struct access *racc)
2888 decl = get_unrenamed_access_replacement (racc);
2890 repl = gimple_default_def (cfun, decl);
2893 repl = make_ssa_name (decl, gimple_build_nop ());
2894 set_default_def (decl, repl);
2900 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2904 contains_bitfld_comp_ref_p (const_tree ref)
2906 while (handled_component_p (ref))
2908 if (TREE_CODE (ref) == COMPONENT_REF
2909 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2911 ref = TREE_OPERAND (ref, 0);
2917 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2918 bit-field field declaration somewhere in it. */
2921 contains_vce_or_bfcref_p (const_tree ref)
2923 while (handled_component_p (ref))
2925 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2926 || (TREE_CODE (ref) == COMPONENT_REF
2927 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2929 ref = TREE_OPERAND (ref, 0);
2935 /* Examine both sides of the assignment statement pointed to by STMT, replace
2936 them with a scalare replacement if there is one and generate copying of
2937 replacements if scalarized aggregates have been used in the assignment. GSI
2938 is used to hold generated statements for type conversions and subtree
2941 static enum assignment_mod_result
2942 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2944 struct access *lacc, *racc;
2946 bool modify_this_stmt = false;
2947 bool force_gimple_rhs = false;
2949 gimple_stmt_iterator orig_gsi = *gsi;
2951 if (!gimple_assign_single_p (*stmt))
2953 lhs = gimple_assign_lhs (*stmt);
2954 rhs = gimple_assign_rhs1 (*stmt);
2956 if (TREE_CODE (rhs) == CONSTRUCTOR)
2957 return sra_modify_constructor_assign (stmt, gsi);
2959 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2960 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2961 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2963 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2965 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2967 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2970 lacc = get_access_for_expr (lhs);
2971 racc = get_access_for_expr (rhs);
2975 loc = gimple_location (*stmt);
2976 if (lacc && lacc->grp_to_be_replaced)
2978 lhs = get_access_replacement (lacc);
2979 gimple_assign_set_lhs (*stmt, lhs);
2980 modify_this_stmt = true;
2981 if (lacc->grp_partial_lhs)
2982 force_gimple_rhs = true;
2986 if (racc && racc->grp_to_be_replaced)
2988 rhs = get_access_replacement (racc);
2989 modify_this_stmt = true;
2990 if (racc->grp_partial_lhs)
2991 force_gimple_rhs = true;
2995 && !access_has_children_p (racc)
2996 && !racc->grp_to_be_replaced
2997 && !racc->grp_unscalarized_data
2998 && TREE_CODE (lhs) == SSA_NAME)
3000 rhs = get_repl_default_def_ssa_name (racc);
3001 modify_this_stmt = true;
3005 if (modify_this_stmt)
3007 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3009 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
3010 ??? This should move to fold_stmt which we simply should
3011 call after building a VIEW_CONVERT_EXPR here. */
3012 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
3013 && !contains_bitfld_comp_ref_p (lhs)
3014 && !access_has_children_p (lacc))
3016 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
3017 gimple_assign_set_lhs (*stmt, lhs);
3019 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
3020 && !contains_vce_or_bfcref_p (rhs)
3021 && !access_has_children_p (racc))
3022 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
3024 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3026 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
3028 if (is_gimple_reg_type (TREE_TYPE (lhs))
3029 && TREE_CODE (lhs) != SSA_NAME)
3030 force_gimple_rhs = true;
3035 /* From this point on, the function deals with assignments in between
3036 aggregates when at least one has scalar reductions of some of its
3037 components. There are three possible scenarios: Both the LHS and RHS have
3038 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
3040 In the first case, we would like to load the LHS components from RHS
3041 components whenever possible. If that is not possible, we would like to
3042 read it directly from the RHS (after updating it by storing in it its own
3043 components). If there are some necessary unscalarized data in the LHS,
3044 those will be loaded by the original assignment too. If neither of these
3045 cases happen, the original statement can be removed. Most of this is done
3046 by load_assign_lhs_subreplacements.
3048 In the second case, we would like to store all RHS scalarized components
3049 directly into LHS and if they cover the aggregate completely, remove the
3050 statement too. In the third case, we want the LHS components to be loaded
3051 directly from the RHS (DSE will remove the original statement if it
3054 This is a bit complex but manageable when types match and when unions do
3055 not cause confusion in a way that we cannot really load a component of LHS
3056 from the RHS or vice versa (the access representing this level can have
3057 subaccesses that are accessible only through a different union field at a
3058 higher level - different from the one used in the examined expression).
3061 Therefore, I specially handle a fourth case, happening when there is a
3062 specific type cast or it is impossible to locate a scalarized subaccess on
3063 the other side of the expression. If that happens, I simply "refresh" the
3064 RHS by storing in it is scalarized components leave the original statement
3065 there to do the copying and then load the scalar replacements of the LHS.
3066 This is what the first branch does. */
3068 if (modify_this_stmt
3069 || gimple_has_volatile_ops (*stmt)
3070 || contains_vce_or_bfcref_p (rhs)
3071 || contains_vce_or_bfcref_p (lhs))
3073 if (access_has_children_p (racc))
3074 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3075 gsi, false, false, loc);
3076 if (access_has_children_p (lacc))
3077 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
3078 gsi, true, true, loc);
3079 sra_stats.separate_lhs_rhs_handling++;
3081 /* This gimplification must be done after generate_subtree_copies,
3082 lest we insert the subtree copies in the middle of the gimplified
3084 if (force_gimple_rhs)
3085 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
3086 true, GSI_SAME_STMT);
3087 if (gimple_assign_rhs1 (*stmt) != rhs)
3089 modify_this_stmt = true;
3090 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
3091 gcc_assert (*stmt == gsi_stmt (orig_gsi));
3094 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3098 if (access_has_children_p (lacc) && access_has_children_p (racc))
3100 gimple_stmt_iterator orig_gsi = *gsi;
3101 enum unscalarized_data_handling refreshed;
3103 if (lacc->grp_read && !lacc->grp_covered)
3104 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
3106 refreshed = SRA_UDH_NONE;
3108 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
3109 &orig_gsi, gsi, &refreshed);
3110 if (refreshed != SRA_UDH_RIGHT)
3113 unlink_stmt_vdef (*stmt);
3114 gsi_remove (&orig_gsi, true);
3115 sra_stats.deleted++;
3116 return SRA_AM_REMOVED;
3121 if (access_has_children_p (racc)
3122 && !racc->grp_unscalarized_data)
3126 fprintf (dump_file, "Removing load: ");
3127 print_gimple_stmt (dump_file, *stmt, 0, 0);
3129 generate_subtree_copies (racc->first_child, lhs,
3130 racc->offset, 0, 0, gsi,
3132 gcc_assert (*stmt == gsi_stmt (*gsi));
3133 unlink_stmt_vdef (*stmt);
3134 gsi_remove (gsi, true);
3135 sra_stats.deleted++;
3136 return SRA_AM_REMOVED;
3138 /* Restore the aggregate RHS from its components so the
3139 prevailing aggregate copy does the right thing. */
3140 if (access_has_children_p (racc))
3141 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3142 gsi, false, false, loc);
3143 /* Re-load the components of the aggregate copy destination.
3144 But use the RHS aggregate to load from to expose more
3145 optimization opportunities. */
3146 if (access_has_children_p (lacc))
3147 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
3148 0, 0, gsi, true, true, loc);
3155 /* Traverse the function body and all modifications as decided in
3156 analyze_all_variable_accesses. Return true iff the CFG has been
3160 sra_modify_function_body (void)
3162 bool cfg_changed = false;
3167 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3168 while (!gsi_end_p (gsi))
3170 gimple stmt = gsi_stmt (gsi);
3171 enum assignment_mod_result assign_result;
3172 bool modified = false, deleted = false;
3176 switch (gimple_code (stmt))
3179 t = gimple_return_retval_ptr (stmt);
3180 if (*t != NULL_TREE)
3181 modified |= sra_modify_expr (t, &gsi, false);
3185 assign_result = sra_modify_assign (&stmt, &gsi);
3186 modified |= assign_result == SRA_AM_MODIFIED;
3187 deleted = assign_result == SRA_AM_REMOVED;
3191 /* Operands must be processed before the lhs. */
3192 for (i = 0; i < gimple_call_num_args (stmt); i++)
3194 t = gimple_call_arg_ptr (stmt, i);
3195 modified |= sra_modify_expr (t, &gsi, false);
3198 if (gimple_call_lhs (stmt))
3200 t = gimple_call_lhs_ptr (stmt);
3201 modified |= sra_modify_expr (t, &gsi, true);
3206 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3208 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3209 modified |= sra_modify_expr (t, &gsi, false);
3211 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3213 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3214 modified |= sra_modify_expr (t, &gsi, true);
3225 if (maybe_clean_eh_stmt (stmt)
3226 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3237 /* Generate statements initializing scalar replacements of parts of function
3241 initialize_parameter_reductions (void)
3243 gimple_stmt_iterator gsi;
3244 gimple_seq seq = NULL;
3247 for (parm = DECL_ARGUMENTS (current_function_decl);
3249 parm = DECL_CHAIN (parm))
3251 VEC (access_p, heap) *access_vec;
3252 struct access *access;
3254 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3256 access_vec = get_base_access_vector (parm);
3262 seq = gimple_seq_alloc ();
3263 gsi = gsi_start (seq);
3266 for (access = VEC_index (access_p, access_vec, 0);
3268 access = access->next_grp)
3269 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3270 EXPR_LOCATION (parm));
3274 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3277 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3278 it reveals there are components of some aggregates to be scalarized, it runs
3279 the required transformations. */
3281 perform_intra_sra (void)
3286 if (!find_var_candidates ())
3289 if (!scan_function ())
3292 if (!analyze_all_variable_accesses ())
3295 if (sra_modify_function_body ())
3296 ret = TODO_update_ssa | TODO_cleanup_cfg;
3298 ret = TODO_update_ssa;
3299 initialize_parameter_reductions ();
3301 statistics_counter_event (cfun, "Scalar replacements created",
3302 sra_stats.replacements);
3303 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3304 statistics_counter_event (cfun, "Subtree copy stmts",
3305 sra_stats.subtree_copies);
3306 statistics_counter_event (cfun, "Subreplacement stmts",
3307 sra_stats.subreplacements);
3308 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3309 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3310 sra_stats.separate_lhs_rhs_handling);
3313 sra_deinitialize ();
3317 /* Perform early intraprocedural SRA. */
3319 early_intra_sra (void)
3321 sra_mode = SRA_MODE_EARLY_INTRA;
3322 return perform_intra_sra ();
3325 /* Perform "late" intraprocedural SRA. */
3327 late_intra_sra (void)
3329 sra_mode = SRA_MODE_INTRA;
3330 return perform_intra_sra ();
3335 gate_intra_sra (void)
3337 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3341 struct gimple_opt_pass pass_sra_early =
3346 gate_intra_sra, /* gate */
3347 early_intra_sra, /* execute */
3350 0, /* static_pass_number */
3351 TV_TREE_SRA, /* tv_id */
3352 PROP_cfg | PROP_ssa, /* properties_required */
3353 0, /* properties_provided */
3354 0, /* properties_destroyed */
3355 0, /* todo_flags_start */
3358 | TODO_verify_ssa /* todo_flags_finish */
3362 struct gimple_opt_pass pass_sra =
3367 gate_intra_sra, /* gate */
3368 late_intra_sra, /* execute */
3371 0, /* static_pass_number */
3372 TV_TREE_SRA, /* tv_id */
3373 PROP_cfg | PROP_ssa, /* properties_required */
3374 0, /* properties_provided */
3375 0, /* properties_destroyed */
3376 TODO_update_address_taken, /* todo_flags_start */
3379 | TODO_verify_ssa /* todo_flags_finish */
3384 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3388 is_unused_scalar_param (tree parm)
3391 return (is_gimple_reg (parm)
3392 && (!(name = gimple_default_def (cfun, parm))
3393 || has_zero_uses (name)));
3396 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3397 examine whether there are any direct or otherwise infeasible ones. If so,
3398 return true, otherwise return false. PARM must be a gimple register with a
3399 non-NULL default definition. */
3402 ptr_parm_has_direct_uses (tree parm)
3404 imm_use_iterator ui;
3406 tree name = gimple_default_def (cfun, parm);
3409 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3412 use_operand_p use_p;
3414 if (is_gimple_debug (stmt))
3417 /* Valid uses include dereferences on the lhs and the rhs. */
3418 if (gimple_has_lhs (stmt))
3420 tree lhs = gimple_get_lhs (stmt);
3421 while (handled_component_p (lhs))
3422 lhs = TREE_OPERAND (lhs, 0);
3423 if (TREE_CODE (lhs) == MEM_REF
3424 && TREE_OPERAND (lhs, 0) == name
3425 && integer_zerop (TREE_OPERAND (lhs, 1))
3426 && types_compatible_p (TREE_TYPE (lhs),
3427 TREE_TYPE (TREE_TYPE (name)))
3428 && !TREE_THIS_VOLATILE (lhs))
3431 if (gimple_assign_single_p (stmt))
3433 tree rhs = gimple_assign_rhs1 (stmt);
3434 while (handled_component_p (rhs))
3435 rhs = TREE_OPERAND (rhs, 0);
3436 if (TREE_CODE (rhs) == MEM_REF
3437 && TREE_OPERAND (rhs, 0) == name
3438 && integer_zerop (TREE_OPERAND (rhs, 1))
3439 && types_compatible_p (TREE_TYPE (rhs),
3440 TREE_TYPE (TREE_TYPE (name)))
3441 && !TREE_THIS_VOLATILE (rhs))
3444 else if (is_gimple_call (stmt))
3447 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3449 tree arg = gimple_call_arg (stmt, i);
3450 while (handled_component_p (arg))
3451 arg = TREE_OPERAND (arg, 0);
3452 if (TREE_CODE (arg) == MEM_REF
3453 && TREE_OPERAND (arg, 0) == name
3454 && integer_zerop (TREE_OPERAND (arg, 1))
3455 && types_compatible_p (TREE_TYPE (arg),
3456 TREE_TYPE (TREE_TYPE (name)))
3457 && !TREE_THIS_VOLATILE (arg))
3462 /* If the number of valid uses does not match the number of
3463 uses in this stmt there is an unhandled use. */
3464 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3471 BREAK_FROM_IMM_USE_STMT (ui);
3477 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3478 them in candidate_bitmap. Note that these do not necessarily include
3479 parameter which are unused and thus can be removed. Return true iff any
3480 such candidate has been found. */
3483 find_param_candidates (void)
3490 for (parm = DECL_ARGUMENTS (current_function_decl);
3492 parm = DECL_CHAIN (parm))
3494 tree type = TREE_TYPE (parm);
3498 if (TREE_THIS_VOLATILE (parm)
3499 || TREE_ADDRESSABLE (parm)
3500 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3503 if (is_unused_scalar_param (parm))
3509 if (POINTER_TYPE_P (type))
3511 type = TREE_TYPE (type);
3513 if (TREE_CODE (type) == FUNCTION_TYPE
3514 || TYPE_VOLATILE (type)
3515 || (TREE_CODE (type) == ARRAY_TYPE
3516 && TYPE_NONALIASED_COMPONENT (type))
3517 || !is_gimple_reg (parm)
3518 || is_va_list_type (type)
3519 || ptr_parm_has_direct_uses (parm))
3522 else if (!AGGREGATE_TYPE_P (type))
3525 if (!COMPLETE_TYPE_P (type)
3526 || !host_integerp (TYPE_SIZE (type), 1)
3527 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3528 || (AGGREGATE_TYPE_P (type)
3529 && type_internals_preclude_sra_p (type, &msg)))
3532 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3534 if (dump_file && (dump_flags & TDF_DETAILS))
3536 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3537 print_generic_expr (dump_file, parm, 0);
3538 fprintf (dump_file, "\n");
3542 func_param_count = count;
3546 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3550 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3553 struct access *repr = (struct access *) data;
3555 repr->grp_maybe_modified = 1;
3559 /* Analyze what representatives (in linked lists accessible from
3560 REPRESENTATIVES) can be modified by side effects of statements in the
3561 current function. */
3564 analyze_modified_params (VEC (access_p, heap) *representatives)
3568 for (i = 0; i < func_param_count; i++)
3570 struct access *repr;
3572 for (repr = VEC_index (access_p, representatives, i);
3574 repr = repr->next_grp)
3576 struct access *access;
3580 if (no_accesses_p (repr))
3582 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3583 || repr->grp_maybe_modified)
3586 ao_ref_init (&ar, repr->expr);
3587 visited = BITMAP_ALLOC (NULL);
3588 for (access = repr; access; access = access->next_sibling)
3590 /* All accesses are read ones, otherwise grp_maybe_modified would
3591 be trivially set. */
3592 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3593 mark_maybe_modified, repr, &visited);
3594 if (repr->grp_maybe_modified)
3597 BITMAP_FREE (visited);
3602 /* Propagate distances in bb_dereferences in the opposite direction than the
3603 control flow edges, in each step storing the maximum of the current value
3604 and the minimum of all successors. These steps are repeated until the table
3605 stabilizes. Note that BBs which might terminate the functions (according to
3606 final_bbs bitmap) never updated in this way. */
3609 propagate_dereference_distances (void)
3611 VEC (basic_block, heap) *queue;
3614 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3615 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3618 VEC_quick_push (basic_block, queue, bb);
3622 while (!VEC_empty (basic_block, queue))
3626 bool change = false;
3629 bb = VEC_pop (basic_block, queue);
3632 if (bitmap_bit_p (final_bbs, bb->index))
3635 for (i = 0; i < func_param_count; i++)
3637 int idx = bb->index * func_param_count + i;
3639 HOST_WIDE_INT inh = 0;
3641 FOR_EACH_EDGE (e, ei, bb->succs)
3643 int succ_idx = e->dest->index * func_param_count + i;
3645 if (e->src == EXIT_BLOCK_PTR)
3651 inh = bb_dereferences [succ_idx];
3653 else if (bb_dereferences [succ_idx] < inh)
3654 inh = bb_dereferences [succ_idx];
3657 if (!first && bb_dereferences[idx] < inh)
3659 bb_dereferences[idx] = inh;
3664 if (change && !bitmap_bit_p (final_bbs, bb->index))
3665 FOR_EACH_EDGE (e, ei, bb->preds)
3670 e->src->aux = e->src;
3671 VEC_quick_push (basic_block, queue, e->src);
3675 VEC_free (basic_block, heap, queue);
3678 /* Dump a dereferences TABLE with heading STR to file F. */
3681 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3685 fprintf (dump_file, str);
3686 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3688 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3689 if (bb != EXIT_BLOCK_PTR)
3692 for (i = 0; i < func_param_count; i++)
3694 int idx = bb->index * func_param_count + i;
3695 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3700 fprintf (dump_file, "\n");
3703 /* Determine what (parts of) parameters passed by reference that are not
3704 assigned to are not certainly dereferenced in this function and thus the
3705 dereferencing cannot be safely moved to the caller without potentially
3706 introducing a segfault. Mark such REPRESENTATIVES as
3707 grp_not_necessarilly_dereferenced.
3709 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3710 part is calculated rather than simple booleans are calculated for each
3711 pointer parameter to handle cases when only a fraction of the whole
3712 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3715 The maximum dereference distances for each pointer parameter and BB are
3716 already stored in bb_dereference. This routine simply propagates these
3717 values upwards by propagate_dereference_distances and then compares the
3718 distances of individual parameters in the ENTRY BB to the equivalent
3719 distances of each representative of a (fraction of a) parameter. */
3722 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3726 if (dump_file && (dump_flags & TDF_DETAILS))
3727 dump_dereferences_table (dump_file,
3728 "Dereference table before propagation:\n",
3731 propagate_dereference_distances ();
3733 if (dump_file && (dump_flags & TDF_DETAILS))
3734 dump_dereferences_table (dump_file,
3735 "Dereference table after propagation:\n",
3738 for (i = 0; i < func_param_count; i++)
3740 struct access *repr = VEC_index (access_p, representatives, i);
3741 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3743 if (!repr || no_accesses_p (repr))
3748 if ((repr->offset + repr->size) > bb_dereferences[idx])
3749 repr->grp_not_necessarilly_dereferenced = 1;
3750 repr = repr->next_grp;
3756 /* Return the representative access for the parameter declaration PARM if it is
3757 a scalar passed by reference which is not written to and the pointer value
3758 is not used directly. Thus, if it is legal to dereference it in the caller
3759 and we can rule out modifications through aliases, such parameter should be
3760 turned into one passed by value. Return NULL otherwise. */
3762 static struct access *
3763 unmodified_by_ref_scalar_representative (tree parm)
3765 int i, access_count;
3766 struct access *repr;
3767 VEC (access_p, heap) *access_vec;
3769 access_vec = get_base_access_vector (parm);
3770 gcc_assert (access_vec);
3771 repr = VEC_index (access_p, access_vec, 0);
3774 repr->group_representative = repr;
3776 access_count = VEC_length (access_p, access_vec);
3777 for (i = 1; i < access_count; i++)
3779 struct access *access = VEC_index (access_p, access_vec, i);
3782 access->group_representative = repr;
3783 access->next_sibling = repr->next_sibling;
3784 repr->next_sibling = access;
3788 repr->grp_scalar_ptr = 1;
3792 /* Return true iff this access precludes IPA-SRA of the parameter it is
3796 access_precludes_ipa_sra_p (struct access *access)
3798 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3799 is incompatible assign in a call statement (and possibly even in asm
3800 statements). This can be relaxed by using a new temporary but only for
3801 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3802 intraprocedural SRA we deal with this by keeping the old aggregate around,
3803 something we cannot do in IPA-SRA.) */
3805 && (is_gimple_call (access->stmt)
3806 || gimple_code (access->stmt) == GIMPLE_ASM))
3809 if (STRICT_ALIGNMENT
3810 && tree_non_aligned_mem_p (access->expr, TYPE_ALIGN (access->type)))
3817 /* Sort collected accesses for parameter PARM, identify representatives for
3818 each accessed region and link them together. Return NULL if there are
3819 different but overlapping accesses, return the special ptr value meaning
3820 there are no accesses for this parameter if that is the case and return the
3821 first representative otherwise. Set *RO_GRP if there is a group of accesses
3822 with only read (i.e. no write) accesses. */
3824 static struct access *
3825 splice_param_accesses (tree parm, bool *ro_grp)
3827 int i, j, access_count, group_count;
3828 int agg_size, total_size = 0;
3829 struct access *access, *res, **prev_acc_ptr = &res;
3830 VEC (access_p, heap) *access_vec;
3832 access_vec = get_base_access_vector (parm);
3834 return &no_accesses_representant;
3835 access_count = VEC_length (access_p, access_vec);
3837 VEC_qsort (access_p, access_vec, compare_access_positions);
3842 while (i < access_count)
3846 access = VEC_index (access_p, access_vec, i);
3847 modification = access->write;
3848 if (access_precludes_ipa_sra_p (access))
3850 a1_alias_type = reference_alias_ptr_type (access->expr);
3852 /* Access is about to become group representative unless we find some
3853 nasty overlap which would preclude us from breaking this parameter
3857 while (j < access_count)
3859 struct access *ac2 = VEC_index (access_p, access_vec, j);
3860 if (ac2->offset != access->offset)
3862 /* All or nothing law for parameters. */
3863 if (access->offset + access->size > ac2->offset)
3868 else if (ac2->size != access->size)
3871 if (access_precludes_ipa_sra_p (ac2)
3872 || (ac2->type != access->type
3873 && (TREE_ADDRESSABLE (ac2->type)
3874 || TREE_ADDRESSABLE (access->type)))
3875 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3878 modification |= ac2->write;
3879 ac2->group_representative = access;
3880 ac2->next_sibling = access->next_sibling;
3881 access->next_sibling = ac2;
3886 access->grp_maybe_modified = modification;
3889 *prev_acc_ptr = access;
3890 prev_acc_ptr = &access->next_grp;
3891 total_size += access->size;
3895 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3896 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3898 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3899 if (total_size >= agg_size)
3902 gcc_assert (group_count > 0);
3906 /* Decide whether parameters with representative accesses given by REPR should
3907 be reduced into components. */
3910 decide_one_param_reduction (struct access *repr)
3912 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3917 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3918 gcc_assert (cur_parm_size > 0);
3920 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3923 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3928 agg_size = cur_parm_size;
3934 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3935 print_generic_expr (dump_file, parm, 0);
3936 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3937 for (acc = repr; acc; acc = acc->next_grp)
3938 dump_access (dump_file, acc, true);
3942 new_param_count = 0;
3944 for (; repr; repr = repr->next_grp)
3946 gcc_assert (parm == repr->base);
3948 /* Taking the address of a non-addressable field is verboten. */
3949 if (by_ref && repr->non_addressable)
3952 /* Do not decompose a non-BLKmode param in a way that would
3953 create BLKmode params. Especially for by-reference passing
3954 (thus, pointer-type param) this is hardly worthwhile. */
3955 if (DECL_MODE (parm) != BLKmode
3956 && TYPE_MODE (repr->type) == BLKmode)
3959 if (!by_ref || (!repr->grp_maybe_modified
3960 && !repr->grp_not_necessarilly_dereferenced))
3961 total_size += repr->size;
3963 total_size += cur_parm_size;
3968 gcc_assert (new_param_count > 0);
3970 if (optimize_function_for_size_p (cfun))
3971 parm_size_limit = cur_parm_size;
3973 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3976 if (total_size < agg_size
3977 && total_size <= parm_size_limit)
3980 fprintf (dump_file, " ....will be split into %i components\n",
3982 return new_param_count;
3988 /* The order of the following enums is important, we need to do extra work for
3989 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3990 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3991 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3993 /* Identify representatives of all accesses to all candidate parameters for
3994 IPA-SRA. Return result based on what representatives have been found. */
3996 static enum ipa_splicing_result
3997 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3999 enum ipa_splicing_result result = NO_GOOD_ACCESS;
4001 struct access *repr;
4003 *representatives = VEC_alloc (access_p, heap, func_param_count);
4005 for (parm = DECL_ARGUMENTS (current_function_decl);
4007 parm = DECL_CHAIN (parm))
4009 if (is_unused_scalar_param (parm))
4011 VEC_quick_push (access_p, *representatives,
4012 &no_accesses_representant);
4013 if (result == NO_GOOD_ACCESS)
4014 result = UNUSED_PARAMS;
4016 else if (POINTER_TYPE_P (TREE_TYPE (parm))
4017 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
4018 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4020 repr = unmodified_by_ref_scalar_representative (parm);
4021 VEC_quick_push (access_p, *representatives, repr);
4023 result = UNMODIF_BY_REF_ACCESSES;
4025 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4027 bool ro_grp = false;
4028 repr = splice_param_accesses (parm, &ro_grp);
4029 VEC_quick_push (access_p, *representatives, repr);
4031 if (repr && !no_accesses_p (repr))
4033 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4036 result = UNMODIF_BY_REF_ACCESSES;
4037 else if (result < MODIF_BY_REF_ACCESSES)
4038 result = MODIF_BY_REF_ACCESSES;
4040 else if (result < BY_VAL_ACCESSES)
4041 result = BY_VAL_ACCESSES;
4043 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
4044 result = UNUSED_PARAMS;
4047 VEC_quick_push (access_p, *representatives, NULL);
4050 if (result == NO_GOOD_ACCESS)
4052 VEC_free (access_p, heap, *representatives);
4053 *representatives = NULL;
4054 return NO_GOOD_ACCESS;
4060 /* Return the index of BASE in PARMS. Abort if it is not found. */
4063 get_param_index (tree base, VEC(tree, heap) *parms)
4067 len = VEC_length (tree, parms);
4068 for (i = 0; i < len; i++)
4069 if (VEC_index (tree, parms, i) == base)
4074 /* Convert the decisions made at the representative level into compact
4075 parameter adjustments. REPRESENTATIVES are pointers to first
4076 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4077 final number of adjustments. */
4079 static ipa_parm_adjustment_vec
4080 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
4081 int adjustments_count)
4083 VEC (tree, heap) *parms;
4084 ipa_parm_adjustment_vec adjustments;
4088 gcc_assert (adjustments_count > 0);
4089 parms = ipa_get_vector_of_formal_parms (current_function_decl);
4090 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
4091 parm = DECL_ARGUMENTS (current_function_decl);
4092 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
4094 struct access *repr = VEC_index (access_p, representatives, i);
4096 if (!repr || no_accesses_p (repr))
4098 struct ipa_parm_adjustment *adj;
4100 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4101 memset (adj, 0, sizeof (*adj));
4102 adj->base_index = get_param_index (parm, parms);
4105 adj->copy_param = 1;
4107 adj->remove_param = 1;
4111 struct ipa_parm_adjustment *adj;
4112 int index = get_param_index (parm, parms);
4114 for (; repr; repr = repr->next_grp)
4116 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4117 memset (adj, 0, sizeof (*adj));
4118 gcc_assert (repr->base == parm);
4119 adj->base_index = index;
4120 adj->base = repr->base;
4121 adj->type = repr->type;
4122 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
4123 adj->offset = repr->offset;
4124 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
4125 && (repr->grp_maybe_modified
4126 || repr->grp_not_necessarilly_dereferenced));
4131 VEC_free (tree, heap, parms);
4135 /* Analyze the collected accesses and produce a plan what to do with the
4136 parameters in the form of adjustments, NULL meaning nothing. */
4138 static ipa_parm_adjustment_vec
4139 analyze_all_param_acesses (void)
4141 enum ipa_splicing_result repr_state;
4142 bool proceed = false;
4143 int i, adjustments_count = 0;
4144 VEC (access_p, heap) *representatives;
4145 ipa_parm_adjustment_vec adjustments;
4147 repr_state = splice_all_param_accesses (&representatives);
4148 if (repr_state == NO_GOOD_ACCESS)
4151 /* If there are any parameters passed by reference which are not modified
4152 directly, we need to check whether they can be modified indirectly. */
4153 if (repr_state == UNMODIF_BY_REF_ACCESSES)
4155 analyze_caller_dereference_legality (representatives);
4156 analyze_modified_params (representatives);
4159 for (i = 0; i < func_param_count; i++)
4161 struct access *repr = VEC_index (access_p, representatives, i);
4163 if (repr && !no_accesses_p (repr))
4165 if (repr->grp_scalar_ptr)
4167 adjustments_count++;
4168 if (repr->grp_not_necessarilly_dereferenced
4169 || repr->grp_maybe_modified)
4170 VEC_replace (access_p, representatives, i, NULL);
4174 sra_stats.scalar_by_ref_to_by_val++;
4179 int new_components = decide_one_param_reduction (repr);
4181 if (new_components == 0)
4183 VEC_replace (access_p, representatives, i, NULL);
4184 adjustments_count++;
4188 adjustments_count += new_components;
4189 sra_stats.aggregate_params_reduced++;
4190 sra_stats.param_reductions_created += new_components;
4197 if (no_accesses_p (repr))
4200 sra_stats.deleted_unused_parameters++;
4202 adjustments_count++;
4206 if (!proceed && dump_file)
4207 fprintf (dump_file, "NOT proceeding to change params.\n");
4210 adjustments = turn_representatives_into_adjustments (representatives,
4215 VEC_free (access_p, heap, representatives);
4219 /* If a parameter replacement identified by ADJ does not yet exist in the form
4220 of declaration, create it and record it, otherwise return the previously
4224 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4227 if (!adj->new_ssa_base)
4229 char *pretty_name = make_fancy_name (adj->base);
4231 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4232 DECL_NAME (repl) = get_identifier (pretty_name);
4233 obstack_free (&name_obstack, pretty_name);
4235 add_referenced_var (repl);
4236 adj->new_ssa_base = repl;
4239 repl = adj->new_ssa_base;
4243 /* Find the first adjustment for a particular parameter BASE in a vector of
4244 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4247 static struct ipa_parm_adjustment *
4248 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4252 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4253 for (i = 0; i < len; i++)
4255 struct ipa_parm_adjustment *adj;
4257 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4258 if (!adj->copy_param && adj->base == base)
4265 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4266 removed because its value is not used, replace the SSA_NAME with a one
4267 relating to a created VAR_DECL together all of its uses and return true.
4268 ADJUSTMENTS is a pointer to an adjustments vector. */
4271 replace_removed_params_ssa_names (gimple stmt,
4272 ipa_parm_adjustment_vec adjustments)
4274 struct ipa_parm_adjustment *adj;
4275 tree lhs, decl, repl, name;
4277 if (gimple_code (stmt) == GIMPLE_PHI)
4278 lhs = gimple_phi_result (stmt);
4279 else if (is_gimple_assign (stmt))
4280 lhs = gimple_assign_lhs (stmt);
4281 else if (is_gimple_call (stmt))
4282 lhs = gimple_call_lhs (stmt);
4286 if (TREE_CODE (lhs) != SSA_NAME)
4288 decl = SSA_NAME_VAR (lhs);
4289 if (TREE_CODE (decl) != PARM_DECL)
4292 adj = get_adjustment_for_base (adjustments, decl);
4296 repl = get_replaced_param_substitute (adj);
4297 name = make_ssa_name (repl, stmt);
4301 fprintf (dump_file, "replacing an SSA name of a removed param ");
4302 print_generic_expr (dump_file, lhs, 0);
4303 fprintf (dump_file, " with ");
4304 print_generic_expr (dump_file, name, 0);
4305 fprintf (dump_file, "\n");
4308 if (is_gimple_assign (stmt))
4309 gimple_assign_set_lhs (stmt, name);
4310 else if (is_gimple_call (stmt))
4311 gimple_call_set_lhs (stmt, name);
4313 gimple_phi_set_result (stmt, name);
4315 replace_uses_by (lhs, name);
4316 release_ssa_name (lhs);
4320 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4321 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4322 specifies whether the function should care about type incompatibility the
4323 current and new expressions. If it is false, the function will leave
4324 incompatibility issues to the caller. Return true iff the expression
4328 sra_ipa_modify_expr (tree *expr, bool convert,
4329 ipa_parm_adjustment_vec adjustments)
4332 struct ipa_parm_adjustment *adj, *cand = NULL;
4333 HOST_WIDE_INT offset, size, max_size;
4336 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4338 if (TREE_CODE (*expr) == BIT_FIELD_REF
4339 || TREE_CODE (*expr) == IMAGPART_EXPR
4340 || TREE_CODE (*expr) == REALPART_EXPR)
4342 expr = &TREE_OPERAND (*expr, 0);
4346 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4347 if (!base || size == -1 || max_size == -1)
4350 if (TREE_CODE (base) == MEM_REF)
4352 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4353 base = TREE_OPERAND (base, 0);
4356 base = get_ssa_base_param (base);
4357 if (!base || TREE_CODE (base) != PARM_DECL)
4360 for (i = 0; i < len; i++)
4362 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4364 if (adj->base == base &&
4365 (adj->offset == offset || adj->remove_param))
4371 if (!cand || cand->copy_param || cand->remove_param)
4375 src = build_simple_mem_ref (cand->reduction);
4377 src = cand->reduction;
4379 if (dump_file && (dump_flags & TDF_DETAILS))
4381 fprintf (dump_file, "About to replace expr ");
4382 print_generic_expr (dump_file, *expr, 0);
4383 fprintf (dump_file, " with ");
4384 print_generic_expr (dump_file, src, 0);
4385 fprintf (dump_file, "\n");
4388 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4390 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4398 /* If the statement pointed to by STMT_PTR contains any expressions that need
4399 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4400 potential type incompatibilities (GSI is used to accommodate conversion
4401 statements and must point to the statement). Return true iff the statement
4405 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4406 ipa_parm_adjustment_vec adjustments)
4408 gimple stmt = *stmt_ptr;
4409 tree *lhs_p, *rhs_p;
4412 if (!gimple_assign_single_p (stmt))
4415 rhs_p = gimple_assign_rhs1_ptr (stmt);
4416 lhs_p = gimple_assign_lhs_ptr (stmt);
4418 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4419 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4422 tree new_rhs = NULL_TREE;
4424 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4426 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4428 /* V_C_Es of constructors can cause trouble (PR 42714). */
4429 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4430 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4432 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4435 new_rhs = fold_build1_loc (gimple_location (stmt),
4436 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4439 else if (REFERENCE_CLASS_P (*rhs_p)
4440 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4441 && !is_gimple_reg (*lhs_p))
4442 /* This can happen when an assignment in between two single field
4443 structures is turned into an assignment in between two pointers to
4444 scalars (PR 42237). */
4449 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4450 true, GSI_SAME_STMT);
4452 gimple_assign_set_rhs_from_tree (gsi, tmp);
4461 /* Traverse the function body and all modifications as described in
4462 ADJUSTMENTS. Return true iff the CFG has been changed. */
4465 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4467 bool cfg_changed = false;
4472 gimple_stmt_iterator gsi;
4474 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4475 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4477 gsi = gsi_start_bb (bb);
4478 while (!gsi_end_p (gsi))
4480 gimple stmt = gsi_stmt (gsi);
4481 bool modified = false;
4485 switch (gimple_code (stmt))
4488 t = gimple_return_retval_ptr (stmt);
4489 if (*t != NULL_TREE)
4490 modified |= sra_ipa_modify_expr (t, true, adjustments);
4494 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4495 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4499 /* Operands must be processed before the lhs. */
4500 for (i = 0; i < gimple_call_num_args (stmt); i++)
4502 t = gimple_call_arg_ptr (stmt, i);
4503 modified |= sra_ipa_modify_expr (t, true, adjustments);
4506 if (gimple_call_lhs (stmt))
4508 t = gimple_call_lhs_ptr (stmt);
4509 modified |= sra_ipa_modify_expr (t, false, adjustments);
4510 modified |= replace_removed_params_ssa_names (stmt,
4516 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4518 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4519 modified |= sra_ipa_modify_expr (t, true, adjustments);
4521 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4523 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4524 modified |= sra_ipa_modify_expr (t, false, adjustments);
4535 if (maybe_clean_eh_stmt (stmt)
4536 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4546 /* Call gimple_debug_bind_reset_value on all debug statements describing
4547 gimple register parameters that are being removed or replaced. */
4550 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4553 gimple_stmt_iterator *gsip = NULL, gsi;
4555 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR))
4557 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
4560 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4561 for (i = 0; i < len; i++)
4563 struct ipa_parm_adjustment *adj;
4564 imm_use_iterator ui;
4565 gimple stmt, def_temp;
4566 tree name, vexpr, copy = NULL_TREE;
4567 use_operand_p use_p;
4569 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4570 if (adj->copy_param || !is_gimple_reg (adj->base))
4572 name = gimple_default_def (cfun, adj->base);
4575 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4577 /* All other users must have been removed by
4578 ipa_sra_modify_function_body. */
4579 gcc_assert (is_gimple_debug (stmt));
4580 if (vexpr == NULL && gsip != NULL)
4582 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4583 vexpr = make_node (DEBUG_EXPR_DECL);
4584 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4586 DECL_ARTIFICIAL (vexpr) = 1;
4587 TREE_TYPE (vexpr) = TREE_TYPE (name);
4588 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4589 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4593 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4594 SET_USE (use_p, vexpr);
4597 gimple_debug_bind_reset_value (stmt);
4600 /* Create a VAR_DECL for debug info purposes. */
4601 if (!DECL_IGNORED_P (adj->base))
4603 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4604 VAR_DECL, DECL_NAME (adj->base),
4605 TREE_TYPE (adj->base));
4606 if (DECL_PT_UID_SET_P (adj->base))
4607 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4608 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4609 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4610 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4611 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4612 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4613 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4614 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4615 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4616 SET_DECL_RTL (copy, 0);
4617 TREE_USED (copy) = 1;
4618 DECL_CONTEXT (copy) = current_function_decl;
4619 add_referenced_var (copy);
4620 add_local_decl (cfun, copy);
4622 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4623 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4625 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4627 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4629 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4631 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4633 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4638 /* Return false iff all callers have at least as many actual arguments as there
4639 are formal parameters in the current function. */
4642 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4643 void *data ATTRIBUTE_UNUSED)
4645 struct cgraph_edge *cs;
4646 for (cs = node->callers; cs; cs = cs->next_caller)
4647 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4653 /* Convert all callers of NODE. */
4656 convert_callers_for_node (struct cgraph_node *node,
4659 ipa_parm_adjustment_vec adjustments = (ipa_parm_adjustment_vec)data;
4660 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4661 struct cgraph_edge *cs;
4663 for (cs = node->callers; cs; cs = cs->next_caller)
4665 current_function_decl = cs->caller->decl;
4666 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4669 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4670 cs->caller->uid, cs->callee->uid,
4671 cgraph_node_name (cs->caller),
4672 cgraph_node_name (cs->callee));
4674 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4679 for (cs = node->callers; cs; cs = cs->next_caller)
4680 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4681 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4682 compute_inline_parameters (cs->caller, true);
4683 BITMAP_FREE (recomputed_callers);
4688 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4691 convert_callers (struct cgraph_node *node, tree old_decl,
4692 ipa_parm_adjustment_vec adjustments)
4694 tree old_cur_fndecl = current_function_decl;
4695 basic_block this_block;
4697 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4698 adjustments, false);
4700 current_function_decl = old_cur_fndecl;
4702 if (!encountered_recursive_call)
4705 FOR_EACH_BB (this_block)
4707 gimple_stmt_iterator gsi;
4709 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4711 gimple stmt = gsi_stmt (gsi);
4713 if (gimple_code (stmt) != GIMPLE_CALL)
4715 call_fndecl = gimple_call_fndecl (stmt);
4716 if (call_fndecl == old_decl)
4719 fprintf (dump_file, "Adjusting recursive call");
4720 gimple_call_set_fndecl (stmt, node->decl);
4721 ipa_modify_call_arguments (NULL, stmt, adjustments);
4729 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4730 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4733 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4735 struct cgraph_node *new_node;
4737 VEC (cgraph_edge_p, heap) * redirect_callers = collect_callers_of_node (node);
4739 rebuild_cgraph_edges ();
4740 free_dominance_info (CDI_DOMINATORS);
4742 current_function_decl = NULL_TREE;
4744 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4745 false, NULL, NULL, "isra");
4746 current_function_decl = new_node->decl;
4747 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4749 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4750 cfg_changed = ipa_sra_modify_function_body (adjustments);
4751 sra_ipa_reset_debug_stmts (adjustments);
4752 convert_callers (new_node, node->decl, adjustments);
4753 cgraph_make_node_local (new_node);
4757 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4758 attributes, return true otherwise. NODE is the cgraph node of the current
4762 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4764 if (!cgraph_node_can_be_local_p (node))
4767 fprintf (dump_file, "Function not local to this compilation unit.\n");
4771 if (!node->local.can_change_signature)
4774 fprintf (dump_file, "Function can not change signature.\n");
4778 if (!tree_versionable_function_p (node->decl))
4781 fprintf (dump_file, "Function is not versionable.\n");
4785 if (DECL_VIRTUAL_P (current_function_decl))
4788 fprintf (dump_file, "Function is a virtual method.\n");
4792 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4793 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4796 fprintf (dump_file, "Function too big to be made truly local.\n");
4804 "Function has no callers in this compilation unit.\n");
4811 fprintf (dump_file, "Function uses stdarg. \n");
4815 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4821 /* Perform early interprocedural SRA. */
4824 ipa_early_sra (void)
4826 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4827 ipa_parm_adjustment_vec adjustments;
4830 if (!ipa_sra_preliminary_function_checks (node))
4834 sra_mode = SRA_MODE_EARLY_IPA;
4836 if (!find_param_candidates ())
4839 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4843 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4847 fprintf (dump_file, "There are callers with insufficient number of "
4852 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4854 * last_basic_block_for_function (cfun));
4855 final_bbs = BITMAP_ALLOC (NULL);
4858 if (encountered_apply_args)
4861 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4865 if (encountered_unchangable_recursive_call)
4868 fprintf (dump_file, "Function calls itself with insufficient "
4869 "number of arguments.\n");
4873 adjustments = analyze_all_param_acesses ();
4877 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4879 if (modify_function (node, adjustments))
4880 ret = TODO_update_ssa | TODO_cleanup_cfg;
4882 ret = TODO_update_ssa;
4883 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4885 statistics_counter_event (cfun, "Unused parameters deleted",
4886 sra_stats.deleted_unused_parameters);
4887 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4888 sra_stats.scalar_by_ref_to_by_val);
4889 statistics_counter_event (cfun, "Aggregate parameters broken up",
4890 sra_stats.aggregate_params_reduced);
4891 statistics_counter_event (cfun, "Aggregate parameter components created",
4892 sra_stats.param_reductions_created);
4895 BITMAP_FREE (final_bbs);
4896 free (bb_dereferences);
4898 sra_deinitialize ();
4902 /* Return if early ipa sra shall be performed. */
4904 ipa_early_sra_gate (void)
4906 return flag_ipa_sra && dbg_cnt (eipa_sra);
4909 struct gimple_opt_pass pass_early_ipa_sra =
4913 "eipa_sra", /* name */
4914 ipa_early_sra_gate, /* gate */
4915 ipa_early_sra, /* execute */
4918 0, /* static_pass_number */
4919 TV_IPA_SRA, /* tv_id */
4920 0, /* properties_required */
4921 0, /* properties_provided */
4922 0, /* properties_destroyed */
4923 0, /* todo_flags_start */
4924 TODO_dump_cgraph /* todo_flags_finish */