Daily bump.
[official-gcc.git] / gcc / tree-sra.c
blob4992b4c13bf4e7615de485ca4a35353b9239e8ae
1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
3 optimizers.
4 Copyright (C) 2008-2014 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
12 version.
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
17 for more details.
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
31 conversions.
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
49 accesses.
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. */
74 #include "config.h"
75 #include "system.h"
76 #include "coretypes.h"
77 #include "hash-table.h"
78 #include "alloc-pool.h"
79 #include "tm.h"
80 #include "tree.h"
81 #include "pointer-set.h"
82 #include "basic-block.h"
83 #include "tree-ssa-alias.h"
84 #include "internal-fn.h"
85 #include "tree-eh.h"
86 #include "gimple-expr.h"
87 #include "is-a.h"
88 #include "gimple.h"
89 #include "stor-layout.h"
90 #include "gimplify.h"
91 #include "gimple-iterator.h"
92 #include "gimplify-me.h"
93 #include "gimple-walk.h"
94 #include "bitmap.h"
95 #include "gimple-ssa.h"
96 #include "tree-cfg.h"
97 #include "tree-phinodes.h"
98 #include "ssa-iterators.h"
99 #include "stringpool.h"
100 #include "tree-ssanames.h"
101 #include "expr.h"
102 #include "tree-dfa.h"
103 #include "tree-ssa.h"
104 #include "tree-pass.h"
105 #include "ipa-prop.h"
106 #include "statistics.h"
107 #include "params.h"
108 #include "target.h"
109 #include "flags.h"
110 #include "dbgcnt.h"
111 #include "tree-inline.h"
112 #include "gimple-pretty-print.h"
113 #include "ipa-inline.h"
114 #include "ipa-utils.h"
116 /* Enumeration of all aggregate reductions we can do. */
117 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
118 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
119 SRA_MODE_INTRA }; /* late intraprocedural SRA */
121 /* Global variable describing which aggregate reduction we are performing at
122 the moment. */
123 static enum sra_mode sra_mode;
125 struct assign_link;
127 /* ACCESS represents each access to an aggregate variable (as a whole or a
128 part). It can also represent a group of accesses that refer to exactly the
129 same fragment of an aggregate (i.e. those that have exactly the same offset
130 and size). Such representatives for a single aggregate, once determined,
131 are linked in a linked list and have the group fields set.
133 Moreover, when doing intraprocedural SRA, a tree is built from those
134 representatives (by the means of first_child and next_sibling pointers), in
135 which all items in a subtree are "within" the root, i.e. their offset is
136 greater or equal to offset of the root and offset+size is smaller or equal
137 to offset+size of the root. Children of an access are sorted by offset.
139 Note that accesses to parts of vector and complex number types always
140 represented by an access to the whole complex number or a vector. It is a
141 duty of the modifying functions to replace them appropriately. */
143 struct access
145 /* Values returned by `get_ref_base_and_extent' for each component reference
146 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
147 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
148 HOST_WIDE_INT offset;
149 HOST_WIDE_INT size;
150 tree base;
152 /* Expression. It is context dependent so do not use it to create new
153 expressions to access the original aggregate. See PR 42154 for a
154 testcase. */
155 tree expr;
156 /* Type. */
157 tree type;
159 /* The statement this access belongs to. */
160 gimple stmt;
162 /* Next group representative for this aggregate. */
163 struct access *next_grp;
165 /* Pointer to the group representative. Pointer to itself if the struct is
166 the representative. */
167 struct access *group_representative;
169 /* If this access has any children (in terms of the definition above), this
170 points to the first one. */
171 struct access *first_child;
173 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
174 described above. In IPA-SRA this is a pointer to the next access
175 belonging to the same group (having the same representative). */
176 struct access *next_sibling;
178 /* Pointers to the first and last element in the linked list of assign
179 links. */
180 struct assign_link *first_link, *last_link;
182 /* Pointer to the next access in the work queue. */
183 struct access *next_queued;
185 /* Replacement variable for this access "region." Never to be accessed
186 directly, always only by the means of get_access_replacement() and only
187 when grp_to_be_replaced flag is set. */
188 tree replacement_decl;
190 /* Is this particular access write access? */
191 unsigned write : 1;
193 /* Is this access an access to a non-addressable field? */
194 unsigned non_addressable : 1;
196 /* Is this access currently in the work queue? */
197 unsigned grp_queued : 1;
199 /* Does this group contain a write access? This flag is propagated down the
200 access tree. */
201 unsigned grp_write : 1;
203 /* Does this group contain a read access? This flag is propagated down the
204 access tree. */
205 unsigned grp_read : 1;
207 /* Does this group contain a read access that comes from an assignment
208 statement? This flag is propagated down the access tree. */
209 unsigned grp_assignment_read : 1;
211 /* Does this group contain a write access that comes from an assignment
212 statement? This flag is propagated down the access tree. */
213 unsigned grp_assignment_write : 1;
215 /* Does this group contain a read access through a scalar type? This flag is
216 not propagated in the access tree in any direction. */
217 unsigned grp_scalar_read : 1;
219 /* Does this group contain a write access through a scalar type? This flag
220 is not propagated in the access tree in any direction. */
221 unsigned grp_scalar_write : 1;
223 /* Is this access an artificial one created to scalarize some record
224 entirely? */
225 unsigned grp_total_scalarization : 1;
227 /* Other passes of the analysis use this bit to make function
228 analyze_access_subtree create scalar replacements for this group if
229 possible. */
230 unsigned grp_hint : 1;
232 /* Is the subtree rooted in this access fully covered by scalar
233 replacements? */
234 unsigned grp_covered : 1;
236 /* If set to true, this access and all below it in an access tree must not be
237 scalarized. */
238 unsigned grp_unscalarizable_region : 1;
240 /* Whether data have been written to parts of the aggregate covered by this
241 access which is not to be scalarized. This flag is propagated up in the
242 access tree. */
243 unsigned grp_unscalarized_data : 1;
245 /* Does this access and/or group contain a write access through a
246 BIT_FIELD_REF? */
247 unsigned grp_partial_lhs : 1;
249 /* Set when a scalar replacement should be created for this variable. */
250 unsigned grp_to_be_replaced : 1;
252 /* Set when we want a replacement for the sole purpose of having it in
253 generated debug statements. */
254 unsigned grp_to_be_debug_replaced : 1;
256 /* Should TREE_NO_WARNING of a replacement be set? */
257 unsigned grp_no_warning : 1;
259 /* Is it possible that the group refers to data which might be (directly or
260 otherwise) modified? */
261 unsigned grp_maybe_modified : 1;
263 /* Set when this is a representative of a pointer to scalar (i.e. by
264 reference) parameter which we consider for turning into a plain scalar
265 (i.e. a by value parameter). */
266 unsigned grp_scalar_ptr : 1;
268 /* Set when we discover that this pointer is not safe to dereference in the
269 caller. */
270 unsigned grp_not_necessarilly_dereferenced : 1;
273 typedef struct access *access_p;
276 /* Alloc pool for allocating access structures. */
277 static alloc_pool access_pool;
279 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
280 are used to propagate subaccesses from rhs to lhs as long as they don't
281 conflict with what is already there. */
282 struct assign_link
284 struct access *lacc, *racc;
285 struct assign_link *next;
288 /* Alloc pool for allocating assign link structures. */
289 static alloc_pool link_pool;
291 /* Base (tree) -> Vector (vec<access_p> *) map. */
292 static struct pointer_map_t *base_access_vec;
294 /* Candidate hash table helpers. */
296 struct uid_decl_hasher : typed_noop_remove <tree_node>
298 typedef tree_node value_type;
299 typedef tree_node compare_type;
300 static inline hashval_t hash (const value_type *);
301 static inline bool equal (const value_type *, const compare_type *);
304 /* Hash a tree in a uid_decl_map. */
306 inline hashval_t
307 uid_decl_hasher::hash (const value_type *item)
309 return item->decl_minimal.uid;
312 /* Return true if the DECL_UID in both trees are equal. */
314 inline bool
315 uid_decl_hasher::equal (const value_type *a, const compare_type *b)
317 return (a->decl_minimal.uid == b->decl_minimal.uid);
320 /* Set of candidates. */
321 static bitmap candidate_bitmap;
322 static hash_table <uid_decl_hasher> candidates;
324 /* For a candidate UID return the candidates decl. */
326 static inline tree
327 candidate (unsigned uid)
329 tree_node t;
330 t.decl_minimal.uid = uid;
331 return candidates.find_with_hash (&t, static_cast <hashval_t> (uid));
334 /* Bitmap of candidates which we should try to entirely scalarize away and
335 those which cannot be (because they are and need be used as a whole). */
336 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
338 /* Obstack for creation of fancy names. */
339 static struct obstack name_obstack;
341 /* Head of a linked list of accesses that need to have its subaccesses
342 propagated to their assignment counterparts. */
343 static struct access *work_queue_head;
345 /* Number of parameters of the analyzed function when doing early ipa SRA. */
346 static int func_param_count;
348 /* scan_function sets the following to true if it encounters a call to
349 __builtin_apply_args. */
350 static bool encountered_apply_args;
352 /* Set by scan_function when it finds a recursive call. */
353 static bool encountered_recursive_call;
355 /* Set by scan_function when it finds a recursive call with less actual
356 arguments than formal parameters.. */
357 static bool encountered_unchangable_recursive_call;
359 /* This is a table in which for each basic block and parameter there is a
360 distance (offset + size) in that parameter which is dereferenced and
361 accessed in that BB. */
362 static HOST_WIDE_INT *bb_dereferences;
363 /* Bitmap of BBs that can cause the function to "stop" progressing by
364 returning, throwing externally, looping infinitely or calling a function
365 which might abort etc.. */
366 static bitmap final_bbs;
368 /* Representative of no accesses at all. */
369 static struct access no_accesses_representant;
371 /* Predicate to test the special value. */
373 static inline bool
374 no_accesses_p (struct access *access)
376 return access == &no_accesses_representant;
379 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
380 representative fields are dumped, otherwise those which only describe the
381 individual access are. */
383 static struct
385 /* Number of processed aggregates is readily available in
386 analyze_all_variable_accesses and so is not stored here. */
388 /* Number of created scalar replacements. */
389 int replacements;
391 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
392 expression. */
393 int exprs;
395 /* Number of statements created by generate_subtree_copies. */
396 int subtree_copies;
398 /* Number of statements created by load_assign_lhs_subreplacements. */
399 int subreplacements;
401 /* Number of times sra_modify_assign has deleted a statement. */
402 int deleted;
404 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
405 RHS reparately due to type conversions or nonexistent matching
406 references. */
407 int separate_lhs_rhs_handling;
409 /* Number of parameters that were removed because they were unused. */
410 int deleted_unused_parameters;
412 /* Number of scalars passed as parameters by reference that have been
413 converted to be passed by value. */
414 int scalar_by_ref_to_by_val;
416 /* Number of aggregate parameters that were replaced by one or more of their
417 components. */
418 int aggregate_params_reduced;
420 /* Numbber of components created when splitting aggregate parameters. */
421 int param_reductions_created;
422 } sra_stats;
424 static void
425 dump_access (FILE *f, struct access *access, bool grp)
427 fprintf (f, "access { ");
428 fprintf (f, "base = (%d)'", DECL_UID (access->base));
429 print_generic_expr (f, access->base, 0);
430 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
431 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
432 fprintf (f, ", expr = ");
433 print_generic_expr (f, access->expr, 0);
434 fprintf (f, ", type = ");
435 print_generic_expr (f, access->type, 0);
436 if (grp)
437 fprintf (f, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
438 "grp_assignment_write = %d, grp_scalar_read = %d, "
439 "grp_scalar_write = %d, grp_total_scalarization = %d, "
440 "grp_hint = %d, grp_covered = %d, "
441 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
442 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
443 "grp_to_be_debug_replaced = %d, grp_maybe_modified = %d, "
444 "grp_not_necessarilly_dereferenced = %d\n",
445 access->grp_read, access->grp_write, access->grp_assignment_read,
446 access->grp_assignment_write, access->grp_scalar_read,
447 access->grp_scalar_write, access->grp_total_scalarization,
448 access->grp_hint, access->grp_covered,
449 access->grp_unscalarizable_region, access->grp_unscalarized_data,
450 access->grp_partial_lhs, access->grp_to_be_replaced,
451 access->grp_to_be_debug_replaced, access->grp_maybe_modified,
452 access->grp_not_necessarilly_dereferenced);
453 else
454 fprintf (f, ", write = %d, grp_total_scalarization = %d, "
455 "grp_partial_lhs = %d\n",
456 access->write, access->grp_total_scalarization,
457 access->grp_partial_lhs);
460 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
462 static void
463 dump_access_tree_1 (FILE *f, struct access *access, int level)
467 int i;
469 for (i = 0; i < level; i++)
470 fputs ("* ", dump_file);
472 dump_access (f, access, true);
474 if (access->first_child)
475 dump_access_tree_1 (f, access->first_child, level + 1);
477 access = access->next_sibling;
479 while (access);
482 /* Dump all access trees for a variable, given the pointer to the first root in
483 ACCESS. */
485 static void
486 dump_access_tree (FILE *f, struct access *access)
488 for (; access; access = access->next_grp)
489 dump_access_tree_1 (f, access, 0);
492 /* Return true iff ACC is non-NULL and has subaccesses. */
494 static inline bool
495 access_has_children_p (struct access *acc)
497 return acc && acc->first_child;
500 /* Return true iff ACC is (partly) covered by at least one replacement. */
502 static bool
503 access_has_replacements_p (struct access *acc)
505 struct access *child;
506 if (acc->grp_to_be_replaced)
507 return true;
508 for (child = acc->first_child; child; child = child->next_sibling)
509 if (access_has_replacements_p (child))
510 return true;
511 return false;
514 /* Return a vector of pointers to accesses for the variable given in BASE or
515 NULL if there is none. */
517 static vec<access_p> *
518 get_base_access_vector (tree base)
520 void **slot;
522 slot = pointer_map_contains (base_access_vec, base);
523 if (!slot)
524 return NULL;
525 else
526 return *(vec<access_p> **) slot;
529 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
530 in ACCESS. Return NULL if it cannot be found. */
532 static struct access *
533 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
534 HOST_WIDE_INT size)
536 while (access && (access->offset != offset || access->size != size))
538 struct access *child = access->first_child;
540 while (child && (child->offset + child->size <= offset))
541 child = child->next_sibling;
542 access = child;
545 return access;
548 /* Return the first group representative for DECL or NULL if none exists. */
550 static struct access *
551 get_first_repr_for_decl (tree base)
553 vec<access_p> *access_vec;
555 access_vec = get_base_access_vector (base);
556 if (!access_vec)
557 return NULL;
559 return (*access_vec)[0];
562 /* Find an access representative for the variable BASE and given OFFSET and
563 SIZE. Requires that access trees have already been built. Return NULL if
564 it cannot be found. */
566 static struct access *
567 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
568 HOST_WIDE_INT size)
570 struct access *access;
572 access = get_first_repr_for_decl (base);
573 while (access && (access->offset + access->size <= offset))
574 access = access->next_grp;
575 if (!access)
576 return NULL;
578 return find_access_in_subtree (access, offset, size);
581 /* Add LINK to the linked list of assign links of RACC. */
582 static void
583 add_link_to_rhs (struct access *racc, struct assign_link *link)
585 gcc_assert (link->racc == racc);
587 if (!racc->first_link)
589 gcc_assert (!racc->last_link);
590 racc->first_link = link;
592 else
593 racc->last_link->next = link;
595 racc->last_link = link;
596 link->next = NULL;
599 /* Move all link structures in their linked list in OLD_RACC to the linked list
600 in NEW_RACC. */
601 static void
602 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
604 if (!old_racc->first_link)
606 gcc_assert (!old_racc->last_link);
607 return;
610 if (new_racc->first_link)
612 gcc_assert (!new_racc->last_link->next);
613 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
615 new_racc->last_link->next = old_racc->first_link;
616 new_racc->last_link = old_racc->last_link;
618 else
620 gcc_assert (!new_racc->last_link);
622 new_racc->first_link = old_racc->first_link;
623 new_racc->last_link = old_racc->last_link;
625 old_racc->first_link = old_racc->last_link = NULL;
628 /* Add ACCESS to the work queue (which is actually a stack). */
630 static void
631 add_access_to_work_queue (struct access *access)
633 if (!access->grp_queued)
635 gcc_assert (!access->next_queued);
636 access->next_queued = work_queue_head;
637 access->grp_queued = 1;
638 work_queue_head = access;
642 /* Pop an access from the work queue, and return it, assuming there is one. */
644 static struct access *
645 pop_access_from_work_queue (void)
647 struct access *access = work_queue_head;
649 work_queue_head = access->next_queued;
650 access->next_queued = NULL;
651 access->grp_queued = 0;
652 return access;
656 /* Allocate necessary structures. */
658 static void
659 sra_initialize (void)
661 candidate_bitmap = BITMAP_ALLOC (NULL);
662 candidates.create (vec_safe_length (cfun->local_decls) / 2);
663 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
664 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
665 gcc_obstack_init (&name_obstack);
666 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
667 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
668 base_access_vec = pointer_map_create ();
669 memset (&sra_stats, 0, sizeof (sra_stats));
670 encountered_apply_args = false;
671 encountered_recursive_call = false;
672 encountered_unchangable_recursive_call = false;
675 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
677 static bool
678 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
679 void *data ATTRIBUTE_UNUSED)
681 vec<access_p> *access_vec = (vec<access_p> *) *value;
682 vec_free (access_vec);
683 return true;
686 /* Deallocate all general structures. */
688 static void
689 sra_deinitialize (void)
691 BITMAP_FREE (candidate_bitmap);
692 candidates.dispose ();
693 BITMAP_FREE (should_scalarize_away_bitmap);
694 BITMAP_FREE (cannot_scalarize_away_bitmap);
695 free_alloc_pool (access_pool);
696 free_alloc_pool (link_pool);
697 obstack_free (&name_obstack, NULL);
699 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
700 pointer_map_destroy (base_access_vec);
703 /* Remove DECL from candidates for SRA and write REASON to the dump file if
704 there is one. */
705 static void
706 disqualify_candidate (tree decl, const char *reason)
708 if (bitmap_clear_bit (candidate_bitmap, DECL_UID (decl)))
709 candidates.clear_slot (candidates.find_slot_with_hash (decl,
710 DECL_UID (decl),
711 NO_INSERT));
713 if (dump_file && (dump_flags & TDF_DETAILS))
715 fprintf (dump_file, "! Disqualifying ");
716 print_generic_expr (dump_file, decl, 0);
717 fprintf (dump_file, " - %s\n", reason);
721 /* Return true iff the type contains a field or an element which does not allow
722 scalarization. */
724 static bool
725 type_internals_preclude_sra_p (tree type, const char **msg)
727 tree fld;
728 tree et;
730 switch (TREE_CODE (type))
732 case RECORD_TYPE:
733 case UNION_TYPE:
734 case QUAL_UNION_TYPE:
735 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
736 if (TREE_CODE (fld) == FIELD_DECL)
738 tree ft = TREE_TYPE (fld);
740 if (TREE_THIS_VOLATILE (fld))
742 *msg = "volatile structure field";
743 return true;
745 if (!DECL_FIELD_OFFSET (fld))
747 *msg = "no structure field offset";
748 return true;
750 if (!DECL_SIZE (fld))
752 *msg = "zero structure field size";
753 return true;
755 if (!tree_fits_uhwi_p (DECL_FIELD_OFFSET (fld)))
757 *msg = "structure field offset not fixed";
758 return true;
760 if (!tree_fits_uhwi_p (DECL_SIZE (fld)))
762 *msg = "structure field size not fixed";
763 return true;
765 if (!tree_fits_shwi_p (bit_position (fld)))
767 *msg = "structure field size too big";
768 return true;
770 if (AGGREGATE_TYPE_P (ft)
771 && int_bit_position (fld) % BITS_PER_UNIT != 0)
773 *msg = "structure field is bit field";
774 return true;
777 if (AGGREGATE_TYPE_P (ft) && type_internals_preclude_sra_p (ft, msg))
778 return true;
781 return false;
783 case ARRAY_TYPE:
784 et = TREE_TYPE (type);
786 if (TYPE_VOLATILE (et))
788 *msg = "element type is volatile";
789 return true;
792 if (AGGREGATE_TYPE_P (et) && type_internals_preclude_sra_p (et, msg))
793 return true;
795 return false;
797 default:
798 return false;
802 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
803 base variable if it is. Return T if it is not an SSA_NAME. */
805 static tree
806 get_ssa_base_param (tree t)
808 if (TREE_CODE (t) == SSA_NAME)
810 if (SSA_NAME_IS_DEFAULT_DEF (t))
811 return SSA_NAME_VAR (t);
812 else
813 return NULL_TREE;
815 return t;
818 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
819 belongs to, unless the BB has already been marked as a potentially
820 final. */
822 static void
823 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
825 basic_block bb = gimple_bb (stmt);
826 int idx, parm_index = 0;
827 tree parm;
829 if (bitmap_bit_p (final_bbs, bb->index))
830 return;
832 for (parm = DECL_ARGUMENTS (current_function_decl);
833 parm && parm != base;
834 parm = DECL_CHAIN (parm))
835 parm_index++;
837 gcc_assert (parm_index < func_param_count);
839 idx = bb->index * func_param_count + parm_index;
840 if (bb_dereferences[idx] < dist)
841 bb_dereferences[idx] = dist;
844 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
845 the three fields. Also add it to the vector of accesses corresponding to
846 the base. Finally, return the new access. */
848 static struct access *
849 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
851 vec<access_p> *v;
852 struct access *access;
853 void **slot;
855 access = (struct access *) pool_alloc (access_pool);
856 memset (access, 0, sizeof (struct access));
857 access->base = base;
858 access->offset = offset;
859 access->size = size;
861 slot = pointer_map_contains (base_access_vec, base);
862 if (slot)
863 v = (vec<access_p> *) *slot;
864 else
865 vec_alloc (v, 32);
867 v->safe_push (access);
869 *((vec<access_p> **)
870 pointer_map_insert (base_access_vec, base)) = v;
872 return access;
875 /* Create and insert access for EXPR. Return created access, or NULL if it is
876 not possible. */
878 static struct access *
879 create_access (tree expr, gimple stmt, bool write)
881 struct access *access;
882 HOST_WIDE_INT offset, size, max_size;
883 tree base = expr;
884 bool ptr, unscalarizable_region = false;
886 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
888 if (sra_mode == SRA_MODE_EARLY_IPA
889 && TREE_CODE (base) == MEM_REF)
891 base = get_ssa_base_param (TREE_OPERAND (base, 0));
892 if (!base)
893 return NULL;
894 ptr = true;
896 else
897 ptr = false;
899 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
900 return NULL;
902 if (sra_mode == SRA_MODE_EARLY_IPA)
904 if (size < 0 || size != max_size)
906 disqualify_candidate (base, "Encountered a variable sized access.");
907 return NULL;
909 if (TREE_CODE (expr) == COMPONENT_REF
910 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
912 disqualify_candidate (base, "Encountered a bit-field access.");
913 return NULL;
915 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
917 if (ptr)
918 mark_parm_dereference (base, offset + size, stmt);
920 else
922 if (size != max_size)
924 size = max_size;
925 unscalarizable_region = true;
927 if (size < 0)
929 disqualify_candidate (base, "Encountered an unconstrained access.");
930 return NULL;
934 access = create_access_1 (base, offset, size);
935 access->expr = expr;
936 access->type = TREE_TYPE (expr);
937 access->write = write;
938 access->grp_unscalarizable_region = unscalarizable_region;
939 access->stmt = stmt;
941 if (TREE_CODE (expr) == COMPONENT_REF
942 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
943 access->non_addressable = 1;
945 return access;
949 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
950 register types or (recursively) records with only these two kinds of fields.
951 It also returns false if any of these records contains a bit-field. */
953 static bool
954 type_consists_of_records_p (tree type)
956 tree fld;
958 if (TREE_CODE (type) != RECORD_TYPE)
959 return false;
961 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
962 if (TREE_CODE (fld) == FIELD_DECL)
964 tree ft = TREE_TYPE (fld);
966 if (DECL_BIT_FIELD (fld))
967 return false;
969 if (!is_gimple_reg_type (ft)
970 && !type_consists_of_records_p (ft))
971 return false;
974 return true;
977 /* Create total_scalarization accesses for all scalar type fields in DECL that
978 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
979 must be the top-most VAR_DECL representing the variable, OFFSET must be the
980 offset of DECL within BASE. REF must be the memory reference expression for
981 the given decl. */
983 static void
984 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
985 tree ref)
987 tree fld, decl_type = TREE_TYPE (decl);
989 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
990 if (TREE_CODE (fld) == FIELD_DECL)
992 HOST_WIDE_INT pos = offset + int_bit_position (fld);
993 tree ft = TREE_TYPE (fld);
994 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
995 NULL_TREE);
997 if (is_gimple_reg_type (ft))
999 struct access *access;
1000 HOST_WIDE_INT size;
1002 size = tree_to_uhwi (DECL_SIZE (fld));
1003 access = create_access_1 (base, pos, size);
1004 access->expr = nref;
1005 access->type = ft;
1006 access->grp_total_scalarization = 1;
1007 /* Accesses for intraprocedural SRA can have their stmt NULL. */
1009 else
1010 completely_scalarize_record (base, fld, pos, nref);
1014 /* Create total_scalarization accesses for all scalar type fields in VAR and
1015 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
1016 type_consists_of_records_p. */
1018 static void
1019 completely_scalarize_var (tree var)
1021 HOST_WIDE_INT size = tree_to_uhwi (DECL_SIZE (var));
1022 struct access *access;
1024 access = create_access_1 (var, 0, size);
1025 access->expr = var;
1026 access->type = TREE_TYPE (var);
1027 access->grp_total_scalarization = 1;
1029 completely_scalarize_record (var, var, 0, var);
1032 /* Return true if REF has an VIEW_CONVERT_EXPR somewhere in it. */
1034 static inline bool
1035 contains_view_convert_expr_p (const_tree ref)
1037 while (handled_component_p (ref))
1039 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR)
1040 return true;
1041 ref = TREE_OPERAND (ref, 0);
1044 return false;
1047 /* Search the given tree for a declaration by skipping handled components and
1048 exclude it from the candidates. */
1050 static void
1051 disqualify_base_of_expr (tree t, const char *reason)
1053 t = get_base_address (t);
1054 if (sra_mode == SRA_MODE_EARLY_IPA
1055 && TREE_CODE (t) == MEM_REF)
1056 t = get_ssa_base_param (TREE_OPERAND (t, 0));
1058 if (t && DECL_P (t))
1059 disqualify_candidate (t, reason);
1062 /* Scan expression EXPR and create access structures for all accesses to
1063 candidates for scalarization. Return the created access or NULL if none is
1064 created. */
1066 static struct access *
1067 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
1069 struct access *ret = NULL;
1070 bool partial_ref;
1072 if (TREE_CODE (expr) == BIT_FIELD_REF
1073 || TREE_CODE (expr) == IMAGPART_EXPR
1074 || TREE_CODE (expr) == REALPART_EXPR)
1076 expr = TREE_OPERAND (expr, 0);
1077 partial_ref = true;
1079 else
1080 partial_ref = false;
1082 /* We need to dive through V_C_Es in order to get the size of its parameter
1083 and not the result type. Ada produces such statements. We are also
1084 capable of handling the topmost V_C_E but not any of those buried in other
1085 handled components. */
1086 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
1087 expr = TREE_OPERAND (expr, 0);
1089 if (contains_view_convert_expr_p (expr))
1091 disqualify_base_of_expr (expr, "V_C_E under a different handled "
1092 "component.");
1093 return NULL;
1096 switch (TREE_CODE (expr))
1098 case MEM_REF:
1099 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
1100 && sra_mode != SRA_MODE_EARLY_IPA)
1101 return NULL;
1102 /* fall through */
1103 case VAR_DECL:
1104 case PARM_DECL:
1105 case RESULT_DECL:
1106 case COMPONENT_REF:
1107 case ARRAY_REF:
1108 case ARRAY_RANGE_REF:
1109 ret = create_access (expr, stmt, write);
1110 break;
1112 default:
1113 break;
1116 if (write && partial_ref && ret)
1117 ret->grp_partial_lhs = 1;
1119 return ret;
1122 /* Scan expression EXPR and create access structures for all accesses to
1123 candidates for scalarization. Return true if any access has been inserted.
1124 STMT must be the statement from which the expression is taken, WRITE must be
1125 true if the expression is a store and false otherwise. */
1127 static bool
1128 build_access_from_expr (tree expr, gimple stmt, bool write)
1130 struct access *access;
1132 access = build_access_from_expr_1 (expr, stmt, write);
1133 if (access)
1135 /* This means the aggregate is accesses as a whole in a way other than an
1136 assign statement and thus cannot be removed even if we had a scalar
1137 replacement for everything. */
1138 if (cannot_scalarize_away_bitmap)
1139 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1140 return true;
1142 return false;
1145 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1146 modes in which it matters, return true iff they have been disqualified. RHS
1147 may be NULL, in that case ignore it. If we scalarize an aggregate in
1148 intra-SRA we may need to add statements after each statement. This is not
1149 possible if a statement unconditionally has to end the basic block. */
1150 static bool
1151 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1153 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1154 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1156 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1157 if (rhs)
1158 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1159 return true;
1161 return false;
1164 /* Scan expressions occurring in STMT, create access structures for all accesses
1165 to candidates for scalarization and remove those candidates which occur in
1166 statements or expressions that prevent them from being split apart. Return
1167 true if any access has been inserted. */
1169 static bool
1170 build_accesses_from_assign (gimple stmt)
1172 tree lhs, rhs;
1173 struct access *lacc, *racc;
1175 if (!gimple_assign_single_p (stmt)
1176 /* Scope clobbers don't influence scalarization. */
1177 || gimple_clobber_p (stmt))
1178 return false;
1180 lhs = gimple_assign_lhs (stmt);
1181 rhs = gimple_assign_rhs1 (stmt);
1183 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1184 return false;
1186 racc = build_access_from_expr_1 (rhs, stmt, false);
1187 lacc = build_access_from_expr_1 (lhs, stmt, true);
1189 if (lacc)
1190 lacc->grp_assignment_write = 1;
1192 if (racc)
1194 racc->grp_assignment_read = 1;
1195 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1196 && !is_gimple_reg_type (racc->type))
1197 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1200 if (lacc && racc
1201 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1202 && !lacc->grp_unscalarizable_region
1203 && !racc->grp_unscalarizable_region
1204 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1205 && lacc->size == racc->size
1206 && useless_type_conversion_p (lacc->type, racc->type))
1208 struct assign_link *link;
1210 link = (struct assign_link *) pool_alloc (link_pool);
1211 memset (link, 0, sizeof (struct assign_link));
1213 link->lacc = lacc;
1214 link->racc = racc;
1216 add_link_to_rhs (racc, link);
1219 return lacc || racc;
1222 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1223 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1225 static bool
1226 asm_visit_addr (gimple, tree op, tree, void *)
1228 op = get_base_address (op);
1229 if (op
1230 && DECL_P (op))
1231 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1233 return false;
1236 /* Return true iff callsite CALL has at least as many actual arguments as there
1237 are formal parameters of the function currently processed by IPA-SRA. */
1239 static inline bool
1240 callsite_has_enough_arguments_p (gimple call)
1242 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1245 /* Scan function and look for interesting expressions and create access
1246 structures for them. Return true iff any access is created. */
1248 static bool
1249 scan_function (void)
1251 basic_block bb;
1252 bool ret = false;
1254 FOR_EACH_BB_FN (bb, cfun)
1256 gimple_stmt_iterator gsi;
1257 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1259 gimple stmt = gsi_stmt (gsi);
1260 tree t;
1261 unsigned i;
1263 if (final_bbs && stmt_can_throw_external (stmt))
1264 bitmap_set_bit (final_bbs, bb->index);
1265 switch (gimple_code (stmt))
1267 case GIMPLE_RETURN:
1268 t = gimple_return_retval (stmt);
1269 if (t != NULL_TREE)
1270 ret |= build_access_from_expr (t, stmt, false);
1271 if (final_bbs)
1272 bitmap_set_bit (final_bbs, bb->index);
1273 break;
1275 case GIMPLE_ASSIGN:
1276 ret |= build_accesses_from_assign (stmt);
1277 break;
1279 case GIMPLE_CALL:
1280 for (i = 0; i < gimple_call_num_args (stmt); i++)
1281 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1282 stmt, false);
1284 if (sra_mode == SRA_MODE_EARLY_IPA)
1286 tree dest = gimple_call_fndecl (stmt);
1287 int flags = gimple_call_flags (stmt);
1289 if (dest)
1291 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1292 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1293 encountered_apply_args = true;
1294 if (recursive_call_p (current_function_decl, dest))
1296 encountered_recursive_call = true;
1297 if (!callsite_has_enough_arguments_p (stmt))
1298 encountered_unchangable_recursive_call = true;
1302 if (final_bbs
1303 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1304 bitmap_set_bit (final_bbs, bb->index);
1307 t = gimple_call_lhs (stmt);
1308 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1309 ret |= build_access_from_expr (t, stmt, true);
1310 break;
1312 case GIMPLE_ASM:
1313 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1314 asm_visit_addr);
1315 if (final_bbs)
1316 bitmap_set_bit (final_bbs, bb->index);
1318 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1320 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1321 ret |= build_access_from_expr (t, stmt, false);
1323 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1325 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1326 ret |= build_access_from_expr (t, stmt, true);
1328 break;
1330 default:
1331 break;
1336 return ret;
1339 /* Helper of QSORT function. There are pointers to accesses in the array. An
1340 access is considered smaller than another if it has smaller offset or if the
1341 offsets are the same but is size is bigger. */
1343 static int
1344 compare_access_positions (const void *a, const void *b)
1346 const access_p *fp1 = (const access_p *) a;
1347 const access_p *fp2 = (const access_p *) b;
1348 const access_p f1 = *fp1;
1349 const access_p f2 = *fp2;
1351 if (f1->offset != f2->offset)
1352 return f1->offset < f2->offset ? -1 : 1;
1354 if (f1->size == f2->size)
1356 if (f1->type == f2->type)
1357 return 0;
1358 /* Put any non-aggregate type before any aggregate type. */
1359 else if (!is_gimple_reg_type (f1->type)
1360 && is_gimple_reg_type (f2->type))
1361 return 1;
1362 else if (is_gimple_reg_type (f1->type)
1363 && !is_gimple_reg_type (f2->type))
1364 return -1;
1365 /* Put any complex or vector type before any other scalar type. */
1366 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1367 && TREE_CODE (f1->type) != VECTOR_TYPE
1368 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1369 || TREE_CODE (f2->type) == VECTOR_TYPE))
1370 return 1;
1371 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1372 || TREE_CODE (f1->type) == VECTOR_TYPE)
1373 && TREE_CODE (f2->type) != COMPLEX_TYPE
1374 && TREE_CODE (f2->type) != VECTOR_TYPE)
1375 return -1;
1376 /* Put the integral type with the bigger precision first. */
1377 else if (INTEGRAL_TYPE_P (f1->type)
1378 && INTEGRAL_TYPE_P (f2->type))
1379 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1380 /* Put any integral type with non-full precision last. */
1381 else if (INTEGRAL_TYPE_P (f1->type)
1382 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1383 != TYPE_PRECISION (f1->type)))
1384 return 1;
1385 else if (INTEGRAL_TYPE_P (f2->type)
1386 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1387 != TYPE_PRECISION (f2->type)))
1388 return -1;
1389 /* Stabilize the sort. */
1390 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1393 /* We want the bigger accesses first, thus the opposite operator in the next
1394 line: */
1395 return f1->size > f2->size ? -1 : 1;
1399 /* Append a name of the declaration to the name obstack. A helper function for
1400 make_fancy_name. */
1402 static void
1403 make_fancy_decl_name (tree decl)
1405 char buffer[32];
1407 tree name = DECL_NAME (decl);
1408 if (name)
1409 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1410 IDENTIFIER_LENGTH (name));
1411 else
1413 sprintf (buffer, "D%u", DECL_UID (decl));
1414 obstack_grow (&name_obstack, buffer, strlen (buffer));
1418 /* Helper for make_fancy_name. */
1420 static void
1421 make_fancy_name_1 (tree expr)
1423 char buffer[32];
1424 tree index;
1426 if (DECL_P (expr))
1428 make_fancy_decl_name (expr);
1429 return;
1432 switch (TREE_CODE (expr))
1434 case COMPONENT_REF:
1435 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1436 obstack_1grow (&name_obstack, '$');
1437 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1438 break;
1440 case ARRAY_REF:
1441 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1442 obstack_1grow (&name_obstack, '$');
1443 /* Arrays with only one element may not have a constant as their
1444 index. */
1445 index = TREE_OPERAND (expr, 1);
1446 if (TREE_CODE (index) != INTEGER_CST)
1447 break;
1448 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1449 obstack_grow (&name_obstack, buffer, strlen (buffer));
1450 break;
1452 case ADDR_EXPR:
1453 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1454 break;
1456 case MEM_REF:
1457 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1458 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1460 obstack_1grow (&name_obstack, '$');
1461 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1462 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1463 obstack_grow (&name_obstack, buffer, strlen (buffer));
1465 break;
1467 case BIT_FIELD_REF:
1468 case REALPART_EXPR:
1469 case IMAGPART_EXPR:
1470 gcc_unreachable (); /* we treat these as scalars. */
1471 break;
1472 default:
1473 break;
1477 /* Create a human readable name for replacement variable of ACCESS. */
1479 static char *
1480 make_fancy_name (tree expr)
1482 make_fancy_name_1 (expr);
1483 obstack_1grow (&name_obstack, '\0');
1484 return XOBFINISH (&name_obstack, char *);
1487 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1488 EXP_TYPE at the given OFFSET. If BASE is something for which
1489 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1490 to insert new statements either before or below the current one as specified
1491 by INSERT_AFTER. This function is not capable of handling bitfields.
1493 BASE must be either a declaration or a memory reference that has correct
1494 alignment ifformation embeded in it (e.g. a pre-existing one in SRA). */
1496 tree
1497 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1498 tree exp_type, gimple_stmt_iterator *gsi,
1499 bool insert_after)
1501 tree prev_base = base;
1502 tree off;
1503 tree mem_ref;
1504 HOST_WIDE_INT base_offset;
1505 unsigned HOST_WIDE_INT misalign;
1506 unsigned int align;
1508 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1509 get_object_alignment_1 (base, &align, &misalign);
1510 base = get_addr_base_and_unit_offset (base, &base_offset);
1512 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1513 offset such as array[var_index]. */
1514 if (!base)
1516 gimple stmt;
1517 tree tmp, addr;
1519 gcc_checking_assert (gsi);
1520 tmp = make_ssa_name (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1521 addr = build_fold_addr_expr (unshare_expr (prev_base));
1522 STRIP_USELESS_TYPE_CONVERSION (addr);
1523 stmt = gimple_build_assign (tmp, addr);
1524 gimple_set_location (stmt, loc);
1525 if (insert_after)
1526 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1527 else
1528 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1530 off = build_int_cst (reference_alias_ptr_type (prev_base),
1531 offset / BITS_PER_UNIT);
1532 base = tmp;
1534 else if (TREE_CODE (base) == MEM_REF)
1536 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1537 base_offset + offset / BITS_PER_UNIT);
1538 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1539 base = unshare_expr (TREE_OPERAND (base, 0));
1541 else
1543 off = build_int_cst (reference_alias_ptr_type (base),
1544 base_offset + offset / BITS_PER_UNIT);
1545 base = build_fold_addr_expr (unshare_expr (base));
1548 misalign = (misalign + offset) & (align - 1);
1549 if (misalign != 0)
1550 align = (misalign & -misalign);
1551 if (align < TYPE_ALIGN (exp_type))
1552 exp_type = build_aligned_type (exp_type, align);
1554 mem_ref = fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1555 if (TREE_THIS_VOLATILE (prev_base))
1556 TREE_THIS_VOLATILE (mem_ref) = 1;
1557 if (TREE_SIDE_EFFECTS (prev_base))
1558 TREE_SIDE_EFFECTS (mem_ref) = 1;
1559 return mem_ref;
1562 /* Construct a memory reference to a part of an aggregate BASE at the given
1563 OFFSET and of the same type as MODEL. In case this is a reference to a
1564 bit-field, the function will replicate the last component_ref of model's
1565 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1566 build_ref_for_offset. */
1568 static tree
1569 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1570 struct access *model, gimple_stmt_iterator *gsi,
1571 bool insert_after)
1573 if (TREE_CODE (model->expr) == COMPONENT_REF
1574 && DECL_BIT_FIELD (TREE_OPERAND (model->expr, 1)))
1576 /* This access represents a bit-field. */
1577 tree t, exp_type, fld = TREE_OPERAND (model->expr, 1);
1579 offset -= int_bit_position (fld);
1580 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1581 t = build_ref_for_offset (loc, base, offset, exp_type, gsi, insert_after);
1582 return fold_build3_loc (loc, COMPONENT_REF, TREE_TYPE (fld), t, fld,
1583 NULL_TREE);
1585 else
1586 return build_ref_for_offset (loc, base, offset, model->type,
1587 gsi, insert_after);
1590 /* Attempt to build a memory reference that we could but into a gimple
1591 debug_bind statement. Similar to build_ref_for_model but punts if it has to
1592 create statements and return s NULL instead. This function also ignores
1593 alignment issues and so its results should never end up in non-debug
1594 statements. */
1596 static tree
1597 build_debug_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1598 struct access *model)
1600 HOST_WIDE_INT base_offset;
1601 tree off;
1603 if (TREE_CODE (model->expr) == COMPONENT_REF
1604 && DECL_BIT_FIELD (TREE_OPERAND (model->expr, 1)))
1605 return NULL_TREE;
1607 base = get_addr_base_and_unit_offset (base, &base_offset);
1608 if (!base)
1609 return NULL_TREE;
1610 if (TREE_CODE (base) == MEM_REF)
1612 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1613 base_offset + offset / BITS_PER_UNIT);
1614 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1615 base = unshare_expr (TREE_OPERAND (base, 0));
1617 else
1619 off = build_int_cst (reference_alias_ptr_type (base),
1620 base_offset + offset / BITS_PER_UNIT);
1621 base = build_fold_addr_expr (unshare_expr (base));
1624 return fold_build2_loc (loc, MEM_REF, model->type, base, off);
1627 /* Construct a memory reference consisting of component_refs and array_refs to
1628 a part of an aggregate *RES (which is of type TYPE). The requested part
1629 should have type EXP_TYPE at be the given OFFSET. This function might not
1630 succeed, it returns true when it does and only then *RES points to something
1631 meaningful. This function should be used only to build expressions that we
1632 might need to present to user (e.g. in warnings). In all other situations,
1633 build_ref_for_model or build_ref_for_offset should be used instead. */
1635 static bool
1636 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1637 tree exp_type)
1639 while (1)
1641 tree fld;
1642 tree tr_size, index, minidx;
1643 HOST_WIDE_INT el_size;
1645 if (offset == 0 && exp_type
1646 && types_compatible_p (exp_type, type))
1647 return true;
1649 switch (TREE_CODE (type))
1651 case UNION_TYPE:
1652 case QUAL_UNION_TYPE:
1653 case RECORD_TYPE:
1654 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1656 HOST_WIDE_INT pos, size;
1657 tree tr_pos, expr, *expr_ptr;
1659 if (TREE_CODE (fld) != FIELD_DECL)
1660 continue;
1662 tr_pos = bit_position (fld);
1663 if (!tr_pos || !tree_fits_uhwi_p (tr_pos))
1664 continue;
1665 pos = tree_to_uhwi (tr_pos);
1666 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1667 tr_size = DECL_SIZE (fld);
1668 if (!tr_size || !tree_fits_uhwi_p (tr_size))
1669 continue;
1670 size = tree_to_uhwi (tr_size);
1671 if (size == 0)
1673 if (pos != offset)
1674 continue;
1676 else if (pos > offset || (pos + size) <= offset)
1677 continue;
1679 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1680 NULL_TREE);
1681 expr_ptr = &expr;
1682 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1683 offset - pos, exp_type))
1685 *res = expr;
1686 return true;
1689 return false;
1691 case ARRAY_TYPE:
1692 tr_size = TYPE_SIZE (TREE_TYPE (type));
1693 if (!tr_size || !tree_fits_uhwi_p (tr_size))
1694 return false;
1695 el_size = tree_to_uhwi (tr_size);
1697 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1698 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1699 return false;
1700 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1701 if (!integer_zerop (minidx))
1702 index = int_const_binop (PLUS_EXPR, index, minidx);
1703 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1704 NULL_TREE, NULL_TREE);
1705 offset = offset % el_size;
1706 type = TREE_TYPE (type);
1707 break;
1709 default:
1710 if (offset != 0)
1711 return false;
1713 if (exp_type)
1714 return false;
1715 else
1716 return true;
1721 /* Return true iff TYPE is stdarg va_list type. */
1723 static inline bool
1724 is_va_list_type (tree type)
1726 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1729 /* Print message to dump file why a variable was rejected. */
1731 static void
1732 reject (tree var, const char *msg)
1734 if (dump_file && (dump_flags & TDF_DETAILS))
1736 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1737 print_generic_expr (dump_file, var, 0);
1738 fprintf (dump_file, "\n");
1742 /* Return true if VAR is a candidate for SRA. */
1744 static bool
1745 maybe_add_sra_candidate (tree var)
1747 tree type = TREE_TYPE (var);
1748 const char *msg;
1749 tree_node **slot;
1751 if (!AGGREGATE_TYPE_P (type))
1753 reject (var, "not aggregate");
1754 return false;
1756 if (needs_to_live_in_memory (var))
1758 reject (var, "needs to live in memory");
1759 return false;
1761 if (TREE_THIS_VOLATILE (var))
1763 reject (var, "is volatile");
1764 return false;
1766 if (!COMPLETE_TYPE_P (type))
1768 reject (var, "has incomplete type");
1769 return false;
1771 if (!tree_fits_uhwi_p (TYPE_SIZE (type)))
1773 reject (var, "type size not fixed");
1774 return false;
1776 if (tree_to_uhwi (TYPE_SIZE (type)) == 0)
1778 reject (var, "type size is zero");
1779 return false;
1781 if (type_internals_preclude_sra_p (type, &msg))
1783 reject (var, msg);
1784 return false;
1786 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1787 we also want to schedule it rather late. Thus we ignore it in
1788 the early pass. */
1789 (sra_mode == SRA_MODE_EARLY_INTRA
1790 && is_va_list_type (type)))
1792 reject (var, "is va_list");
1793 return false;
1796 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1797 slot = candidates.find_slot_with_hash (var, DECL_UID (var), INSERT);
1798 *slot = var;
1800 if (dump_file && (dump_flags & TDF_DETAILS))
1802 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1803 print_generic_expr (dump_file, var, 0);
1804 fprintf (dump_file, "\n");
1807 return true;
1810 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1811 those with type which is suitable for scalarization. */
1813 static bool
1814 find_var_candidates (void)
1816 tree var, parm;
1817 unsigned int i;
1818 bool ret = false;
1820 for (parm = DECL_ARGUMENTS (current_function_decl);
1821 parm;
1822 parm = DECL_CHAIN (parm))
1823 ret |= maybe_add_sra_candidate (parm);
1825 FOR_EACH_LOCAL_DECL (cfun, i, var)
1827 if (TREE_CODE (var) != VAR_DECL)
1828 continue;
1830 ret |= maybe_add_sra_candidate (var);
1833 return ret;
1836 /* Sort all accesses for the given variable, check for partial overlaps and
1837 return NULL if there are any. If there are none, pick a representative for
1838 each combination of offset and size and create a linked list out of them.
1839 Return the pointer to the first representative and make sure it is the first
1840 one in the vector of accesses. */
1842 static struct access *
1843 sort_and_splice_var_accesses (tree var)
1845 int i, j, access_count;
1846 struct access *res, **prev_acc_ptr = &res;
1847 vec<access_p> *access_vec;
1848 bool first = true;
1849 HOST_WIDE_INT low = -1, high = 0;
1851 access_vec = get_base_access_vector (var);
1852 if (!access_vec)
1853 return NULL;
1854 access_count = access_vec->length ();
1856 /* Sort by <OFFSET, SIZE>. */
1857 access_vec->qsort (compare_access_positions);
1859 i = 0;
1860 while (i < access_count)
1862 struct access *access = (*access_vec)[i];
1863 bool grp_write = access->write;
1864 bool grp_read = !access->write;
1865 bool grp_scalar_write = access->write
1866 && is_gimple_reg_type (access->type);
1867 bool grp_scalar_read = !access->write
1868 && is_gimple_reg_type (access->type);
1869 bool grp_assignment_read = access->grp_assignment_read;
1870 bool grp_assignment_write = access->grp_assignment_write;
1871 bool multiple_scalar_reads = false;
1872 bool total_scalarization = access->grp_total_scalarization;
1873 bool grp_partial_lhs = access->grp_partial_lhs;
1874 bool first_scalar = is_gimple_reg_type (access->type);
1875 bool unscalarizable_region = access->grp_unscalarizable_region;
1877 if (first || access->offset >= high)
1879 first = false;
1880 low = access->offset;
1881 high = access->offset + access->size;
1883 else if (access->offset > low && access->offset + access->size > high)
1884 return NULL;
1885 else
1886 gcc_assert (access->offset >= low
1887 && access->offset + access->size <= high);
1889 j = i + 1;
1890 while (j < access_count)
1892 struct access *ac2 = (*access_vec)[j];
1893 if (ac2->offset != access->offset || ac2->size != access->size)
1894 break;
1895 if (ac2->write)
1897 grp_write = true;
1898 grp_scalar_write = (grp_scalar_write
1899 || is_gimple_reg_type (ac2->type));
1901 else
1903 grp_read = true;
1904 if (is_gimple_reg_type (ac2->type))
1906 if (grp_scalar_read)
1907 multiple_scalar_reads = true;
1908 else
1909 grp_scalar_read = true;
1912 grp_assignment_read |= ac2->grp_assignment_read;
1913 grp_assignment_write |= ac2->grp_assignment_write;
1914 grp_partial_lhs |= ac2->grp_partial_lhs;
1915 unscalarizable_region |= ac2->grp_unscalarizable_region;
1916 total_scalarization |= ac2->grp_total_scalarization;
1917 relink_to_new_repr (access, ac2);
1919 /* If there are both aggregate-type and scalar-type accesses with
1920 this combination of size and offset, the comparison function
1921 should have put the scalars first. */
1922 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1923 ac2->group_representative = access;
1924 j++;
1927 i = j;
1929 access->group_representative = access;
1930 access->grp_write = grp_write;
1931 access->grp_read = grp_read;
1932 access->grp_scalar_read = grp_scalar_read;
1933 access->grp_scalar_write = grp_scalar_write;
1934 access->grp_assignment_read = grp_assignment_read;
1935 access->grp_assignment_write = grp_assignment_write;
1936 access->grp_hint = multiple_scalar_reads || total_scalarization;
1937 access->grp_total_scalarization = total_scalarization;
1938 access->grp_partial_lhs = grp_partial_lhs;
1939 access->grp_unscalarizable_region = unscalarizable_region;
1940 if (access->first_link)
1941 add_access_to_work_queue (access);
1943 *prev_acc_ptr = access;
1944 prev_acc_ptr = &access->next_grp;
1947 gcc_assert (res == (*access_vec)[0]);
1948 return res;
1951 /* Create a variable for the given ACCESS which determines the type, name and a
1952 few other properties. Return the variable declaration and store it also to
1953 ACCESS->replacement. */
1955 static tree
1956 create_access_replacement (struct access *access)
1958 tree repl;
1960 if (access->grp_to_be_debug_replaced)
1962 repl = create_tmp_var_raw (access->type, NULL);
1963 DECL_CONTEXT (repl) = current_function_decl;
1965 else
1966 repl = create_tmp_var (access->type, "SR");
1967 if (TREE_CODE (access->type) == COMPLEX_TYPE
1968 || TREE_CODE (access->type) == VECTOR_TYPE)
1970 if (!access->grp_partial_lhs)
1971 DECL_GIMPLE_REG_P (repl) = 1;
1973 else if (access->grp_partial_lhs
1974 && is_gimple_reg_type (access->type))
1975 TREE_ADDRESSABLE (repl) = 1;
1977 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1978 DECL_ARTIFICIAL (repl) = 1;
1979 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1981 if (DECL_NAME (access->base)
1982 && !DECL_IGNORED_P (access->base)
1983 && !DECL_ARTIFICIAL (access->base))
1985 char *pretty_name = make_fancy_name (access->expr);
1986 tree debug_expr = unshare_expr_without_location (access->expr), d;
1987 bool fail = false;
1989 DECL_NAME (repl) = get_identifier (pretty_name);
1990 obstack_free (&name_obstack, pretty_name);
1992 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1993 as DECL_DEBUG_EXPR isn't considered when looking for still
1994 used SSA_NAMEs and thus they could be freed. All debug info
1995 generation cares is whether something is constant or variable
1996 and that get_ref_base_and_extent works properly on the
1997 expression. It cannot handle accesses at a non-constant offset
1998 though, so just give up in those cases. */
1999 for (d = debug_expr;
2000 !fail && (handled_component_p (d) || TREE_CODE (d) == MEM_REF);
2001 d = TREE_OPERAND (d, 0))
2002 switch (TREE_CODE (d))
2004 case ARRAY_REF:
2005 case ARRAY_RANGE_REF:
2006 if (TREE_OPERAND (d, 1)
2007 && TREE_CODE (TREE_OPERAND (d, 1)) != INTEGER_CST)
2008 fail = true;
2009 if (TREE_OPERAND (d, 3)
2010 && TREE_CODE (TREE_OPERAND (d, 3)) != INTEGER_CST)
2011 fail = true;
2012 /* FALLTHRU */
2013 case COMPONENT_REF:
2014 if (TREE_OPERAND (d, 2)
2015 && TREE_CODE (TREE_OPERAND (d, 2)) != INTEGER_CST)
2016 fail = true;
2017 break;
2018 case MEM_REF:
2019 if (TREE_CODE (TREE_OPERAND (d, 0)) != ADDR_EXPR)
2020 fail = true;
2021 else
2022 d = TREE_OPERAND (d, 0);
2023 break;
2024 default:
2025 break;
2027 if (!fail)
2029 SET_DECL_DEBUG_EXPR (repl, debug_expr);
2030 DECL_HAS_DEBUG_EXPR_P (repl) = 1;
2032 if (access->grp_no_warning)
2033 TREE_NO_WARNING (repl) = 1;
2034 else
2035 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
2037 else
2038 TREE_NO_WARNING (repl) = 1;
2040 if (dump_file)
2042 if (access->grp_to_be_debug_replaced)
2044 fprintf (dump_file, "Created a debug-only replacement for ");
2045 print_generic_expr (dump_file, access->base, 0);
2046 fprintf (dump_file, " offset: %u, size: %u\n",
2047 (unsigned) access->offset, (unsigned) access->size);
2049 else
2051 fprintf (dump_file, "Created a replacement for ");
2052 print_generic_expr (dump_file, access->base, 0);
2053 fprintf (dump_file, " offset: %u, size: %u: ",
2054 (unsigned) access->offset, (unsigned) access->size);
2055 print_generic_expr (dump_file, repl, 0);
2056 fprintf (dump_file, "\n");
2059 sra_stats.replacements++;
2061 return repl;
2064 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
2066 static inline tree
2067 get_access_replacement (struct access *access)
2069 gcc_checking_assert (access->replacement_decl);
2070 return access->replacement_decl;
2074 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
2075 linked list along the way. Stop when *ACCESS is NULL or the access pointed
2076 to it is not "within" the root. Return false iff some accesses partially
2077 overlap. */
2079 static bool
2080 build_access_subtree (struct access **access)
2082 struct access *root = *access, *last_child = NULL;
2083 HOST_WIDE_INT limit = root->offset + root->size;
2085 *access = (*access)->next_grp;
2086 while (*access && (*access)->offset + (*access)->size <= limit)
2088 if (!last_child)
2089 root->first_child = *access;
2090 else
2091 last_child->next_sibling = *access;
2092 last_child = *access;
2094 if (!build_access_subtree (access))
2095 return false;
2098 if (*access && (*access)->offset < limit)
2099 return false;
2101 return true;
2104 /* Build a tree of access representatives, ACCESS is the pointer to the first
2105 one, others are linked in a list by the next_grp field. Return false iff
2106 some accesses partially overlap. */
2108 static bool
2109 build_access_trees (struct access *access)
2111 while (access)
2113 struct access *root = access;
2115 if (!build_access_subtree (&access))
2116 return false;
2117 root->next_grp = access;
2119 return true;
2122 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2123 array. */
2125 static bool
2126 expr_with_var_bounded_array_refs_p (tree expr)
2128 while (handled_component_p (expr))
2130 if (TREE_CODE (expr) == ARRAY_REF
2131 && !tree_fits_shwi_p (array_ref_low_bound (expr)))
2132 return true;
2133 expr = TREE_OPERAND (expr, 0);
2135 return false;
2138 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2139 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2140 sorts of access flags appropriately along the way, notably always set
2141 grp_read and grp_assign_read according to MARK_READ and grp_write when
2142 MARK_WRITE is true.
2144 Creating a replacement for a scalar access is considered beneficial if its
2145 grp_hint is set (this means we are either attempting total scalarization or
2146 there is more than one direct read access) or according to the following
2147 table:
2149 Access written to through a scalar type (once or more times)
2151 | Written to in an assignment statement
2153 | | Access read as scalar _once_
2154 | | |
2155 | | | Read in an assignment statement
2156 | | | |
2157 | | | | Scalarize Comment
2158 -----------------------------------------------------------------------------
2159 0 0 0 0 No access for the scalar
2160 0 0 0 1 No access for the scalar
2161 0 0 1 0 No Single read - won't help
2162 0 0 1 1 No The same case
2163 0 1 0 0 No access for the scalar
2164 0 1 0 1 No access for the scalar
2165 0 1 1 0 Yes s = *g; return s.i;
2166 0 1 1 1 Yes The same case as above
2167 1 0 0 0 No Won't help
2168 1 0 0 1 Yes s.i = 1; *g = s;
2169 1 0 1 0 Yes s.i = 5; g = s.i;
2170 1 0 1 1 Yes The same case as above
2171 1 1 0 0 No Won't help.
2172 1 1 0 1 Yes s.i = 1; *g = s;
2173 1 1 1 0 Yes s = *g; return s.i;
2174 1 1 1 1 Yes Any of the above yeses */
2176 static bool
2177 analyze_access_subtree (struct access *root, struct access *parent,
2178 bool allow_replacements)
2180 struct access *child;
2181 HOST_WIDE_INT limit = root->offset + root->size;
2182 HOST_WIDE_INT covered_to = root->offset;
2183 bool scalar = is_gimple_reg_type (root->type);
2184 bool hole = false, sth_created = false;
2186 if (parent)
2188 if (parent->grp_read)
2189 root->grp_read = 1;
2190 if (parent->grp_assignment_read)
2191 root->grp_assignment_read = 1;
2192 if (parent->grp_write)
2193 root->grp_write = 1;
2194 if (parent->grp_assignment_write)
2195 root->grp_assignment_write = 1;
2196 if (parent->grp_total_scalarization)
2197 root->grp_total_scalarization = 1;
2200 if (root->grp_unscalarizable_region)
2201 allow_replacements = false;
2203 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2204 allow_replacements = false;
2206 for (child = root->first_child; child; child = child->next_sibling)
2208 hole |= covered_to < child->offset;
2209 sth_created |= analyze_access_subtree (child, root,
2210 allow_replacements && !scalar);
2212 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2213 root->grp_total_scalarization &= child->grp_total_scalarization;
2214 if (child->grp_covered)
2215 covered_to += child->size;
2216 else
2217 hole = true;
2220 if (allow_replacements && scalar && !root->first_child
2221 && (root->grp_hint
2222 || ((root->grp_scalar_read || root->grp_assignment_read)
2223 && (root->grp_scalar_write || root->grp_assignment_write))))
2225 /* Always create access replacements that cover the whole access.
2226 For integral types this means the precision has to match.
2227 Avoid assumptions based on the integral type kind, too. */
2228 if (INTEGRAL_TYPE_P (root->type)
2229 && (TREE_CODE (root->type) != INTEGER_TYPE
2230 || TYPE_PRECISION (root->type) != root->size)
2231 /* But leave bitfield accesses alone. */
2232 && (TREE_CODE (root->expr) != COMPONENT_REF
2233 || !DECL_BIT_FIELD (TREE_OPERAND (root->expr, 1))))
2235 tree rt = root->type;
2236 gcc_assert ((root->offset % BITS_PER_UNIT) == 0
2237 && (root->size % BITS_PER_UNIT) == 0);
2238 root->type = build_nonstandard_integer_type (root->size,
2239 TYPE_UNSIGNED (rt));
2240 root->expr = build_ref_for_offset (UNKNOWN_LOCATION,
2241 root->base, root->offset,
2242 root->type, NULL, false);
2244 if (dump_file && (dump_flags & TDF_DETAILS))
2246 fprintf (dump_file, "Changing the type of a replacement for ");
2247 print_generic_expr (dump_file, root->base, 0);
2248 fprintf (dump_file, " offset: %u, size: %u ",
2249 (unsigned) root->offset, (unsigned) root->size);
2250 fprintf (dump_file, " to an integer.\n");
2254 root->grp_to_be_replaced = 1;
2255 root->replacement_decl = create_access_replacement (root);
2256 sth_created = true;
2257 hole = false;
2259 else
2261 if (allow_replacements
2262 && scalar && !root->first_child
2263 && (root->grp_scalar_write || root->grp_assignment_write)
2264 && !bitmap_bit_p (cannot_scalarize_away_bitmap,
2265 DECL_UID (root->base)))
2267 gcc_checking_assert (!root->grp_scalar_read
2268 && !root->grp_assignment_read);
2269 sth_created = true;
2270 if (MAY_HAVE_DEBUG_STMTS)
2272 root->grp_to_be_debug_replaced = 1;
2273 root->replacement_decl = create_access_replacement (root);
2277 if (covered_to < limit)
2278 hole = true;
2279 if (scalar)
2280 root->grp_total_scalarization = 0;
2283 if (!hole || root->grp_total_scalarization)
2284 root->grp_covered = 1;
2285 else if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2286 root->grp_unscalarized_data = 1; /* not covered and written to */
2287 return sth_created;
2290 /* Analyze all access trees linked by next_grp by the means of
2291 analyze_access_subtree. */
2292 static bool
2293 analyze_access_trees (struct access *access)
2295 bool ret = false;
2297 while (access)
2299 if (analyze_access_subtree (access, NULL, true))
2300 ret = true;
2301 access = access->next_grp;
2304 return ret;
2307 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2308 SIZE would conflict with an already existing one. If exactly such a child
2309 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2311 static bool
2312 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2313 HOST_WIDE_INT size, struct access **exact_match)
2315 struct access *child;
2317 for (child = lacc->first_child; child; child = child->next_sibling)
2319 if (child->offset == norm_offset && child->size == size)
2321 *exact_match = child;
2322 return true;
2325 if (child->offset < norm_offset + size
2326 && child->offset + child->size > norm_offset)
2327 return true;
2330 return false;
2333 /* Create a new child access of PARENT, with all properties just like MODEL
2334 except for its offset and with its grp_write false and grp_read true.
2335 Return the new access or NULL if it cannot be created. Note that this access
2336 is created long after all splicing and sorting, it's not located in any
2337 access vector and is automatically a representative of its group. */
2339 static struct access *
2340 create_artificial_child_access (struct access *parent, struct access *model,
2341 HOST_WIDE_INT new_offset)
2343 struct access *access;
2344 struct access **child;
2345 tree expr = parent->base;
2347 gcc_assert (!model->grp_unscalarizable_region);
2349 access = (struct access *) pool_alloc (access_pool);
2350 memset (access, 0, sizeof (struct access));
2351 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2352 model->type))
2354 access->grp_no_warning = true;
2355 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2356 new_offset, model, NULL, false);
2359 access->base = parent->base;
2360 access->expr = expr;
2361 access->offset = new_offset;
2362 access->size = model->size;
2363 access->type = model->type;
2364 access->grp_write = true;
2365 access->grp_read = false;
2367 child = &parent->first_child;
2368 while (*child && (*child)->offset < new_offset)
2369 child = &(*child)->next_sibling;
2371 access->next_sibling = *child;
2372 *child = access;
2374 return access;
2378 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2379 true if any new subaccess was created. Additionally, if RACC is a scalar
2380 access but LACC is not, change the type of the latter, if possible. */
2382 static bool
2383 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2385 struct access *rchild;
2386 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2387 bool ret = false;
2389 if (is_gimple_reg_type (lacc->type)
2390 || lacc->grp_unscalarizable_region
2391 || racc->grp_unscalarizable_region)
2392 return false;
2394 if (is_gimple_reg_type (racc->type))
2396 if (!lacc->first_child && !racc->first_child)
2398 tree t = lacc->base;
2400 lacc->type = racc->type;
2401 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t),
2402 lacc->offset, racc->type))
2403 lacc->expr = t;
2404 else
2406 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2407 lacc->base, lacc->offset,
2408 racc, NULL, false);
2409 lacc->grp_no_warning = true;
2412 return false;
2415 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2417 struct access *new_acc = NULL;
2418 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2420 if (rchild->grp_unscalarizable_region)
2421 continue;
2423 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2424 &new_acc))
2426 if (new_acc)
2428 rchild->grp_hint = 1;
2429 new_acc->grp_hint |= new_acc->grp_read;
2430 if (rchild->first_child)
2431 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2433 continue;
2436 rchild->grp_hint = 1;
2437 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2438 if (new_acc)
2440 ret = true;
2441 if (racc->first_child)
2442 propagate_subaccesses_across_link (new_acc, rchild);
2446 return ret;
2449 /* Propagate all subaccesses across assignment links. */
2451 static void
2452 propagate_all_subaccesses (void)
2454 while (work_queue_head)
2456 struct access *racc = pop_access_from_work_queue ();
2457 struct assign_link *link;
2459 gcc_assert (racc->first_link);
2461 for (link = racc->first_link; link; link = link->next)
2463 struct access *lacc = link->lacc;
2465 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2466 continue;
2467 lacc = lacc->group_representative;
2468 if (propagate_subaccesses_across_link (lacc, racc)
2469 && lacc->first_link)
2470 add_access_to_work_queue (lacc);
2475 /* Go through all accesses collected throughout the (intraprocedural) analysis
2476 stage, exclude overlapping ones, identify representatives and build trees
2477 out of them, making decisions about scalarization on the way. Return true
2478 iff there are any to-be-scalarized variables after this stage. */
2480 static bool
2481 analyze_all_variable_accesses (void)
2483 int res = 0;
2484 bitmap tmp = BITMAP_ALLOC (NULL);
2485 bitmap_iterator bi;
2486 unsigned i, max_total_scalarization_size;
2488 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2489 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2491 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2492 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2493 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2495 tree var = candidate (i);
2497 if (TREE_CODE (var) == VAR_DECL
2498 && type_consists_of_records_p (TREE_TYPE (var)))
2500 if (tree_to_uhwi (TYPE_SIZE (TREE_TYPE (var)))
2501 <= max_total_scalarization_size)
2503 completely_scalarize_var (var);
2504 if (dump_file && (dump_flags & TDF_DETAILS))
2506 fprintf (dump_file, "Will attempt to totally scalarize ");
2507 print_generic_expr (dump_file, var, 0);
2508 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2511 else if (dump_file && (dump_flags & TDF_DETAILS))
2513 fprintf (dump_file, "Too big to totally scalarize: ");
2514 print_generic_expr (dump_file, var, 0);
2515 fprintf (dump_file, " (UID: %u)\n", DECL_UID (var));
2520 bitmap_copy (tmp, candidate_bitmap);
2521 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2523 tree var = candidate (i);
2524 struct access *access;
2526 access = sort_and_splice_var_accesses (var);
2527 if (!access || !build_access_trees (access))
2528 disqualify_candidate (var,
2529 "No or inhibitingly overlapping accesses.");
2532 propagate_all_subaccesses ();
2534 bitmap_copy (tmp, candidate_bitmap);
2535 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2537 tree var = candidate (i);
2538 struct access *access = get_first_repr_for_decl (var);
2540 if (analyze_access_trees (access))
2542 res++;
2543 if (dump_file && (dump_flags & TDF_DETAILS))
2545 fprintf (dump_file, "\nAccess trees for ");
2546 print_generic_expr (dump_file, var, 0);
2547 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2548 dump_access_tree (dump_file, access);
2549 fprintf (dump_file, "\n");
2552 else
2553 disqualify_candidate (var, "No scalar replacements to be created.");
2556 BITMAP_FREE (tmp);
2558 if (res)
2560 statistics_counter_event (cfun, "Scalarized aggregates", res);
2561 return true;
2563 else
2564 return false;
2567 /* Generate statements copying scalar replacements of accesses within a subtree
2568 into or out of AGG. ACCESS, all its children, siblings and their children
2569 are to be processed. AGG is an aggregate type expression (can be a
2570 declaration but does not have to be, it can for example also be a mem_ref or
2571 a series of handled components). TOP_OFFSET is the offset of the processed
2572 subtree which has to be subtracted from offsets of individual accesses to
2573 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2574 replacements in the interval <start_offset, start_offset + chunk_size>,
2575 otherwise copy all. GSI is a statement iterator used to place the new
2576 statements. WRITE should be true when the statements should write from AGG
2577 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2578 statements will be added after the current statement in GSI, they will be
2579 added before the statement otherwise. */
2581 static void
2582 generate_subtree_copies (struct access *access, tree agg,
2583 HOST_WIDE_INT top_offset,
2584 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2585 gimple_stmt_iterator *gsi, bool write,
2586 bool insert_after, location_t loc)
2590 if (chunk_size && access->offset >= start_offset + chunk_size)
2591 return;
2593 if (access->grp_to_be_replaced
2594 && (chunk_size == 0
2595 || access->offset + access->size > start_offset))
2597 tree expr, repl = get_access_replacement (access);
2598 gimple stmt;
2600 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2601 access, gsi, insert_after);
2603 if (write)
2605 if (access->grp_partial_lhs)
2606 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2607 !insert_after,
2608 insert_after ? GSI_NEW_STMT
2609 : GSI_SAME_STMT);
2610 stmt = gimple_build_assign (repl, expr);
2612 else
2614 TREE_NO_WARNING (repl) = 1;
2615 if (access->grp_partial_lhs)
2616 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2617 !insert_after,
2618 insert_after ? GSI_NEW_STMT
2619 : GSI_SAME_STMT);
2620 stmt = gimple_build_assign (expr, repl);
2622 gimple_set_location (stmt, loc);
2624 if (insert_after)
2625 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2626 else
2627 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2628 update_stmt (stmt);
2629 sra_stats.subtree_copies++;
2631 else if (write
2632 && access->grp_to_be_debug_replaced
2633 && (chunk_size == 0
2634 || access->offset + access->size > start_offset))
2636 gimple ds;
2637 tree drhs = build_debug_ref_for_model (loc, agg,
2638 access->offset - top_offset,
2639 access);
2640 ds = gimple_build_debug_bind (get_access_replacement (access),
2641 drhs, gsi_stmt (*gsi));
2642 if (insert_after)
2643 gsi_insert_after (gsi, ds, GSI_NEW_STMT);
2644 else
2645 gsi_insert_before (gsi, ds, GSI_SAME_STMT);
2648 if (access->first_child)
2649 generate_subtree_copies (access->first_child, agg, top_offset,
2650 start_offset, chunk_size, gsi,
2651 write, insert_after, loc);
2653 access = access->next_sibling;
2655 while (access);
2658 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2659 the root of the subtree to be processed. GSI is the statement iterator used
2660 for inserting statements which are added after the current statement if
2661 INSERT_AFTER is true or before it otherwise. */
2663 static void
2664 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2665 bool insert_after, location_t loc)
2668 struct access *child;
2670 if (access->grp_to_be_replaced)
2672 gimple stmt;
2674 stmt = gimple_build_assign (get_access_replacement (access),
2675 build_zero_cst (access->type));
2676 if (insert_after)
2677 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2678 else
2679 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2680 update_stmt (stmt);
2681 gimple_set_location (stmt, loc);
2683 else if (access->grp_to_be_debug_replaced)
2685 gimple ds = gimple_build_debug_bind (get_access_replacement (access),
2686 build_zero_cst (access->type),
2687 gsi_stmt (*gsi));
2688 if (insert_after)
2689 gsi_insert_after (gsi, ds, GSI_NEW_STMT);
2690 else
2691 gsi_insert_before (gsi, ds, GSI_SAME_STMT);
2694 for (child = access->first_child; child; child = child->next_sibling)
2695 init_subtree_with_zero (child, gsi, insert_after, loc);
2698 /* Search for an access representative for the given expression EXPR and
2699 return it or NULL if it cannot be found. */
2701 static struct access *
2702 get_access_for_expr (tree expr)
2704 HOST_WIDE_INT offset, size, max_size;
2705 tree base;
2707 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2708 a different size than the size of its argument and we need the latter
2709 one. */
2710 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2711 expr = TREE_OPERAND (expr, 0);
2713 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2714 if (max_size == -1 || !DECL_P (base))
2715 return NULL;
2717 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2718 return NULL;
2720 return get_var_base_offset_size_access (base, offset, max_size);
2723 /* Replace the expression EXPR with a scalar replacement if there is one and
2724 generate other statements to do type conversion or subtree copying if
2725 necessary. GSI is used to place newly created statements, WRITE is true if
2726 the expression is being written to (it is on a LHS of a statement or output
2727 in an assembly statement). */
2729 static bool
2730 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2732 location_t loc;
2733 struct access *access;
2734 tree type, bfr;
2736 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2738 bfr = *expr;
2739 expr = &TREE_OPERAND (*expr, 0);
2741 else
2742 bfr = NULL_TREE;
2744 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2745 expr = &TREE_OPERAND (*expr, 0);
2746 access = get_access_for_expr (*expr);
2747 if (!access)
2748 return false;
2749 type = TREE_TYPE (*expr);
2751 loc = gimple_location (gsi_stmt (*gsi));
2752 if (access->grp_to_be_replaced)
2754 tree repl = get_access_replacement (access);
2755 /* If we replace a non-register typed access simply use the original
2756 access expression to extract the scalar component afterwards.
2757 This happens if scalarizing a function return value or parameter
2758 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2759 gcc.c-torture/compile/20011217-1.c.
2761 We also want to use this when accessing a complex or vector which can
2762 be accessed as a different type too, potentially creating a need for
2763 type conversion (see PR42196) and when scalarized unions are involved
2764 in assembler statements (see PR42398). */
2765 if (!useless_type_conversion_p (type, access->type))
2767 tree ref;
2769 ref = build_ref_for_model (loc, access->base, access->offset, access,
2770 NULL, false);
2772 if (write)
2774 gimple stmt;
2776 if (access->grp_partial_lhs)
2777 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2778 false, GSI_NEW_STMT);
2779 stmt = gimple_build_assign (repl, ref);
2780 gimple_set_location (stmt, loc);
2781 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2783 else
2785 gimple stmt;
2787 if (access->grp_partial_lhs)
2788 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2789 true, GSI_SAME_STMT);
2790 stmt = gimple_build_assign (ref, repl);
2791 gimple_set_location (stmt, loc);
2792 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2795 else
2796 *expr = repl;
2797 sra_stats.exprs++;
2799 else if (write && access->grp_to_be_debug_replaced)
2801 gimple ds = gimple_build_debug_bind (get_access_replacement (access),
2802 NULL_TREE,
2803 gsi_stmt (*gsi));
2804 gsi_insert_after (gsi, ds, GSI_NEW_STMT);
2807 if (access->first_child)
2809 HOST_WIDE_INT start_offset, chunk_size;
2810 if (bfr
2811 && tree_fits_uhwi_p (TREE_OPERAND (bfr, 1))
2812 && tree_fits_uhwi_p (TREE_OPERAND (bfr, 2)))
2814 chunk_size = tree_to_uhwi (TREE_OPERAND (bfr, 1));
2815 start_offset = access->offset
2816 + tree_to_uhwi (TREE_OPERAND (bfr, 2));
2818 else
2819 start_offset = chunk_size = 0;
2821 generate_subtree_copies (access->first_child, access->base, 0,
2822 start_offset, chunk_size, gsi, write, write,
2823 loc);
2825 return true;
2828 /* Where scalar replacements of the RHS have been written to when a replacement
2829 of a LHS of an assigments cannot be direclty loaded from a replacement of
2830 the RHS. */
2831 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2832 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2833 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2835 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2836 base aggregate if there are unscalarized data or directly to LHS of the
2837 statement that is pointed to by GSI otherwise. */
2839 static enum unscalarized_data_handling
2840 handle_unscalarized_data_in_subtree (struct access *top_racc,
2841 gimple_stmt_iterator *gsi)
2843 if (top_racc->grp_unscalarized_data)
2845 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2846 gsi, false, false,
2847 gimple_location (gsi_stmt (*gsi)));
2848 return SRA_UDH_RIGHT;
2850 else
2852 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2853 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2854 0, 0, gsi, false, false,
2855 gimple_location (gsi_stmt (*gsi)));
2856 return SRA_UDH_LEFT;
2861 /* Try to generate statements to load all sub-replacements in an access subtree
2862 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2863 If that is not possible, refresh the TOP_RACC base aggregate and load the
2864 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2865 copied. NEW_GSI is stmt iterator used for statement insertions after the
2866 original assignment, OLD_GSI is used to insert statements before the
2867 assignment. *REFRESHED keeps the information whether we have needed to
2868 refresh replacements of the LHS and from which side of the assignments this
2869 takes place. */
2871 static void
2872 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2873 HOST_WIDE_INT left_offset,
2874 gimple_stmt_iterator *old_gsi,
2875 gimple_stmt_iterator *new_gsi,
2876 enum unscalarized_data_handling *refreshed)
2878 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2879 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2881 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2883 if (lacc->grp_to_be_replaced)
2885 struct access *racc;
2886 gimple stmt;
2887 tree rhs;
2889 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2890 if (racc && racc->grp_to_be_replaced)
2892 rhs = get_access_replacement (racc);
2893 if (!useless_type_conversion_p (lacc->type, racc->type))
2894 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2896 if (racc->grp_partial_lhs && lacc->grp_partial_lhs)
2897 rhs = force_gimple_operand_gsi (old_gsi, rhs, true, NULL_TREE,
2898 true, GSI_SAME_STMT);
2900 else
2902 /* No suitable access on the right hand side, need to load from
2903 the aggregate. See if we have to update it first... */
2904 if (*refreshed == SRA_UDH_NONE)
2905 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2906 old_gsi);
2908 if (*refreshed == SRA_UDH_LEFT)
2909 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2910 new_gsi, true);
2911 else
2912 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2913 new_gsi, true);
2914 if (lacc->grp_partial_lhs)
2915 rhs = force_gimple_operand_gsi (new_gsi, rhs, true, NULL_TREE,
2916 false, GSI_NEW_STMT);
2919 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2920 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2921 gimple_set_location (stmt, loc);
2922 update_stmt (stmt);
2923 sra_stats.subreplacements++;
2925 else
2927 if (*refreshed == SRA_UDH_NONE
2928 && lacc->grp_read && !lacc->grp_covered)
2929 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2930 old_gsi);
2931 if (lacc && lacc->grp_to_be_debug_replaced)
2933 gimple ds;
2934 tree drhs;
2935 struct access *racc = find_access_in_subtree (top_racc, offset,
2936 lacc->size);
2938 if (racc && racc->grp_to_be_replaced)
2940 if (racc->grp_write)
2941 drhs = get_access_replacement (racc);
2942 else
2943 drhs = NULL;
2945 else if (*refreshed == SRA_UDH_LEFT)
2946 drhs = build_debug_ref_for_model (loc, lacc->base, lacc->offset,
2947 lacc);
2948 else if (*refreshed == SRA_UDH_RIGHT)
2949 drhs = build_debug_ref_for_model (loc, top_racc->base, offset,
2950 lacc);
2951 else
2952 drhs = NULL_TREE;
2953 ds = gimple_build_debug_bind (get_access_replacement (lacc),
2954 drhs, gsi_stmt (*old_gsi));
2955 gsi_insert_after (new_gsi, ds, GSI_NEW_STMT);
2959 if (lacc->first_child)
2960 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2961 old_gsi, new_gsi, refreshed);
2965 /* Result code for SRA assignment modification. */
2966 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2967 SRA_AM_MODIFIED, /* stmt changed but not
2968 removed */
2969 SRA_AM_REMOVED }; /* stmt eliminated */
2971 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2972 to the assignment and GSI is the statement iterator pointing at it. Returns
2973 the same values as sra_modify_assign. */
2975 static enum assignment_mod_result
2976 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2978 tree lhs = gimple_assign_lhs (*stmt);
2979 struct access *acc;
2980 location_t loc;
2982 acc = get_access_for_expr (lhs);
2983 if (!acc)
2984 return SRA_AM_NONE;
2986 if (gimple_clobber_p (*stmt))
2988 /* Remove clobbers of fully scalarized variables, otherwise
2989 do nothing. */
2990 if (acc->grp_covered)
2992 unlink_stmt_vdef (*stmt);
2993 gsi_remove (gsi, true);
2994 release_defs (*stmt);
2995 return SRA_AM_REMOVED;
2997 else
2998 return SRA_AM_NONE;
3001 loc = gimple_location (*stmt);
3002 if (vec_safe_length (CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
3004 /* I have never seen this code path trigger but if it can happen the
3005 following should handle it gracefully. */
3006 if (access_has_children_p (acc))
3007 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
3008 true, true, loc);
3009 return SRA_AM_MODIFIED;
3012 if (acc->grp_covered)
3014 init_subtree_with_zero (acc, gsi, false, loc);
3015 unlink_stmt_vdef (*stmt);
3016 gsi_remove (gsi, true);
3017 release_defs (*stmt);
3018 return SRA_AM_REMOVED;
3020 else
3022 init_subtree_with_zero (acc, gsi, true, loc);
3023 return SRA_AM_MODIFIED;
3027 /* Create and return a new suitable default definition SSA_NAME for RACC which
3028 is an access describing an uninitialized part of an aggregate that is being
3029 loaded. */
3031 static tree
3032 get_repl_default_def_ssa_name (struct access *racc)
3034 gcc_checking_assert (!racc->grp_to_be_replaced
3035 && !racc->grp_to_be_debug_replaced);
3036 if (!racc->replacement_decl)
3037 racc->replacement_decl = create_access_replacement (racc);
3038 return get_or_create_ssa_default_def (cfun, racc->replacement_decl);
3041 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
3042 bit-field field declaration somewhere in it. */
3044 static inline bool
3045 contains_vce_or_bfcref_p (const_tree ref)
3047 while (handled_component_p (ref))
3049 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
3050 || (TREE_CODE (ref) == COMPONENT_REF
3051 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
3052 return true;
3053 ref = TREE_OPERAND (ref, 0);
3056 return false;
3059 /* Examine both sides of the assignment statement pointed to by STMT, replace
3060 them with a scalare replacement if there is one and generate copying of
3061 replacements if scalarized aggregates have been used in the assignment. GSI
3062 is used to hold generated statements for type conversions and subtree
3063 copying. */
3065 static enum assignment_mod_result
3066 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
3068 struct access *lacc, *racc;
3069 tree lhs, rhs;
3070 bool modify_this_stmt = false;
3071 bool force_gimple_rhs = false;
3072 location_t loc;
3073 gimple_stmt_iterator orig_gsi = *gsi;
3075 if (!gimple_assign_single_p (*stmt))
3076 return SRA_AM_NONE;
3077 lhs = gimple_assign_lhs (*stmt);
3078 rhs = gimple_assign_rhs1 (*stmt);
3080 if (TREE_CODE (rhs) == CONSTRUCTOR)
3081 return sra_modify_constructor_assign (stmt, gsi);
3083 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
3084 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
3085 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
3087 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
3088 gsi, false);
3089 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
3090 gsi, true);
3091 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3094 lacc = get_access_for_expr (lhs);
3095 racc = get_access_for_expr (rhs);
3096 if (!lacc && !racc)
3097 return SRA_AM_NONE;
3099 loc = gimple_location (*stmt);
3100 if (lacc && lacc->grp_to_be_replaced)
3102 lhs = get_access_replacement (lacc);
3103 gimple_assign_set_lhs (*stmt, lhs);
3104 modify_this_stmt = true;
3105 if (lacc->grp_partial_lhs)
3106 force_gimple_rhs = true;
3107 sra_stats.exprs++;
3110 if (racc && racc->grp_to_be_replaced)
3112 rhs = get_access_replacement (racc);
3113 modify_this_stmt = true;
3114 if (racc->grp_partial_lhs)
3115 force_gimple_rhs = true;
3116 sra_stats.exprs++;
3118 else if (racc
3119 && !racc->grp_unscalarized_data
3120 && TREE_CODE (lhs) == SSA_NAME
3121 && !access_has_replacements_p (racc))
3123 rhs = get_repl_default_def_ssa_name (racc);
3124 modify_this_stmt = true;
3125 sra_stats.exprs++;
3128 if (modify_this_stmt)
3130 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3132 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
3133 ??? This should move to fold_stmt which we simply should
3134 call after building a VIEW_CONVERT_EXPR here. */
3135 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
3136 && !contains_bitfld_component_ref_p (lhs))
3138 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
3139 gimple_assign_set_lhs (*stmt, lhs);
3141 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
3142 && !contains_vce_or_bfcref_p (rhs))
3143 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
3145 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3147 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
3148 rhs);
3149 if (is_gimple_reg_type (TREE_TYPE (lhs))
3150 && TREE_CODE (lhs) != SSA_NAME)
3151 force_gimple_rhs = true;
3156 if (lacc && lacc->grp_to_be_debug_replaced)
3158 tree dlhs = get_access_replacement (lacc);
3159 tree drhs = unshare_expr (rhs);
3160 if (!useless_type_conversion_p (TREE_TYPE (dlhs), TREE_TYPE (drhs)))
3162 if (AGGREGATE_TYPE_P (TREE_TYPE (drhs))
3163 && !contains_vce_or_bfcref_p (drhs))
3164 drhs = build_debug_ref_for_model (loc, drhs, 0, lacc);
3165 if (drhs
3166 && !useless_type_conversion_p (TREE_TYPE (dlhs),
3167 TREE_TYPE (drhs)))
3168 drhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR,
3169 TREE_TYPE (dlhs), drhs);
3171 gimple ds = gimple_build_debug_bind (dlhs, drhs, *stmt);
3172 gsi_insert_before (gsi, ds, GSI_SAME_STMT);
3175 /* From this point on, the function deals with assignments in between
3176 aggregates when at least one has scalar reductions of some of its
3177 components. There are three possible scenarios: Both the LHS and RHS have
3178 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
3180 In the first case, we would like to load the LHS components from RHS
3181 components whenever possible. If that is not possible, we would like to
3182 read it directly from the RHS (after updating it by storing in it its own
3183 components). If there are some necessary unscalarized data in the LHS,
3184 those will be loaded by the original assignment too. If neither of these
3185 cases happen, the original statement can be removed. Most of this is done
3186 by load_assign_lhs_subreplacements.
3188 In the second case, we would like to store all RHS scalarized components
3189 directly into LHS and if they cover the aggregate completely, remove the
3190 statement too. In the third case, we want the LHS components to be loaded
3191 directly from the RHS (DSE will remove the original statement if it
3192 becomes redundant).
3194 This is a bit complex but manageable when types match and when unions do
3195 not cause confusion in a way that we cannot really load a component of LHS
3196 from the RHS or vice versa (the access representing this level can have
3197 subaccesses that are accessible only through a different union field at a
3198 higher level - different from the one used in the examined expression).
3199 Unions are fun.
3201 Therefore, I specially handle a fourth case, happening when there is a
3202 specific type cast or it is impossible to locate a scalarized subaccess on
3203 the other side of the expression. If that happens, I simply "refresh" the
3204 RHS by storing in it is scalarized components leave the original statement
3205 there to do the copying and then load the scalar replacements of the LHS.
3206 This is what the first branch does. */
3208 if (modify_this_stmt
3209 || gimple_has_volatile_ops (*stmt)
3210 || contains_vce_or_bfcref_p (rhs)
3211 || contains_vce_or_bfcref_p (lhs))
3213 if (access_has_children_p (racc))
3214 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3215 gsi, false, false, loc);
3216 if (access_has_children_p (lacc))
3217 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
3218 gsi, true, true, loc);
3219 sra_stats.separate_lhs_rhs_handling++;
3221 /* This gimplification must be done after generate_subtree_copies,
3222 lest we insert the subtree copies in the middle of the gimplified
3223 sequence. */
3224 if (force_gimple_rhs)
3225 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
3226 true, GSI_SAME_STMT);
3227 if (gimple_assign_rhs1 (*stmt) != rhs)
3229 modify_this_stmt = true;
3230 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
3231 gcc_assert (*stmt == gsi_stmt (orig_gsi));
3234 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3236 else
3238 if (access_has_children_p (lacc)
3239 && access_has_children_p (racc)
3240 /* When an access represents an unscalarizable region, it usually
3241 represents accesses with variable offset and thus must not be used
3242 to generate new memory accesses. */
3243 && !lacc->grp_unscalarizable_region
3244 && !racc->grp_unscalarizable_region)
3246 gimple_stmt_iterator orig_gsi = *gsi;
3247 enum unscalarized_data_handling refreshed;
3249 if (lacc->grp_read && !lacc->grp_covered)
3250 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
3251 else
3252 refreshed = SRA_UDH_NONE;
3254 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
3255 &orig_gsi, gsi, &refreshed);
3256 if (refreshed != SRA_UDH_RIGHT)
3258 gsi_next (gsi);
3259 unlink_stmt_vdef (*stmt);
3260 gsi_remove (&orig_gsi, true);
3261 release_defs (*stmt);
3262 sra_stats.deleted++;
3263 return SRA_AM_REMOVED;
3266 else
3268 if (access_has_children_p (racc)
3269 && !racc->grp_unscalarized_data)
3271 if (dump_file)
3273 fprintf (dump_file, "Removing load: ");
3274 print_gimple_stmt (dump_file, *stmt, 0, 0);
3276 generate_subtree_copies (racc->first_child, lhs,
3277 racc->offset, 0, 0, gsi,
3278 false, false, loc);
3279 gcc_assert (*stmt == gsi_stmt (*gsi));
3280 unlink_stmt_vdef (*stmt);
3281 gsi_remove (gsi, true);
3282 release_defs (*stmt);
3283 sra_stats.deleted++;
3284 return SRA_AM_REMOVED;
3286 /* Restore the aggregate RHS from its components so the
3287 prevailing aggregate copy does the right thing. */
3288 if (access_has_children_p (racc))
3289 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3290 gsi, false, false, loc);
3291 /* Re-load the components of the aggregate copy destination.
3292 But use the RHS aggregate to load from to expose more
3293 optimization opportunities. */
3294 if (access_has_children_p (lacc))
3295 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
3296 0, 0, gsi, true, true, loc);
3299 return SRA_AM_NONE;
3303 /* Traverse the function body and all modifications as decided in
3304 analyze_all_variable_accesses. Return true iff the CFG has been
3305 changed. */
3307 static bool
3308 sra_modify_function_body (void)
3310 bool cfg_changed = false;
3311 basic_block bb;
3313 FOR_EACH_BB_FN (bb, cfun)
3315 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3316 while (!gsi_end_p (gsi))
3318 gimple stmt = gsi_stmt (gsi);
3319 enum assignment_mod_result assign_result;
3320 bool modified = false, deleted = false;
3321 tree *t;
3322 unsigned i;
3324 switch (gimple_code (stmt))
3326 case GIMPLE_RETURN:
3327 t = gimple_return_retval_ptr (stmt);
3328 if (*t != NULL_TREE)
3329 modified |= sra_modify_expr (t, &gsi, false);
3330 break;
3332 case GIMPLE_ASSIGN:
3333 assign_result = sra_modify_assign (&stmt, &gsi);
3334 modified |= assign_result == SRA_AM_MODIFIED;
3335 deleted = assign_result == SRA_AM_REMOVED;
3336 break;
3338 case GIMPLE_CALL:
3339 /* Operands must be processed before the lhs. */
3340 for (i = 0; i < gimple_call_num_args (stmt); i++)
3342 t = gimple_call_arg_ptr (stmt, i);
3343 modified |= sra_modify_expr (t, &gsi, false);
3346 if (gimple_call_lhs (stmt))
3348 t = gimple_call_lhs_ptr (stmt);
3349 modified |= sra_modify_expr (t, &gsi, true);
3351 break;
3353 case GIMPLE_ASM:
3354 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3356 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3357 modified |= sra_modify_expr (t, &gsi, false);
3359 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3361 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3362 modified |= sra_modify_expr (t, &gsi, true);
3364 break;
3366 default:
3367 break;
3370 if (modified)
3372 update_stmt (stmt);
3373 if (maybe_clean_eh_stmt (stmt)
3374 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3375 cfg_changed = true;
3377 if (!deleted)
3378 gsi_next (&gsi);
3382 return cfg_changed;
3385 /* Generate statements initializing scalar replacements of parts of function
3386 parameters. */
3388 static void
3389 initialize_parameter_reductions (void)
3391 gimple_stmt_iterator gsi;
3392 gimple_seq seq = NULL;
3393 tree parm;
3395 gsi = gsi_start (seq);
3396 for (parm = DECL_ARGUMENTS (current_function_decl);
3397 parm;
3398 parm = DECL_CHAIN (parm))
3400 vec<access_p> *access_vec;
3401 struct access *access;
3403 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3404 continue;
3405 access_vec = get_base_access_vector (parm);
3406 if (!access_vec)
3407 continue;
3409 for (access = (*access_vec)[0];
3410 access;
3411 access = access->next_grp)
3412 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3413 EXPR_LOCATION (parm));
3416 seq = gsi_seq (gsi);
3417 if (seq)
3418 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR_FOR_FN (cfun)), seq);
3421 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3422 it reveals there are components of some aggregates to be scalarized, it runs
3423 the required transformations. */
3424 static unsigned int
3425 perform_intra_sra (void)
3427 int ret = 0;
3428 sra_initialize ();
3430 if (!find_var_candidates ())
3431 goto out;
3433 if (!scan_function ())
3434 goto out;
3436 if (!analyze_all_variable_accesses ())
3437 goto out;
3439 if (sra_modify_function_body ())
3440 ret = TODO_update_ssa | TODO_cleanup_cfg;
3441 else
3442 ret = TODO_update_ssa;
3443 initialize_parameter_reductions ();
3445 statistics_counter_event (cfun, "Scalar replacements created",
3446 sra_stats.replacements);
3447 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3448 statistics_counter_event (cfun, "Subtree copy stmts",
3449 sra_stats.subtree_copies);
3450 statistics_counter_event (cfun, "Subreplacement stmts",
3451 sra_stats.subreplacements);
3452 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3453 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3454 sra_stats.separate_lhs_rhs_handling);
3456 out:
3457 sra_deinitialize ();
3458 return ret;
3461 /* Perform early intraprocedural SRA. */
3462 static unsigned int
3463 early_intra_sra (void)
3465 sra_mode = SRA_MODE_EARLY_INTRA;
3466 return perform_intra_sra ();
3469 /* Perform "late" intraprocedural SRA. */
3470 static unsigned int
3471 late_intra_sra (void)
3473 sra_mode = SRA_MODE_INTRA;
3474 return perform_intra_sra ();
3478 static bool
3479 gate_intra_sra (void)
3481 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3485 namespace {
3487 const pass_data pass_data_sra_early =
3489 GIMPLE_PASS, /* type */
3490 "esra", /* name */
3491 OPTGROUP_NONE, /* optinfo_flags */
3492 true, /* has_gate */
3493 true, /* has_execute */
3494 TV_TREE_SRA, /* tv_id */
3495 ( PROP_cfg | PROP_ssa ), /* properties_required */
3496 0, /* properties_provided */
3497 0, /* properties_destroyed */
3498 0, /* todo_flags_start */
3499 ( TODO_update_ssa | TODO_verify_ssa ), /* todo_flags_finish */
3502 class pass_sra_early : public gimple_opt_pass
3504 public:
3505 pass_sra_early (gcc::context *ctxt)
3506 : gimple_opt_pass (pass_data_sra_early, ctxt)
3509 /* opt_pass methods: */
3510 bool gate () { return gate_intra_sra (); }
3511 unsigned int execute () { return early_intra_sra (); }
3513 }; // class pass_sra_early
3515 } // anon namespace
3517 gimple_opt_pass *
3518 make_pass_sra_early (gcc::context *ctxt)
3520 return new pass_sra_early (ctxt);
3523 namespace {
3525 const pass_data pass_data_sra =
3527 GIMPLE_PASS, /* type */
3528 "sra", /* name */
3529 OPTGROUP_NONE, /* optinfo_flags */
3530 true, /* has_gate */
3531 true, /* has_execute */
3532 TV_TREE_SRA, /* tv_id */
3533 ( PROP_cfg | PROP_ssa ), /* properties_required */
3534 0, /* properties_provided */
3535 0, /* properties_destroyed */
3536 TODO_update_address_taken, /* todo_flags_start */
3537 ( TODO_update_ssa | TODO_verify_ssa ), /* todo_flags_finish */
3540 class pass_sra : public gimple_opt_pass
3542 public:
3543 pass_sra (gcc::context *ctxt)
3544 : gimple_opt_pass (pass_data_sra, ctxt)
3547 /* opt_pass methods: */
3548 bool gate () { return gate_intra_sra (); }
3549 unsigned int execute () { return late_intra_sra (); }
3551 }; // class pass_sra
3553 } // anon namespace
3555 gimple_opt_pass *
3556 make_pass_sra (gcc::context *ctxt)
3558 return new pass_sra (ctxt);
3562 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3563 parameter. */
3565 static bool
3566 is_unused_scalar_param (tree parm)
3568 tree name;
3569 return (is_gimple_reg (parm)
3570 && (!(name = ssa_default_def (cfun, parm))
3571 || has_zero_uses (name)));
3574 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3575 examine whether there are any direct or otherwise infeasible ones. If so,
3576 return true, otherwise return false. PARM must be a gimple register with a
3577 non-NULL default definition. */
3579 static bool
3580 ptr_parm_has_direct_uses (tree parm)
3582 imm_use_iterator ui;
3583 gimple stmt;
3584 tree name = ssa_default_def (cfun, parm);
3585 bool ret = false;
3587 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3589 int uses_ok = 0;
3590 use_operand_p use_p;
3592 if (is_gimple_debug (stmt))
3593 continue;
3595 /* Valid uses include dereferences on the lhs and the rhs. */
3596 if (gimple_has_lhs (stmt))
3598 tree lhs = gimple_get_lhs (stmt);
3599 while (handled_component_p (lhs))
3600 lhs = TREE_OPERAND (lhs, 0);
3601 if (TREE_CODE (lhs) == MEM_REF
3602 && TREE_OPERAND (lhs, 0) == name
3603 && integer_zerop (TREE_OPERAND (lhs, 1))
3604 && types_compatible_p (TREE_TYPE (lhs),
3605 TREE_TYPE (TREE_TYPE (name)))
3606 && !TREE_THIS_VOLATILE (lhs))
3607 uses_ok++;
3609 if (gimple_assign_single_p (stmt))
3611 tree rhs = gimple_assign_rhs1 (stmt);
3612 while (handled_component_p (rhs))
3613 rhs = TREE_OPERAND (rhs, 0);
3614 if (TREE_CODE (rhs) == MEM_REF
3615 && TREE_OPERAND (rhs, 0) == name
3616 && integer_zerop (TREE_OPERAND (rhs, 1))
3617 && types_compatible_p (TREE_TYPE (rhs),
3618 TREE_TYPE (TREE_TYPE (name)))
3619 && !TREE_THIS_VOLATILE (rhs))
3620 uses_ok++;
3622 else if (is_gimple_call (stmt))
3624 unsigned i;
3625 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3627 tree arg = gimple_call_arg (stmt, i);
3628 while (handled_component_p (arg))
3629 arg = TREE_OPERAND (arg, 0);
3630 if (TREE_CODE (arg) == MEM_REF
3631 && TREE_OPERAND (arg, 0) == name
3632 && integer_zerop (TREE_OPERAND (arg, 1))
3633 && types_compatible_p (TREE_TYPE (arg),
3634 TREE_TYPE (TREE_TYPE (name)))
3635 && !TREE_THIS_VOLATILE (arg))
3636 uses_ok++;
3640 /* If the number of valid uses does not match the number of
3641 uses in this stmt there is an unhandled use. */
3642 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3643 --uses_ok;
3645 if (uses_ok != 0)
3646 ret = true;
3648 if (ret)
3649 BREAK_FROM_IMM_USE_STMT (ui);
3652 return ret;
3655 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3656 them in candidate_bitmap. Note that these do not necessarily include
3657 parameter which are unused and thus can be removed. Return true iff any
3658 such candidate has been found. */
3660 static bool
3661 find_param_candidates (void)
3663 tree parm;
3664 int count = 0;
3665 bool ret = false;
3666 const char *msg;
3668 for (parm = DECL_ARGUMENTS (current_function_decl);
3669 parm;
3670 parm = DECL_CHAIN (parm))
3672 tree type = TREE_TYPE (parm);
3673 tree_node **slot;
3675 count++;
3677 if (TREE_THIS_VOLATILE (parm)
3678 || TREE_ADDRESSABLE (parm)
3679 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3680 continue;
3682 if (is_unused_scalar_param (parm))
3684 ret = true;
3685 continue;
3688 if (POINTER_TYPE_P (type))
3690 type = TREE_TYPE (type);
3692 if (TREE_CODE (type) == FUNCTION_TYPE
3693 || TYPE_VOLATILE (type)
3694 || (TREE_CODE (type) == ARRAY_TYPE
3695 && TYPE_NONALIASED_COMPONENT (type))
3696 || !is_gimple_reg (parm)
3697 || is_va_list_type (type)
3698 || ptr_parm_has_direct_uses (parm))
3699 continue;
3701 else if (!AGGREGATE_TYPE_P (type))
3702 continue;
3704 if (!COMPLETE_TYPE_P (type)
3705 || !tree_fits_uhwi_p (TYPE_SIZE (type))
3706 || tree_to_uhwi (TYPE_SIZE (type)) == 0
3707 || (AGGREGATE_TYPE_P (type)
3708 && type_internals_preclude_sra_p (type, &msg)))
3709 continue;
3711 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3712 slot = candidates.find_slot_with_hash (parm, DECL_UID (parm), INSERT);
3713 *slot = parm;
3715 ret = true;
3716 if (dump_file && (dump_flags & TDF_DETAILS))
3718 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3719 print_generic_expr (dump_file, parm, 0);
3720 fprintf (dump_file, "\n");
3724 func_param_count = count;
3725 return ret;
3728 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3729 maybe_modified. */
3731 static bool
3732 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3733 void *data)
3735 struct access *repr = (struct access *) data;
3737 repr->grp_maybe_modified = 1;
3738 return true;
3741 /* Analyze what representatives (in linked lists accessible from
3742 REPRESENTATIVES) can be modified by side effects of statements in the
3743 current function. */
3745 static void
3746 analyze_modified_params (vec<access_p> representatives)
3748 int i;
3750 for (i = 0; i < func_param_count; i++)
3752 struct access *repr;
3754 for (repr = representatives[i];
3755 repr;
3756 repr = repr->next_grp)
3758 struct access *access;
3759 bitmap visited;
3760 ao_ref ar;
3762 if (no_accesses_p (repr))
3763 continue;
3764 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3765 || repr->grp_maybe_modified)
3766 continue;
3768 ao_ref_init (&ar, repr->expr);
3769 visited = BITMAP_ALLOC (NULL);
3770 for (access = repr; access; access = access->next_sibling)
3772 /* All accesses are read ones, otherwise grp_maybe_modified would
3773 be trivially set. */
3774 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3775 mark_maybe_modified, repr, &visited);
3776 if (repr->grp_maybe_modified)
3777 break;
3779 BITMAP_FREE (visited);
3784 /* Propagate distances in bb_dereferences in the opposite direction than the
3785 control flow edges, in each step storing the maximum of the current value
3786 and the minimum of all successors. These steps are repeated until the table
3787 stabilizes. Note that BBs which might terminate the functions (according to
3788 final_bbs bitmap) never updated in this way. */
3790 static void
3791 propagate_dereference_distances (void)
3793 basic_block bb;
3795 auto_vec<basic_block> queue (last_basic_block_for_fn (cfun));
3796 queue.quick_push (ENTRY_BLOCK_PTR_FOR_FN (cfun));
3797 FOR_EACH_BB_FN (bb, cfun)
3799 queue.quick_push (bb);
3800 bb->aux = bb;
3803 while (!queue.is_empty ())
3805 edge_iterator ei;
3806 edge e;
3807 bool change = false;
3808 int i;
3810 bb = queue.pop ();
3811 bb->aux = NULL;
3813 if (bitmap_bit_p (final_bbs, bb->index))
3814 continue;
3816 for (i = 0; i < func_param_count; i++)
3818 int idx = bb->index * func_param_count + i;
3819 bool first = true;
3820 HOST_WIDE_INT inh = 0;
3822 FOR_EACH_EDGE (e, ei, bb->succs)
3824 int succ_idx = e->dest->index * func_param_count + i;
3826 if (e->src == EXIT_BLOCK_PTR_FOR_FN (cfun))
3827 continue;
3829 if (first)
3831 first = false;
3832 inh = bb_dereferences [succ_idx];
3834 else if (bb_dereferences [succ_idx] < inh)
3835 inh = bb_dereferences [succ_idx];
3838 if (!first && bb_dereferences[idx] < inh)
3840 bb_dereferences[idx] = inh;
3841 change = true;
3845 if (change && !bitmap_bit_p (final_bbs, bb->index))
3846 FOR_EACH_EDGE (e, ei, bb->preds)
3848 if (e->src->aux)
3849 continue;
3851 e->src->aux = e->src;
3852 queue.quick_push (e->src);
3857 /* Dump a dereferences TABLE with heading STR to file F. */
3859 static void
3860 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3862 basic_block bb;
3864 fprintf (dump_file, str);
3865 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR_FOR_FN (cfun),
3866 EXIT_BLOCK_PTR_FOR_FN (cfun), next_bb)
3868 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3869 if (bb != EXIT_BLOCK_PTR_FOR_FN (cfun))
3871 int i;
3872 for (i = 0; i < func_param_count; i++)
3874 int idx = bb->index * func_param_count + i;
3875 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3878 fprintf (f, "\n");
3880 fprintf (dump_file, "\n");
3883 /* Determine what (parts of) parameters passed by reference that are not
3884 assigned to are not certainly dereferenced in this function and thus the
3885 dereferencing cannot be safely moved to the caller without potentially
3886 introducing a segfault. Mark such REPRESENTATIVES as
3887 grp_not_necessarilly_dereferenced.
3889 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3890 part is calculated rather than simple booleans are calculated for each
3891 pointer parameter to handle cases when only a fraction of the whole
3892 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3893 an example).
3895 The maximum dereference distances for each pointer parameter and BB are
3896 already stored in bb_dereference. This routine simply propagates these
3897 values upwards by propagate_dereference_distances and then compares the
3898 distances of individual parameters in the ENTRY BB to the equivalent
3899 distances of each representative of a (fraction of a) parameter. */
3901 static void
3902 analyze_caller_dereference_legality (vec<access_p> representatives)
3904 int i;
3906 if (dump_file && (dump_flags & TDF_DETAILS))
3907 dump_dereferences_table (dump_file,
3908 "Dereference table before propagation:\n",
3909 bb_dereferences);
3911 propagate_dereference_distances ();
3913 if (dump_file && (dump_flags & TDF_DETAILS))
3914 dump_dereferences_table (dump_file,
3915 "Dereference table after propagation:\n",
3916 bb_dereferences);
3918 for (i = 0; i < func_param_count; i++)
3920 struct access *repr = representatives[i];
3921 int idx = ENTRY_BLOCK_PTR_FOR_FN (cfun)->index * func_param_count + i;
3923 if (!repr || no_accesses_p (repr))
3924 continue;
3928 if ((repr->offset + repr->size) > bb_dereferences[idx])
3929 repr->grp_not_necessarilly_dereferenced = 1;
3930 repr = repr->next_grp;
3932 while (repr);
3936 /* Return the representative access for the parameter declaration PARM if it is
3937 a scalar passed by reference which is not written to and the pointer value
3938 is not used directly. Thus, if it is legal to dereference it in the caller
3939 and we can rule out modifications through aliases, such parameter should be
3940 turned into one passed by value. Return NULL otherwise. */
3942 static struct access *
3943 unmodified_by_ref_scalar_representative (tree parm)
3945 int i, access_count;
3946 struct access *repr;
3947 vec<access_p> *access_vec;
3949 access_vec = get_base_access_vector (parm);
3950 gcc_assert (access_vec);
3951 repr = (*access_vec)[0];
3952 if (repr->write)
3953 return NULL;
3954 repr->group_representative = repr;
3956 access_count = access_vec->length ();
3957 for (i = 1; i < access_count; i++)
3959 struct access *access = (*access_vec)[i];
3960 if (access->write)
3961 return NULL;
3962 access->group_representative = repr;
3963 access->next_sibling = repr->next_sibling;
3964 repr->next_sibling = access;
3967 repr->grp_read = 1;
3968 repr->grp_scalar_ptr = 1;
3969 return repr;
3972 /* Return true iff this ACCESS precludes IPA-SRA of the parameter it is
3973 associated with. REQ_ALIGN is the minimum required alignment. */
3975 static bool
3976 access_precludes_ipa_sra_p (struct access *access, unsigned int req_align)
3978 unsigned int exp_align;
3979 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3980 is incompatible assign in a call statement (and possibly even in asm
3981 statements). This can be relaxed by using a new temporary but only for
3982 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3983 intraprocedural SRA we deal with this by keeping the old aggregate around,
3984 something we cannot do in IPA-SRA.) */
3985 if (access->write
3986 && (is_gimple_call (access->stmt)
3987 || gimple_code (access->stmt) == GIMPLE_ASM))
3988 return true;
3990 exp_align = get_object_alignment (access->expr);
3991 if (exp_align < req_align)
3992 return true;
3994 return false;
3998 /* Sort collected accesses for parameter PARM, identify representatives for
3999 each accessed region and link them together. Return NULL if there are
4000 different but overlapping accesses, return the special ptr value meaning
4001 there are no accesses for this parameter if that is the case and return the
4002 first representative otherwise. Set *RO_GRP if there is a group of accesses
4003 with only read (i.e. no write) accesses. */
4005 static struct access *
4006 splice_param_accesses (tree parm, bool *ro_grp)
4008 int i, j, access_count, group_count;
4009 int agg_size, total_size = 0;
4010 struct access *access, *res, **prev_acc_ptr = &res;
4011 vec<access_p> *access_vec;
4013 access_vec = get_base_access_vector (parm);
4014 if (!access_vec)
4015 return &no_accesses_representant;
4016 access_count = access_vec->length ();
4018 access_vec->qsort (compare_access_positions);
4020 i = 0;
4021 total_size = 0;
4022 group_count = 0;
4023 while (i < access_count)
4025 bool modification;
4026 tree a1_alias_type;
4027 access = (*access_vec)[i];
4028 modification = access->write;
4029 if (access_precludes_ipa_sra_p (access, TYPE_ALIGN (access->type)))
4030 return NULL;
4031 a1_alias_type = reference_alias_ptr_type (access->expr);
4033 /* Access is about to become group representative unless we find some
4034 nasty overlap which would preclude us from breaking this parameter
4035 apart. */
4037 j = i + 1;
4038 while (j < access_count)
4040 struct access *ac2 = (*access_vec)[j];
4041 if (ac2->offset != access->offset)
4043 /* All or nothing law for parameters. */
4044 if (access->offset + access->size > ac2->offset)
4045 return NULL;
4046 else
4047 break;
4049 else if (ac2->size != access->size)
4050 return NULL;
4052 if (access_precludes_ipa_sra_p (ac2, TYPE_ALIGN (access->type))
4053 || (ac2->type != access->type
4054 && (TREE_ADDRESSABLE (ac2->type)
4055 || TREE_ADDRESSABLE (access->type)))
4056 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
4057 return NULL;
4059 modification |= ac2->write;
4060 ac2->group_representative = access;
4061 ac2->next_sibling = access->next_sibling;
4062 access->next_sibling = ac2;
4063 j++;
4066 group_count++;
4067 access->grp_maybe_modified = modification;
4068 if (!modification)
4069 *ro_grp = true;
4070 *prev_acc_ptr = access;
4071 prev_acc_ptr = &access->next_grp;
4072 total_size += access->size;
4073 i = j;
4076 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4077 agg_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))));
4078 else
4079 agg_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (parm)));
4080 if (total_size >= agg_size)
4081 return NULL;
4083 gcc_assert (group_count > 0);
4084 return res;
4087 /* Decide whether parameters with representative accesses given by REPR should
4088 be reduced into components. */
4090 static int
4091 decide_one_param_reduction (struct access *repr)
4093 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
4094 bool by_ref;
4095 tree parm;
4097 parm = repr->base;
4098 cur_parm_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (parm)));
4099 gcc_assert (cur_parm_size > 0);
4101 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4103 by_ref = true;
4104 agg_size = tree_to_uhwi (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))));
4106 else
4108 by_ref = false;
4109 agg_size = cur_parm_size;
4112 if (dump_file)
4114 struct access *acc;
4115 fprintf (dump_file, "Evaluating PARAM group sizes for ");
4116 print_generic_expr (dump_file, parm, 0);
4117 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
4118 for (acc = repr; acc; acc = acc->next_grp)
4119 dump_access (dump_file, acc, true);
4122 total_size = 0;
4123 new_param_count = 0;
4125 for (; repr; repr = repr->next_grp)
4127 gcc_assert (parm == repr->base);
4129 /* Taking the address of a non-addressable field is verboten. */
4130 if (by_ref && repr->non_addressable)
4131 return 0;
4133 /* Do not decompose a non-BLKmode param in a way that would
4134 create BLKmode params. Especially for by-reference passing
4135 (thus, pointer-type param) this is hardly worthwhile. */
4136 if (DECL_MODE (parm) != BLKmode
4137 && TYPE_MODE (repr->type) == BLKmode)
4138 return 0;
4140 if (!by_ref || (!repr->grp_maybe_modified
4141 && !repr->grp_not_necessarilly_dereferenced))
4142 total_size += repr->size;
4143 else
4144 total_size += cur_parm_size;
4146 new_param_count++;
4149 gcc_assert (new_param_count > 0);
4151 if (optimize_function_for_size_p (cfun))
4152 parm_size_limit = cur_parm_size;
4153 else
4154 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
4155 * cur_parm_size);
4157 if (total_size < agg_size
4158 && total_size <= parm_size_limit)
4160 if (dump_file)
4161 fprintf (dump_file, " ....will be split into %i components\n",
4162 new_param_count);
4163 return new_param_count;
4165 else
4166 return 0;
4169 /* The order of the following enums is important, we need to do extra work for
4170 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
4171 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
4172 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
4174 /* Identify representatives of all accesses to all candidate parameters for
4175 IPA-SRA. Return result based on what representatives have been found. */
4177 static enum ipa_splicing_result
4178 splice_all_param_accesses (vec<access_p> &representatives)
4180 enum ipa_splicing_result result = NO_GOOD_ACCESS;
4181 tree parm;
4182 struct access *repr;
4184 representatives.create (func_param_count);
4186 for (parm = DECL_ARGUMENTS (current_function_decl);
4187 parm;
4188 parm = DECL_CHAIN (parm))
4190 if (is_unused_scalar_param (parm))
4192 representatives.quick_push (&no_accesses_representant);
4193 if (result == NO_GOOD_ACCESS)
4194 result = UNUSED_PARAMS;
4196 else if (POINTER_TYPE_P (TREE_TYPE (parm))
4197 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
4198 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4200 repr = unmodified_by_ref_scalar_representative (parm);
4201 representatives.quick_push (repr);
4202 if (repr)
4203 result = UNMODIF_BY_REF_ACCESSES;
4205 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4207 bool ro_grp = false;
4208 repr = splice_param_accesses (parm, &ro_grp);
4209 representatives.quick_push (repr);
4211 if (repr && !no_accesses_p (repr))
4213 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4215 if (ro_grp)
4216 result = UNMODIF_BY_REF_ACCESSES;
4217 else if (result < MODIF_BY_REF_ACCESSES)
4218 result = MODIF_BY_REF_ACCESSES;
4220 else if (result < BY_VAL_ACCESSES)
4221 result = BY_VAL_ACCESSES;
4223 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
4224 result = UNUSED_PARAMS;
4226 else
4227 representatives.quick_push (NULL);
4230 if (result == NO_GOOD_ACCESS)
4232 representatives.release ();
4233 return NO_GOOD_ACCESS;
4236 return result;
4239 /* Return the index of BASE in PARMS. Abort if it is not found. */
4241 static inline int
4242 get_param_index (tree base, vec<tree> parms)
4244 int i, len;
4246 len = parms.length ();
4247 for (i = 0; i < len; i++)
4248 if (parms[i] == base)
4249 return i;
4250 gcc_unreachable ();
4253 /* Convert the decisions made at the representative level into compact
4254 parameter adjustments. REPRESENTATIVES are pointers to first
4255 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4256 final number of adjustments. */
4258 static ipa_parm_adjustment_vec
4259 turn_representatives_into_adjustments (vec<access_p> representatives,
4260 int adjustments_count)
4262 vec<tree> parms;
4263 ipa_parm_adjustment_vec adjustments;
4264 tree parm;
4265 int i;
4267 gcc_assert (adjustments_count > 0);
4268 parms = ipa_get_vector_of_formal_parms (current_function_decl);
4269 adjustments.create (adjustments_count);
4270 parm = DECL_ARGUMENTS (current_function_decl);
4271 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
4273 struct access *repr = representatives[i];
4275 if (!repr || no_accesses_p (repr))
4277 struct ipa_parm_adjustment adj;
4279 memset (&adj, 0, sizeof (adj));
4280 adj.base_index = get_param_index (parm, parms);
4281 adj.base = parm;
4282 if (!repr)
4283 adj.op = IPA_PARM_OP_COPY;
4284 else
4285 adj.op = IPA_PARM_OP_REMOVE;
4286 adj.arg_prefix = "ISRA";
4287 adjustments.quick_push (adj);
4289 else
4291 struct ipa_parm_adjustment adj;
4292 int index = get_param_index (parm, parms);
4294 for (; repr; repr = repr->next_grp)
4296 memset (&adj, 0, sizeof (adj));
4297 gcc_assert (repr->base == parm);
4298 adj.base_index = index;
4299 adj.base = repr->base;
4300 adj.type = repr->type;
4301 adj.alias_ptr_type = reference_alias_ptr_type (repr->expr);
4302 adj.offset = repr->offset;
4303 adj.by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
4304 && (repr->grp_maybe_modified
4305 || repr->grp_not_necessarilly_dereferenced));
4306 adj.arg_prefix = "ISRA";
4307 adjustments.quick_push (adj);
4311 parms.release ();
4312 return adjustments;
4315 /* Analyze the collected accesses and produce a plan what to do with the
4316 parameters in the form of adjustments, NULL meaning nothing. */
4318 static ipa_parm_adjustment_vec
4319 analyze_all_param_acesses (void)
4321 enum ipa_splicing_result repr_state;
4322 bool proceed = false;
4323 int i, adjustments_count = 0;
4324 vec<access_p> representatives;
4325 ipa_parm_adjustment_vec adjustments;
4327 repr_state = splice_all_param_accesses (representatives);
4328 if (repr_state == NO_GOOD_ACCESS)
4329 return ipa_parm_adjustment_vec ();
4331 /* If there are any parameters passed by reference which are not modified
4332 directly, we need to check whether they can be modified indirectly. */
4333 if (repr_state == UNMODIF_BY_REF_ACCESSES)
4335 analyze_caller_dereference_legality (representatives);
4336 analyze_modified_params (representatives);
4339 for (i = 0; i < func_param_count; i++)
4341 struct access *repr = representatives[i];
4343 if (repr && !no_accesses_p (repr))
4345 if (repr->grp_scalar_ptr)
4347 adjustments_count++;
4348 if (repr->grp_not_necessarilly_dereferenced
4349 || repr->grp_maybe_modified)
4350 representatives[i] = NULL;
4351 else
4353 proceed = true;
4354 sra_stats.scalar_by_ref_to_by_val++;
4357 else
4359 int new_components = decide_one_param_reduction (repr);
4361 if (new_components == 0)
4363 representatives[i] = NULL;
4364 adjustments_count++;
4366 else
4368 adjustments_count += new_components;
4369 sra_stats.aggregate_params_reduced++;
4370 sra_stats.param_reductions_created += new_components;
4371 proceed = true;
4375 else
4377 if (no_accesses_p (repr))
4379 proceed = true;
4380 sra_stats.deleted_unused_parameters++;
4382 adjustments_count++;
4386 if (!proceed && dump_file)
4387 fprintf (dump_file, "NOT proceeding to change params.\n");
4389 if (proceed)
4390 adjustments = turn_representatives_into_adjustments (representatives,
4391 adjustments_count);
4392 else
4393 adjustments = ipa_parm_adjustment_vec ();
4395 representatives.release ();
4396 return adjustments;
4399 /* If a parameter replacement identified by ADJ does not yet exist in the form
4400 of declaration, create it and record it, otherwise return the previously
4401 created one. */
4403 static tree
4404 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4406 tree repl;
4407 if (!adj->new_ssa_base)
4409 char *pretty_name = make_fancy_name (adj->base);
4411 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4412 DECL_NAME (repl) = get_identifier (pretty_name);
4413 obstack_free (&name_obstack, pretty_name);
4415 adj->new_ssa_base = repl;
4417 else
4418 repl = adj->new_ssa_base;
4419 return repl;
4422 /* Find the first adjustment for a particular parameter BASE in a vector of
4423 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4424 adjustment. */
4426 static struct ipa_parm_adjustment *
4427 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4429 int i, len;
4431 len = adjustments.length ();
4432 for (i = 0; i < len; i++)
4434 struct ipa_parm_adjustment *adj;
4436 adj = &adjustments[i];
4437 if (adj->op != IPA_PARM_OP_COPY && adj->base == base)
4438 return adj;
4441 return NULL;
4444 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4445 removed because its value is not used, replace the SSA_NAME with a one
4446 relating to a created VAR_DECL together all of its uses and return true.
4447 ADJUSTMENTS is a pointer to an adjustments vector. */
4449 static bool
4450 replace_removed_params_ssa_names (gimple stmt,
4451 ipa_parm_adjustment_vec adjustments)
4453 struct ipa_parm_adjustment *adj;
4454 tree lhs, decl, repl, name;
4456 if (gimple_code (stmt) == GIMPLE_PHI)
4457 lhs = gimple_phi_result (stmt);
4458 else if (is_gimple_assign (stmt))
4459 lhs = gimple_assign_lhs (stmt);
4460 else if (is_gimple_call (stmt))
4461 lhs = gimple_call_lhs (stmt);
4462 else
4463 gcc_unreachable ();
4465 if (TREE_CODE (lhs) != SSA_NAME)
4466 return false;
4468 decl = SSA_NAME_VAR (lhs);
4469 if (decl == NULL_TREE
4470 || TREE_CODE (decl) != PARM_DECL)
4471 return false;
4473 adj = get_adjustment_for_base (adjustments, decl);
4474 if (!adj)
4475 return false;
4477 repl = get_replaced_param_substitute (adj);
4478 name = make_ssa_name (repl, stmt);
4480 if (dump_file)
4482 fprintf (dump_file, "replacing an SSA name of a removed param ");
4483 print_generic_expr (dump_file, lhs, 0);
4484 fprintf (dump_file, " with ");
4485 print_generic_expr (dump_file, name, 0);
4486 fprintf (dump_file, "\n");
4489 if (is_gimple_assign (stmt))
4490 gimple_assign_set_lhs (stmt, name);
4491 else if (is_gimple_call (stmt))
4492 gimple_call_set_lhs (stmt, name);
4493 else
4494 gimple_phi_set_result (stmt, name);
4496 replace_uses_by (lhs, name);
4497 release_ssa_name (lhs);
4498 return true;
4501 /* If the statement pointed to by STMT_PTR contains any expressions that need
4502 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4503 potential type incompatibilities (GSI is used to accommodate conversion
4504 statements and must point to the statement). Return true iff the statement
4505 was modified. */
4507 static bool
4508 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4509 ipa_parm_adjustment_vec adjustments)
4511 gimple stmt = *stmt_ptr;
4512 tree *lhs_p, *rhs_p;
4513 bool any;
4515 if (!gimple_assign_single_p (stmt))
4516 return false;
4518 rhs_p = gimple_assign_rhs1_ptr (stmt);
4519 lhs_p = gimple_assign_lhs_ptr (stmt);
4521 any = ipa_modify_expr (rhs_p, false, adjustments);
4522 any |= ipa_modify_expr (lhs_p, false, adjustments);
4523 if (any)
4525 tree new_rhs = NULL_TREE;
4527 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4529 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4531 /* V_C_Es of constructors can cause trouble (PR 42714). */
4532 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4533 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4534 else
4535 *rhs_p = build_constructor (TREE_TYPE (*lhs_p),
4536 NULL);
4538 else
4539 new_rhs = fold_build1_loc (gimple_location (stmt),
4540 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4541 *rhs_p);
4543 else if (REFERENCE_CLASS_P (*rhs_p)
4544 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4545 && !is_gimple_reg (*lhs_p))
4546 /* This can happen when an assignment in between two single field
4547 structures is turned into an assignment in between two pointers to
4548 scalars (PR 42237). */
4549 new_rhs = *rhs_p;
4551 if (new_rhs)
4553 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4554 true, GSI_SAME_STMT);
4556 gimple_assign_set_rhs_from_tree (gsi, tmp);
4559 return true;
4562 return false;
4565 /* Traverse the function body and all modifications as described in
4566 ADJUSTMENTS. Return true iff the CFG has been changed. */
4568 bool
4569 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4571 bool cfg_changed = false;
4572 basic_block bb;
4574 FOR_EACH_BB_FN (bb, cfun)
4576 gimple_stmt_iterator gsi;
4578 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4579 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4581 gsi = gsi_start_bb (bb);
4582 while (!gsi_end_p (gsi))
4584 gimple stmt = gsi_stmt (gsi);
4585 bool modified = false;
4586 tree *t;
4587 unsigned i;
4589 switch (gimple_code (stmt))
4591 case GIMPLE_RETURN:
4592 t = gimple_return_retval_ptr (stmt);
4593 if (*t != NULL_TREE)
4594 modified |= ipa_modify_expr (t, true, adjustments);
4595 break;
4597 case GIMPLE_ASSIGN:
4598 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4599 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4600 break;
4602 case GIMPLE_CALL:
4603 /* Operands must be processed before the lhs. */
4604 for (i = 0; i < gimple_call_num_args (stmt); i++)
4606 t = gimple_call_arg_ptr (stmt, i);
4607 modified |= ipa_modify_expr (t, true, adjustments);
4610 if (gimple_call_lhs (stmt))
4612 t = gimple_call_lhs_ptr (stmt);
4613 modified |= ipa_modify_expr (t, false, adjustments);
4614 modified |= replace_removed_params_ssa_names (stmt,
4615 adjustments);
4617 break;
4619 case GIMPLE_ASM:
4620 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4622 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4623 modified |= ipa_modify_expr (t, true, adjustments);
4625 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4627 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4628 modified |= ipa_modify_expr (t, false, adjustments);
4630 break;
4632 default:
4633 break;
4636 if (modified)
4638 update_stmt (stmt);
4639 if (maybe_clean_eh_stmt (stmt)
4640 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4641 cfg_changed = true;
4643 gsi_next (&gsi);
4647 return cfg_changed;
4650 /* Call gimple_debug_bind_reset_value on all debug statements describing
4651 gimple register parameters that are being removed or replaced. */
4653 static void
4654 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4656 int i, len;
4657 gimple_stmt_iterator *gsip = NULL, gsi;
4659 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR_FOR_FN (cfun)))
4661 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR_FOR_FN (cfun)));
4662 gsip = &gsi;
4664 len = adjustments.length ();
4665 for (i = 0; i < len; i++)
4667 struct ipa_parm_adjustment *adj;
4668 imm_use_iterator ui;
4669 gimple stmt, def_temp;
4670 tree name, vexpr, copy = NULL_TREE;
4671 use_operand_p use_p;
4673 adj = &adjustments[i];
4674 if (adj->op == IPA_PARM_OP_COPY || !is_gimple_reg (adj->base))
4675 continue;
4676 name = ssa_default_def (cfun, adj->base);
4677 vexpr = NULL;
4678 if (name)
4679 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4681 if (gimple_clobber_p (stmt))
4683 gimple_stmt_iterator cgsi = gsi_for_stmt (stmt);
4684 unlink_stmt_vdef (stmt);
4685 gsi_remove (&cgsi, true);
4686 release_defs (stmt);
4687 continue;
4689 /* All other users must have been removed by
4690 ipa_sra_modify_function_body. */
4691 gcc_assert (is_gimple_debug (stmt));
4692 if (vexpr == NULL && gsip != NULL)
4694 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4695 vexpr = make_node (DEBUG_EXPR_DECL);
4696 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4697 NULL);
4698 DECL_ARTIFICIAL (vexpr) = 1;
4699 TREE_TYPE (vexpr) = TREE_TYPE (name);
4700 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4701 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4703 if (vexpr)
4705 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4706 SET_USE (use_p, vexpr);
4708 else
4709 gimple_debug_bind_reset_value (stmt);
4710 update_stmt (stmt);
4712 /* Create a VAR_DECL for debug info purposes. */
4713 if (!DECL_IGNORED_P (adj->base))
4715 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4716 VAR_DECL, DECL_NAME (adj->base),
4717 TREE_TYPE (adj->base));
4718 if (DECL_PT_UID_SET_P (adj->base))
4719 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4720 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4721 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4722 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4723 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4724 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4725 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4726 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4727 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4728 SET_DECL_RTL (copy, 0);
4729 TREE_USED (copy) = 1;
4730 DECL_CONTEXT (copy) = current_function_decl;
4731 add_local_decl (cfun, copy);
4732 DECL_CHAIN (copy) =
4733 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4734 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4736 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4738 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4739 if (vexpr)
4740 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4741 else
4742 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4743 NULL);
4744 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4749 /* Return false iff all callers have at least as many actual arguments as there
4750 are formal parameters in the current function. */
4752 static bool
4753 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4754 void *data ATTRIBUTE_UNUSED)
4756 struct cgraph_edge *cs;
4757 for (cs = node->callers; cs; cs = cs->next_caller)
4758 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4759 return true;
4761 return false;
4764 /* Convert all callers of NODE. */
4766 static bool
4767 convert_callers_for_node (struct cgraph_node *node,
4768 void *data)
4770 ipa_parm_adjustment_vec *adjustments = (ipa_parm_adjustment_vec *) data;
4771 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4772 struct cgraph_edge *cs;
4774 for (cs = node->callers; cs; cs = cs->next_caller)
4776 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4778 if (dump_file)
4779 fprintf (dump_file, "Adjusting call %s/%i -> %s/%i\n",
4780 xstrdup (cs->caller->name ()),
4781 cs->caller->order,
4782 xstrdup (cs->callee->name ()),
4783 cs->callee->order);
4785 ipa_modify_call_arguments (cs, cs->call_stmt, *adjustments);
4787 pop_cfun ();
4790 for (cs = node->callers; cs; cs = cs->next_caller)
4791 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4792 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4793 compute_inline_parameters (cs->caller, true);
4794 BITMAP_FREE (recomputed_callers);
4796 return true;
4799 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4801 static void
4802 convert_callers (struct cgraph_node *node, tree old_decl,
4803 ipa_parm_adjustment_vec adjustments)
4805 basic_block this_block;
4807 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4808 &adjustments, false);
4810 if (!encountered_recursive_call)
4811 return;
4813 FOR_EACH_BB_FN (this_block, cfun)
4815 gimple_stmt_iterator gsi;
4817 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4819 gimple stmt = gsi_stmt (gsi);
4820 tree call_fndecl;
4821 if (gimple_code (stmt) != GIMPLE_CALL)
4822 continue;
4823 call_fndecl = gimple_call_fndecl (stmt);
4824 if (call_fndecl == old_decl)
4826 if (dump_file)
4827 fprintf (dump_file, "Adjusting recursive call");
4828 gimple_call_set_fndecl (stmt, node->decl);
4829 ipa_modify_call_arguments (NULL, stmt, adjustments);
4834 return;
4837 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4838 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4840 static bool
4841 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4843 struct cgraph_node *new_node;
4844 bool cfg_changed;
4845 vec<cgraph_edge_p> redirect_callers = collect_callers_of_node (node);
4847 rebuild_cgraph_edges ();
4848 free_dominance_info (CDI_DOMINATORS);
4849 pop_cfun ();
4851 new_node = cgraph_function_versioning (node, redirect_callers,
4852 NULL,
4853 NULL, false, NULL, NULL, "isra");
4854 redirect_callers.release ();
4856 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4857 ipa_modify_formal_parameters (current_function_decl, adjustments);
4858 cfg_changed = ipa_sra_modify_function_body (adjustments);
4859 sra_ipa_reset_debug_stmts (adjustments);
4860 convert_callers (new_node, node->decl, adjustments);
4861 cgraph_make_node_local (new_node);
4862 return cfg_changed;
4865 /* If NODE has a caller, return true. */
4867 static bool
4868 has_caller_p (struct cgraph_node *node, void *data ATTRIBUTE_UNUSED)
4870 if (node->callers)
4871 return true;
4872 return false;
4875 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4876 attributes, return true otherwise. NODE is the cgraph node of the current
4877 function. */
4879 static bool
4880 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4882 if (!cgraph_node_can_be_local_p (node))
4884 if (dump_file)
4885 fprintf (dump_file, "Function not local to this compilation unit.\n");
4886 return false;
4889 if (!node->local.can_change_signature)
4891 if (dump_file)
4892 fprintf (dump_file, "Function can not change signature.\n");
4893 return false;
4896 if (!tree_versionable_function_p (node->decl))
4898 if (dump_file)
4899 fprintf (dump_file, "Function is not versionable.\n");
4900 return false;
4903 if (DECL_VIRTUAL_P (current_function_decl))
4905 if (dump_file)
4906 fprintf (dump_file, "Function is a virtual method.\n");
4907 return false;
4910 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4911 && inline_summary (node)->size >= MAX_INLINE_INSNS_AUTO)
4913 if (dump_file)
4914 fprintf (dump_file, "Function too big to be made truly local.\n");
4915 return false;
4918 if (!cgraph_for_node_and_aliases (node, has_caller_p, NULL, true))
4920 if (dump_file)
4921 fprintf (dump_file,
4922 "Function has no callers in this compilation unit.\n");
4923 return false;
4926 if (cfun->stdarg)
4928 if (dump_file)
4929 fprintf (dump_file, "Function uses stdarg. \n");
4930 return false;
4933 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4934 return false;
4936 return true;
4939 /* Perform early interprocedural SRA. */
4941 static unsigned int
4942 ipa_early_sra (void)
4944 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4945 ipa_parm_adjustment_vec adjustments;
4946 int ret = 0;
4948 if (!ipa_sra_preliminary_function_checks (node))
4949 return 0;
4951 sra_initialize ();
4952 sra_mode = SRA_MODE_EARLY_IPA;
4954 if (!find_param_candidates ())
4956 if (dump_file)
4957 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4958 goto simple_out;
4961 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4962 NULL, true))
4964 if (dump_file)
4965 fprintf (dump_file, "There are callers with insufficient number of "
4966 "arguments.\n");
4967 goto simple_out;
4970 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4971 func_param_count
4972 * last_basic_block_for_fn (cfun));
4973 final_bbs = BITMAP_ALLOC (NULL);
4975 scan_function ();
4976 if (encountered_apply_args)
4978 if (dump_file)
4979 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4980 goto out;
4983 if (encountered_unchangable_recursive_call)
4985 if (dump_file)
4986 fprintf (dump_file, "Function calls itself with insufficient "
4987 "number of arguments.\n");
4988 goto out;
4991 adjustments = analyze_all_param_acesses ();
4992 if (!adjustments.exists ())
4993 goto out;
4994 if (dump_file)
4995 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4997 if (modify_function (node, adjustments))
4998 ret = TODO_update_ssa | TODO_cleanup_cfg;
4999 else
5000 ret = TODO_update_ssa;
5001 adjustments.release ();
5003 statistics_counter_event (cfun, "Unused parameters deleted",
5004 sra_stats.deleted_unused_parameters);
5005 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
5006 sra_stats.scalar_by_ref_to_by_val);
5007 statistics_counter_event (cfun, "Aggregate parameters broken up",
5008 sra_stats.aggregate_params_reduced);
5009 statistics_counter_event (cfun, "Aggregate parameter components created",
5010 sra_stats.param_reductions_created);
5012 out:
5013 BITMAP_FREE (final_bbs);
5014 free (bb_dereferences);
5015 simple_out:
5016 sra_deinitialize ();
5017 return ret;
5020 /* Return if early ipa sra shall be performed. */
5021 static bool
5022 ipa_early_sra_gate (void)
5024 return flag_ipa_sra && dbg_cnt (eipa_sra);
5027 namespace {
5029 const pass_data pass_data_early_ipa_sra =
5031 GIMPLE_PASS, /* type */
5032 "eipa_sra", /* name */
5033 OPTGROUP_NONE, /* optinfo_flags */
5034 true, /* has_gate */
5035 true, /* has_execute */
5036 TV_IPA_SRA, /* tv_id */
5037 0, /* properties_required */
5038 0, /* properties_provided */
5039 0, /* properties_destroyed */
5040 0, /* todo_flags_start */
5041 TODO_dump_symtab, /* todo_flags_finish */
5044 class pass_early_ipa_sra : public gimple_opt_pass
5046 public:
5047 pass_early_ipa_sra (gcc::context *ctxt)
5048 : gimple_opt_pass (pass_data_early_ipa_sra, ctxt)
5051 /* opt_pass methods: */
5052 bool gate () { return ipa_early_sra_gate (); }
5053 unsigned int execute () { return ipa_early_sra (); }
5055 }; // class pass_early_ipa_sra
5057 } // anon namespace
5059 gimple_opt_pass *
5060 make_pass_early_ipa_sra (gcc::context *ctxt)
5062 return new pass_early_ipa_sra (ctxt);