Merge trunk version 194962 into gupc branch.
[official-gcc.git] / gcc / tree-sra.c
blob6c391a587540927fe94df9ad0a6246c8b90739da
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, 2009, 2010, 2011, 2012 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 "alloc-pool.h"
78 #include "tm.h"
79 #include "tree.h"
80 #include "gimple.h"
81 #include "cgraph.h"
82 #include "tree-flow.h"
83 #include "tree-pass.h"
84 #include "ipa-prop.h"
85 #include "statistics.h"
86 #include "params.h"
87 #include "target.h"
88 #include "flags.h"
89 #include "dbgcnt.h"
90 #include "tree-inline.h"
91 #include "gimple-pretty-print.h"
92 #include "ipa-inline.h"
94 /* Enumeration of all aggregate reductions we can do. */
95 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
96 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
97 SRA_MODE_INTRA }; /* late intraprocedural SRA */
99 /* Global variable describing which aggregate reduction we are performing at
100 the moment. */
101 static enum sra_mode sra_mode;
103 struct assign_link;
105 /* ACCESS represents each access to an aggregate variable (as a whole or a
106 part). It can also represent a group of accesses that refer to exactly the
107 same fragment of an aggregate (i.e. those that have exactly the same offset
108 and size). Such representatives for a single aggregate, once determined,
109 are linked in a linked list and have the group fields set.
111 Moreover, when doing intraprocedural SRA, a tree is built from those
112 representatives (by the means of first_child and next_sibling pointers), in
113 which all items in a subtree are "within" the root, i.e. their offset is
114 greater or equal to offset of the root and offset+size is smaller or equal
115 to offset+size of the root. Children of an access are sorted by offset.
117 Note that accesses to parts of vector and complex number types always
118 represented by an access to the whole complex number or a vector. It is a
119 duty of the modifying functions to replace them appropriately. */
121 struct access
123 /* Values returned by `get_ref_base_and_extent' for each component reference
124 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
125 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
126 HOST_WIDE_INT offset;
127 HOST_WIDE_INT size;
128 tree base;
130 /* Expression. It is context dependent so do not use it to create new
131 expressions to access the original aggregate. See PR 42154 for a
132 testcase. */
133 tree expr;
134 /* Type. */
135 tree type;
137 /* The statement this access belongs to. */
138 gimple stmt;
140 /* Next group representative for this aggregate. */
141 struct access *next_grp;
143 /* Pointer to the group representative. Pointer to itself if the struct is
144 the representative. */
145 struct access *group_representative;
147 /* If this access has any children (in terms of the definition above), this
148 points to the first one. */
149 struct access *first_child;
151 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
152 described above. In IPA-SRA this is a pointer to the next access
153 belonging to the same group (having the same representative). */
154 struct access *next_sibling;
156 /* Pointers to the first and last element in the linked list of assign
157 links. */
158 struct assign_link *first_link, *last_link;
160 /* Pointer to the next access in the work queue. */
161 struct access *next_queued;
163 /* Replacement variable for this access "region." Never to be accessed
164 directly, always only by the means of get_access_replacement() and only
165 when grp_to_be_replaced flag is set. */
166 tree replacement_decl;
168 /* Is this particular access write access? */
169 unsigned write : 1;
171 /* Is this access an access to a non-addressable field? */
172 unsigned non_addressable : 1;
174 /* Is this access currently in the work queue? */
175 unsigned grp_queued : 1;
177 /* Does this group contain a write access? This flag is propagated down the
178 access tree. */
179 unsigned grp_write : 1;
181 /* Does this group contain a read access? This flag is propagated down the
182 access tree. */
183 unsigned grp_read : 1;
185 /* Does this group contain a read access that comes from an assignment
186 statement? This flag is propagated down the access tree. */
187 unsigned grp_assignment_read : 1;
189 /* Does this group contain a write access that comes from an assignment
190 statement? This flag is propagated down the access tree. */
191 unsigned grp_assignment_write : 1;
193 /* Does this group contain a read access through a scalar type? This flag is
194 not propagated in the access tree in any direction. */
195 unsigned grp_scalar_read : 1;
197 /* Does this group contain a write access through a scalar type? This flag
198 is not propagated in the access tree in any direction. */
199 unsigned grp_scalar_write : 1;
201 /* Is this access an artificial one created to scalarize some record
202 entirely? */
203 unsigned grp_total_scalarization : 1;
205 /* Other passes of the analysis use this bit to make function
206 analyze_access_subtree create scalar replacements for this group if
207 possible. */
208 unsigned grp_hint : 1;
210 /* Is the subtree rooted in this access fully covered by scalar
211 replacements? */
212 unsigned grp_covered : 1;
214 /* If set to true, this access and all below it in an access tree must not be
215 scalarized. */
216 unsigned grp_unscalarizable_region : 1;
218 /* Whether data have been written to parts of the aggregate covered by this
219 access which is not to be scalarized. This flag is propagated up in the
220 access tree. */
221 unsigned grp_unscalarized_data : 1;
223 /* Does this access and/or group contain a write access through a
224 BIT_FIELD_REF? */
225 unsigned grp_partial_lhs : 1;
227 /* Set when a scalar replacement should be created for this variable. */
228 unsigned grp_to_be_replaced : 1;
230 /* Set when we want a replacement for the sole purpose of having it in
231 generated debug statements. */
232 unsigned grp_to_be_debug_replaced : 1;
234 /* Should TREE_NO_WARNING of a replacement be set? */
235 unsigned grp_no_warning : 1;
237 /* Is it possible that the group refers to data which might be (directly or
238 otherwise) modified? */
239 unsigned grp_maybe_modified : 1;
241 /* Set when this is a representative of a pointer to scalar (i.e. by
242 reference) parameter which we consider for turning into a plain scalar
243 (i.e. a by value parameter). */
244 unsigned grp_scalar_ptr : 1;
246 /* Set when we discover that this pointer is not safe to dereference in the
247 caller. */
248 unsigned grp_not_necessarilly_dereferenced : 1;
251 typedef struct access *access_p;
254 /* Alloc pool for allocating access structures. */
255 static alloc_pool access_pool;
257 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
258 are used to propagate subaccesses from rhs to lhs as long as they don't
259 conflict with what is already there. */
260 struct assign_link
262 struct access *lacc, *racc;
263 struct assign_link *next;
266 /* Alloc pool for allocating assign link structures. */
267 static alloc_pool link_pool;
269 /* Base (tree) -> Vector (vec<access_p> *) map. */
270 static struct pointer_map_t *base_access_vec;
272 /* Set of candidates. */
273 static bitmap candidate_bitmap;
274 static htab_t candidates;
276 /* For a candidate UID return the candidates decl. */
278 static inline tree
279 candidate (unsigned uid)
281 struct tree_decl_minimal t;
282 t.uid = uid;
283 return (tree) htab_find_with_hash (candidates, &t, uid);
286 /* Bitmap of candidates which we should try to entirely scalarize away and
287 those which cannot be (because they are and need be used as a whole). */
288 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
290 /* Obstack for creation of fancy names. */
291 static struct obstack name_obstack;
293 /* Head of a linked list of accesses that need to have its subaccesses
294 propagated to their assignment counterparts. */
295 static struct access *work_queue_head;
297 /* Number of parameters of the analyzed function when doing early ipa SRA. */
298 static int func_param_count;
300 /* scan_function sets the following to true if it encounters a call to
301 __builtin_apply_args. */
302 static bool encountered_apply_args;
304 /* Set by scan_function when it finds a recursive call. */
305 static bool encountered_recursive_call;
307 /* Set by scan_function when it finds a recursive call with less actual
308 arguments than formal parameters.. */
309 static bool encountered_unchangable_recursive_call;
311 /* This is a table in which for each basic block and parameter there is a
312 distance (offset + size) in that parameter which is dereferenced and
313 accessed in that BB. */
314 static HOST_WIDE_INT *bb_dereferences;
315 /* Bitmap of BBs that can cause the function to "stop" progressing by
316 returning, throwing externally, looping infinitely or calling a function
317 which might abort etc.. */
318 static bitmap final_bbs;
320 /* Representative of no accesses at all. */
321 static struct access no_accesses_representant;
323 /* Predicate to test the special value. */
325 static inline bool
326 no_accesses_p (struct access *access)
328 return access == &no_accesses_representant;
331 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
332 representative fields are dumped, otherwise those which only describe the
333 individual access are. */
335 static struct
337 /* Number of processed aggregates is readily available in
338 analyze_all_variable_accesses and so is not stored here. */
340 /* Number of created scalar replacements. */
341 int replacements;
343 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
344 expression. */
345 int exprs;
347 /* Number of statements created by generate_subtree_copies. */
348 int subtree_copies;
350 /* Number of statements created by load_assign_lhs_subreplacements. */
351 int subreplacements;
353 /* Number of times sra_modify_assign has deleted a statement. */
354 int deleted;
356 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
357 RHS reparately due to type conversions or nonexistent matching
358 references. */
359 int separate_lhs_rhs_handling;
361 /* Number of parameters that were removed because they were unused. */
362 int deleted_unused_parameters;
364 /* Number of scalars passed as parameters by reference that have been
365 converted to be passed by value. */
366 int scalar_by_ref_to_by_val;
368 /* Number of aggregate parameters that were replaced by one or more of their
369 components. */
370 int aggregate_params_reduced;
372 /* Numbber of components created when splitting aggregate parameters. */
373 int param_reductions_created;
374 } sra_stats;
376 static void
377 dump_access (FILE *f, struct access *access, bool grp)
379 fprintf (f, "access { ");
380 fprintf (f, "base = (%d)'", DECL_UID (access->base));
381 print_generic_expr (f, access->base, 0);
382 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
383 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
384 fprintf (f, ", expr = ");
385 print_generic_expr (f, access->expr, 0);
386 fprintf (f, ", type = ");
387 print_generic_expr (f, access->type, 0);
388 if (grp)
389 fprintf (f, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
390 "grp_assignment_write = %d, grp_scalar_read = %d, "
391 "grp_scalar_write = %d, grp_total_scalarization = %d, "
392 "grp_hint = %d, grp_covered = %d, "
393 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
394 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
395 "grp_to_be_debug_replaced = %d, grp_maybe_modified = %d, "
396 "grp_not_necessarilly_dereferenced = %d\n",
397 access->grp_read, access->grp_write, access->grp_assignment_read,
398 access->grp_assignment_write, access->grp_scalar_read,
399 access->grp_scalar_write, access->grp_total_scalarization,
400 access->grp_hint, access->grp_covered,
401 access->grp_unscalarizable_region, access->grp_unscalarized_data,
402 access->grp_partial_lhs, access->grp_to_be_replaced,
403 access->grp_to_be_debug_replaced, access->grp_maybe_modified,
404 access->grp_not_necessarilly_dereferenced);
405 else
406 fprintf (f, ", write = %d, grp_total_scalarization = %d, "
407 "grp_partial_lhs = %d\n",
408 access->write, access->grp_total_scalarization,
409 access->grp_partial_lhs);
412 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
414 static void
415 dump_access_tree_1 (FILE *f, struct access *access, int level)
419 int i;
421 for (i = 0; i < level; i++)
422 fputs ("* ", dump_file);
424 dump_access (f, access, true);
426 if (access->first_child)
427 dump_access_tree_1 (f, access->first_child, level + 1);
429 access = access->next_sibling;
431 while (access);
434 /* Dump all access trees for a variable, given the pointer to the first root in
435 ACCESS. */
437 static void
438 dump_access_tree (FILE *f, struct access *access)
440 for (; access; access = access->next_grp)
441 dump_access_tree_1 (f, access, 0);
444 /* Return true iff ACC is non-NULL and has subaccesses. */
446 static inline bool
447 access_has_children_p (struct access *acc)
449 return acc && acc->first_child;
452 /* Return true iff ACC is (partly) covered by at least one replacement. */
454 static bool
455 access_has_replacements_p (struct access *acc)
457 struct access *child;
458 if (acc->grp_to_be_replaced)
459 return true;
460 for (child = acc->first_child; child; child = child->next_sibling)
461 if (access_has_replacements_p (child))
462 return true;
463 return false;
466 /* Return a vector of pointers to accesses for the variable given in BASE or
467 NULL if there is none. */
469 static vec<access_p> *
470 get_base_access_vector (tree base)
472 void **slot;
474 slot = pointer_map_contains (base_access_vec, base);
475 if (!slot)
476 return NULL;
477 else
478 return *(vec<access_p> **) slot;
481 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
482 in ACCESS. Return NULL if it cannot be found. */
484 static struct access *
485 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
486 HOST_WIDE_INT size)
488 while (access && (access->offset != offset || access->size != size))
490 struct access *child = access->first_child;
492 while (child && (child->offset + child->size <= offset))
493 child = child->next_sibling;
494 access = child;
497 return access;
500 /* Return the first group representative for DECL or NULL if none exists. */
502 static struct access *
503 get_first_repr_for_decl (tree base)
505 vec<access_p> *access_vec;
507 access_vec = get_base_access_vector (base);
508 if (!access_vec)
509 return NULL;
511 return (*access_vec)[0];
514 /* Find an access representative for the variable BASE and given OFFSET and
515 SIZE. Requires that access trees have already been built. Return NULL if
516 it cannot be found. */
518 static struct access *
519 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
520 HOST_WIDE_INT size)
522 struct access *access;
524 access = get_first_repr_for_decl (base);
525 while (access && (access->offset + access->size <= offset))
526 access = access->next_grp;
527 if (!access)
528 return NULL;
530 return find_access_in_subtree (access, offset, size);
533 /* Add LINK to the linked list of assign links of RACC. */
534 static void
535 add_link_to_rhs (struct access *racc, struct assign_link *link)
537 gcc_assert (link->racc == racc);
539 if (!racc->first_link)
541 gcc_assert (!racc->last_link);
542 racc->first_link = link;
544 else
545 racc->last_link->next = link;
547 racc->last_link = link;
548 link->next = NULL;
551 /* Move all link structures in their linked list in OLD_RACC to the linked list
552 in NEW_RACC. */
553 static void
554 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
556 if (!old_racc->first_link)
558 gcc_assert (!old_racc->last_link);
559 return;
562 if (new_racc->first_link)
564 gcc_assert (!new_racc->last_link->next);
565 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
567 new_racc->last_link->next = old_racc->first_link;
568 new_racc->last_link = old_racc->last_link;
570 else
572 gcc_assert (!new_racc->last_link);
574 new_racc->first_link = old_racc->first_link;
575 new_racc->last_link = old_racc->last_link;
577 old_racc->first_link = old_racc->last_link = NULL;
580 /* Add ACCESS to the work queue (which is actually a stack). */
582 static void
583 add_access_to_work_queue (struct access *access)
585 if (!access->grp_queued)
587 gcc_assert (!access->next_queued);
588 access->next_queued = work_queue_head;
589 access->grp_queued = 1;
590 work_queue_head = access;
594 /* Pop an access from the work queue, and return it, assuming there is one. */
596 static struct access *
597 pop_access_from_work_queue (void)
599 struct access *access = work_queue_head;
601 work_queue_head = access->next_queued;
602 access->next_queued = NULL;
603 access->grp_queued = 0;
604 return access;
608 /* Allocate necessary structures. */
610 static void
611 sra_initialize (void)
613 candidate_bitmap = BITMAP_ALLOC (NULL);
614 candidates = htab_create (vec_safe_length (cfun->local_decls) / 2,
615 uid_decl_map_hash, uid_decl_map_eq, NULL);
616 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
617 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
618 gcc_obstack_init (&name_obstack);
619 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
620 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
621 base_access_vec = pointer_map_create ();
622 memset (&sra_stats, 0, sizeof (sra_stats));
623 encountered_apply_args = false;
624 encountered_recursive_call = false;
625 encountered_unchangable_recursive_call = false;
628 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
630 static bool
631 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
632 void *data ATTRIBUTE_UNUSED)
634 vec<access_p> *access_vec = (vec<access_p> *) *value;
635 vec_free (access_vec);
636 return true;
639 /* Deallocate all general structures. */
641 static void
642 sra_deinitialize (void)
644 BITMAP_FREE (candidate_bitmap);
645 htab_delete (candidates);
646 BITMAP_FREE (should_scalarize_away_bitmap);
647 BITMAP_FREE (cannot_scalarize_away_bitmap);
648 free_alloc_pool (access_pool);
649 free_alloc_pool (link_pool);
650 obstack_free (&name_obstack, NULL);
652 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
653 pointer_map_destroy (base_access_vec);
656 /* Remove DECL from candidates for SRA and write REASON to the dump file if
657 there is one. */
658 static void
659 disqualify_candidate (tree decl, const char *reason)
661 if (bitmap_clear_bit (candidate_bitmap, DECL_UID (decl)))
662 htab_clear_slot (candidates,
663 htab_find_slot_with_hash (candidates, decl,
664 DECL_UID (decl), NO_INSERT));
666 if (dump_file && (dump_flags & TDF_DETAILS))
668 fprintf (dump_file, "! Disqualifying ");
669 print_generic_expr (dump_file, decl, 0);
670 fprintf (dump_file, " - %s\n", reason);
674 /* Return true iff the type contains a field or an element which does not allow
675 scalarization. */
677 static bool
678 type_internals_preclude_sra_p (tree type, const char **msg)
680 tree fld;
681 tree et;
683 switch (TREE_CODE (type))
685 case RECORD_TYPE:
686 case UNION_TYPE:
687 case QUAL_UNION_TYPE:
688 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
689 if (TREE_CODE (fld) == FIELD_DECL)
691 tree ft = TREE_TYPE (fld);
693 if (TREE_THIS_VOLATILE (fld))
695 *msg = "volatile structure field";
696 return true;
698 if (!DECL_FIELD_OFFSET (fld))
700 *msg = "no structure field offset";
701 return true;
703 if (!DECL_SIZE (fld))
705 *msg = "zero structure field size";
706 return true;
708 if (!host_integerp (DECL_FIELD_OFFSET (fld), 1))
710 *msg = "structure field offset not fixed";
711 return true;
713 if (!host_integerp (DECL_SIZE (fld), 1))
715 *msg = "structure field size not fixed";
716 return true;
718 if (!host_integerp (bit_position (fld), 0))
720 *msg = "structure field size too big";
721 return true;
723 if (AGGREGATE_TYPE_P (ft)
724 && int_bit_position (fld) % BITS_PER_UNIT != 0)
726 *msg = "structure field is bit field";
727 return true;
730 if (AGGREGATE_TYPE_P (ft) && type_internals_preclude_sra_p (ft, msg))
731 return true;
734 return false;
736 case ARRAY_TYPE:
737 et = TREE_TYPE (type);
739 if (TYPE_VOLATILE (et))
741 *msg = "element type is volatile";
742 return true;
745 if (AGGREGATE_TYPE_P (et) && type_internals_preclude_sra_p (et, msg))
746 return true;
748 return false;
750 default:
751 return false;
755 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
756 base variable if it is. Return T if it is not an SSA_NAME. */
758 static tree
759 get_ssa_base_param (tree t)
761 if (TREE_CODE (t) == SSA_NAME)
763 if (SSA_NAME_IS_DEFAULT_DEF (t))
764 return SSA_NAME_VAR (t);
765 else
766 return NULL_TREE;
768 return t;
771 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
772 belongs to, unless the BB has already been marked as a potentially
773 final. */
775 static void
776 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
778 basic_block bb = gimple_bb (stmt);
779 int idx, parm_index = 0;
780 tree parm;
782 if (bitmap_bit_p (final_bbs, bb->index))
783 return;
785 for (parm = DECL_ARGUMENTS (current_function_decl);
786 parm && parm != base;
787 parm = DECL_CHAIN (parm))
788 parm_index++;
790 gcc_assert (parm_index < func_param_count);
792 idx = bb->index * func_param_count + parm_index;
793 if (bb_dereferences[idx] < dist)
794 bb_dereferences[idx] = dist;
797 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
798 the three fields. Also add it to the vector of accesses corresponding to
799 the base. Finally, return the new access. */
801 static struct access *
802 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
804 vec<access_p> *v;
805 struct access *access;
806 void **slot;
808 access = (struct access *) pool_alloc (access_pool);
809 memset (access, 0, sizeof (struct access));
810 access->base = base;
811 access->offset = offset;
812 access->size = size;
814 slot = pointer_map_contains (base_access_vec, base);
815 if (slot)
816 v = (vec<access_p> *) *slot;
817 else
818 vec_alloc (v, 32);
820 v->safe_push (access);
822 *((vec<access_p> **)
823 pointer_map_insert (base_access_vec, base)) = v;
825 return access;
828 /* Create and insert access for EXPR. Return created access, or NULL if it is
829 not possible. */
831 static struct access *
832 create_access (tree expr, gimple stmt, bool write)
834 struct access *access;
835 HOST_WIDE_INT offset, size, max_size;
836 tree base = expr;
837 bool ptr, unscalarizable_region = false;
839 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
841 if (sra_mode == SRA_MODE_EARLY_IPA
842 && TREE_CODE (base) == MEM_REF)
844 base = get_ssa_base_param (TREE_OPERAND (base, 0));
845 if (!base)
846 return NULL;
847 ptr = true;
849 else
850 ptr = false;
852 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
853 return NULL;
855 if (sra_mode == SRA_MODE_EARLY_IPA)
857 if (size < 0 || size != max_size)
859 disqualify_candidate (base, "Encountered a variable sized access.");
860 return NULL;
862 if (TREE_CODE (expr) == COMPONENT_REF
863 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
865 disqualify_candidate (base, "Encountered a bit-field access.");
866 return NULL;
868 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
870 if (ptr)
871 mark_parm_dereference (base, offset + size, stmt);
873 else
875 if (size != max_size)
877 size = max_size;
878 unscalarizable_region = true;
880 if (size < 0)
882 disqualify_candidate (base, "Encountered an unconstrained access.");
883 return NULL;
887 access = create_access_1 (base, offset, size);
888 access->expr = expr;
889 access->type = TREE_TYPE (expr);
890 access->write = write;
891 access->grp_unscalarizable_region = unscalarizable_region;
892 access->stmt = stmt;
894 if (TREE_CODE (expr) == COMPONENT_REF
895 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
896 access->non_addressable = 1;
898 return access;
902 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
903 register types or (recursively) records with only these two kinds of fields.
904 It also returns false if any of these records contains a bit-field. */
906 static bool
907 type_consists_of_records_p (tree type)
909 tree fld;
911 if (TREE_CODE (type) != RECORD_TYPE)
912 return false;
914 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
915 if (TREE_CODE (fld) == FIELD_DECL)
917 tree ft = TREE_TYPE (fld);
919 if (DECL_BIT_FIELD (fld))
920 return false;
922 if (!is_gimple_reg_type (ft)
923 && !type_consists_of_records_p (ft))
924 return false;
927 return true;
930 /* Create total_scalarization accesses for all scalar type fields in DECL that
931 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
932 must be the top-most VAR_DECL representing the variable, OFFSET must be the
933 offset of DECL within BASE. REF must be the memory reference expression for
934 the given decl. */
936 static void
937 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
938 tree ref)
940 tree fld, decl_type = TREE_TYPE (decl);
942 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
943 if (TREE_CODE (fld) == FIELD_DECL)
945 HOST_WIDE_INT pos = offset + int_bit_position (fld);
946 tree ft = TREE_TYPE (fld);
947 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
948 NULL_TREE);
950 if (is_gimple_reg_type (ft))
952 struct access *access;
953 HOST_WIDE_INT size;
955 size = tree_low_cst (DECL_SIZE (fld), 1);
956 access = create_access_1 (base, pos, size);
957 access->expr = nref;
958 access->type = ft;
959 access->grp_total_scalarization = 1;
960 /* Accesses for intraprocedural SRA can have their stmt NULL. */
962 else
963 completely_scalarize_record (base, fld, pos, nref);
967 /* Create total_scalarization accesses for all scalar type fields in VAR and
968 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
969 type_consists_of_records_p. */
971 static void
972 completely_scalarize_var (tree var)
974 HOST_WIDE_INT size = tree_low_cst (DECL_SIZE (var), 1);
975 struct access *access;
977 access = create_access_1 (var, 0, size);
978 access->expr = var;
979 access->type = TREE_TYPE (var);
980 access->grp_total_scalarization = 1;
982 completely_scalarize_record (var, var, 0, var);
985 /* Search the given tree for a declaration by skipping handled components and
986 exclude it from the candidates. */
988 static void
989 disqualify_base_of_expr (tree t, const char *reason)
991 t = get_base_address (t);
992 if (sra_mode == SRA_MODE_EARLY_IPA
993 && TREE_CODE (t) == MEM_REF)
994 t = get_ssa_base_param (TREE_OPERAND (t, 0));
996 if (t && DECL_P (t))
997 disqualify_candidate (t, reason);
1000 /* Scan expression EXPR and create access structures for all accesses to
1001 candidates for scalarization. Return the created access or NULL if none is
1002 created. */
1004 static struct access *
1005 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
1007 struct access *ret = NULL;
1008 bool partial_ref;
1010 if (TREE_CODE (expr) == BIT_FIELD_REF
1011 || TREE_CODE (expr) == IMAGPART_EXPR
1012 || TREE_CODE (expr) == REALPART_EXPR)
1014 expr = TREE_OPERAND (expr, 0);
1015 partial_ref = true;
1017 else
1018 partial_ref = false;
1020 /* We need to dive through V_C_Es in order to get the size of its parameter
1021 and not the result type. Ada produces such statements. We are also
1022 capable of handling the topmost V_C_E but not any of those buried in other
1023 handled components. */
1024 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
1025 expr = TREE_OPERAND (expr, 0);
1027 if (contains_view_convert_expr_p (expr))
1029 disqualify_base_of_expr (expr, "V_C_E under a different handled "
1030 "component.");
1031 return NULL;
1034 switch (TREE_CODE (expr))
1036 case MEM_REF:
1037 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
1038 && sra_mode != SRA_MODE_EARLY_IPA)
1039 return NULL;
1040 /* fall through */
1041 case VAR_DECL:
1042 case PARM_DECL:
1043 case RESULT_DECL:
1044 case COMPONENT_REF:
1045 case ARRAY_REF:
1046 case ARRAY_RANGE_REF:
1047 ret = create_access (expr, stmt, write);
1048 break;
1050 default:
1051 break;
1054 if (write && partial_ref && ret)
1055 ret->grp_partial_lhs = 1;
1057 return ret;
1060 /* Scan expression EXPR and create access structures for all accesses to
1061 candidates for scalarization. Return true if any access has been inserted.
1062 STMT must be the statement from which the expression is taken, WRITE must be
1063 true if the expression is a store and false otherwise. */
1065 static bool
1066 build_access_from_expr (tree expr, gimple stmt, bool write)
1068 struct access *access;
1070 access = build_access_from_expr_1 (expr, stmt, write);
1071 if (access)
1073 /* This means the aggregate is accesses as a whole in a way other than an
1074 assign statement and thus cannot be removed even if we had a scalar
1075 replacement for everything. */
1076 if (cannot_scalarize_away_bitmap)
1077 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1078 return true;
1080 return false;
1083 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1084 modes in which it matters, return true iff they have been disqualified. RHS
1085 may be NULL, in that case ignore it. If we scalarize an aggregate in
1086 intra-SRA we may need to add statements after each statement. This is not
1087 possible if a statement unconditionally has to end the basic block. */
1088 static bool
1089 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1091 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1092 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1094 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1095 if (rhs)
1096 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1097 return true;
1099 return false;
1102 /* Scan expressions occurring in STMT, create access structures for all accesses
1103 to candidates for scalarization and remove those candidates which occur in
1104 statements or expressions that prevent them from being split apart. Return
1105 true if any access has been inserted. */
1107 static bool
1108 build_accesses_from_assign (gimple stmt)
1110 tree lhs, rhs;
1111 struct access *lacc, *racc;
1113 if (!gimple_assign_single_p (stmt)
1114 /* Scope clobbers don't influence scalarization. */
1115 || gimple_clobber_p (stmt))
1116 return false;
1118 lhs = gimple_assign_lhs (stmt);
1119 rhs = gimple_assign_rhs1 (stmt);
1121 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1122 return false;
1124 racc = build_access_from_expr_1 (rhs, stmt, false);
1125 lacc = build_access_from_expr_1 (lhs, stmt, true);
1127 if (lacc)
1128 lacc->grp_assignment_write = 1;
1130 if (racc)
1132 racc->grp_assignment_read = 1;
1133 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1134 && !is_gimple_reg_type (racc->type))
1135 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1138 if (lacc && racc
1139 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1140 && !lacc->grp_unscalarizable_region
1141 && !racc->grp_unscalarizable_region
1142 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1143 && lacc->size == racc->size
1144 && useless_type_conversion_p (lacc->type, racc->type))
1146 struct assign_link *link;
1148 link = (struct assign_link *) pool_alloc (link_pool);
1149 memset (link, 0, sizeof (struct assign_link));
1151 link->lacc = lacc;
1152 link->racc = racc;
1154 add_link_to_rhs (racc, link);
1157 return lacc || racc;
1160 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1161 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1163 static bool
1164 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1165 void *data ATTRIBUTE_UNUSED)
1167 op = get_base_address (op);
1168 if (op
1169 && DECL_P (op))
1170 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1172 return false;
1175 /* Return true iff callsite CALL has at least as many actual arguments as there
1176 are formal parameters of the function currently processed by IPA-SRA. */
1178 static inline bool
1179 callsite_has_enough_arguments_p (gimple call)
1181 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1184 /* Scan function and look for interesting expressions and create access
1185 structures for them. Return true iff any access is created. */
1187 static bool
1188 scan_function (void)
1190 basic_block bb;
1191 bool ret = false;
1193 FOR_EACH_BB (bb)
1195 gimple_stmt_iterator gsi;
1196 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1198 gimple stmt = gsi_stmt (gsi);
1199 tree t;
1200 unsigned i;
1202 if (final_bbs && stmt_can_throw_external (stmt))
1203 bitmap_set_bit (final_bbs, bb->index);
1204 switch (gimple_code (stmt))
1206 case GIMPLE_RETURN:
1207 t = gimple_return_retval (stmt);
1208 if (t != NULL_TREE)
1209 ret |= build_access_from_expr (t, stmt, false);
1210 if (final_bbs)
1211 bitmap_set_bit (final_bbs, bb->index);
1212 break;
1214 case GIMPLE_ASSIGN:
1215 ret |= build_accesses_from_assign (stmt);
1216 break;
1218 case GIMPLE_CALL:
1219 for (i = 0; i < gimple_call_num_args (stmt); i++)
1220 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1221 stmt, false);
1223 if (sra_mode == SRA_MODE_EARLY_IPA)
1225 tree dest = gimple_call_fndecl (stmt);
1226 int flags = gimple_call_flags (stmt);
1228 if (dest)
1230 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1231 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1232 encountered_apply_args = true;
1233 if (cgraph_get_node (dest)
1234 == cgraph_get_node (current_function_decl))
1236 encountered_recursive_call = true;
1237 if (!callsite_has_enough_arguments_p (stmt))
1238 encountered_unchangable_recursive_call = true;
1242 if (final_bbs
1243 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1244 bitmap_set_bit (final_bbs, bb->index);
1247 t = gimple_call_lhs (stmt);
1248 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1249 ret |= build_access_from_expr (t, stmt, true);
1250 break;
1252 case GIMPLE_ASM:
1253 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1254 asm_visit_addr);
1255 if (final_bbs)
1256 bitmap_set_bit (final_bbs, bb->index);
1258 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1260 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1261 ret |= build_access_from_expr (t, stmt, false);
1263 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1265 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1266 ret |= build_access_from_expr (t, stmt, true);
1268 break;
1270 default:
1271 break;
1276 return ret;
1279 /* Helper of QSORT function. There are pointers to accesses in the array. An
1280 access is considered smaller than another if it has smaller offset or if the
1281 offsets are the same but is size is bigger. */
1283 static int
1284 compare_access_positions (const void *a, const void *b)
1286 const access_p *fp1 = (const access_p *) a;
1287 const access_p *fp2 = (const access_p *) b;
1288 const access_p f1 = *fp1;
1289 const access_p f2 = *fp2;
1291 if (f1->offset != f2->offset)
1292 return f1->offset < f2->offset ? -1 : 1;
1294 if (f1->size == f2->size)
1296 if (f1->type == f2->type)
1297 return 0;
1298 /* Put any non-aggregate type before any aggregate type. */
1299 else if (!is_gimple_reg_type (f1->type)
1300 && is_gimple_reg_type (f2->type))
1301 return 1;
1302 else if (is_gimple_reg_type (f1->type)
1303 && !is_gimple_reg_type (f2->type))
1304 return -1;
1305 /* Put any complex or vector type before any other scalar type. */
1306 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1307 && TREE_CODE (f1->type) != VECTOR_TYPE
1308 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1309 || TREE_CODE (f2->type) == VECTOR_TYPE))
1310 return 1;
1311 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1312 || TREE_CODE (f1->type) == VECTOR_TYPE)
1313 && TREE_CODE (f2->type) != COMPLEX_TYPE
1314 && TREE_CODE (f2->type) != VECTOR_TYPE)
1315 return -1;
1316 /* Put the integral type with the bigger precision first. */
1317 else if (INTEGRAL_TYPE_P (f1->type)
1318 && INTEGRAL_TYPE_P (f2->type))
1319 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1320 /* Put any integral type with non-full precision last. */
1321 else if (INTEGRAL_TYPE_P (f1->type)
1322 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1323 != TYPE_PRECISION (f1->type)))
1324 return 1;
1325 else if (INTEGRAL_TYPE_P (f2->type)
1326 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1327 != TYPE_PRECISION (f2->type)))
1328 return -1;
1329 /* Stabilize the sort. */
1330 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1333 /* We want the bigger accesses first, thus the opposite operator in the next
1334 line: */
1335 return f1->size > f2->size ? -1 : 1;
1339 /* Append a name of the declaration to the name obstack. A helper function for
1340 make_fancy_name. */
1342 static void
1343 make_fancy_decl_name (tree decl)
1345 char buffer[32];
1347 tree name = DECL_NAME (decl);
1348 if (name)
1349 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1350 IDENTIFIER_LENGTH (name));
1351 else
1353 sprintf (buffer, "D%u", DECL_UID (decl));
1354 obstack_grow (&name_obstack, buffer, strlen (buffer));
1358 /* Helper for make_fancy_name. */
1360 static void
1361 make_fancy_name_1 (tree expr)
1363 char buffer[32];
1364 tree index;
1366 if (DECL_P (expr))
1368 make_fancy_decl_name (expr);
1369 return;
1372 switch (TREE_CODE (expr))
1374 case COMPONENT_REF:
1375 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1376 obstack_1grow (&name_obstack, '$');
1377 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1378 break;
1380 case ARRAY_REF:
1381 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1382 obstack_1grow (&name_obstack, '$');
1383 /* Arrays with only one element may not have a constant as their
1384 index. */
1385 index = TREE_OPERAND (expr, 1);
1386 if (TREE_CODE (index) != INTEGER_CST)
1387 break;
1388 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1389 obstack_grow (&name_obstack, buffer, strlen (buffer));
1390 break;
1392 case ADDR_EXPR:
1393 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1394 break;
1396 case MEM_REF:
1397 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1398 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1400 obstack_1grow (&name_obstack, '$');
1401 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1402 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1403 obstack_grow (&name_obstack, buffer, strlen (buffer));
1405 break;
1407 case BIT_FIELD_REF:
1408 case REALPART_EXPR:
1409 case IMAGPART_EXPR:
1410 gcc_unreachable (); /* we treat these as scalars. */
1411 break;
1412 default:
1413 break;
1417 /* Create a human readable name for replacement variable of ACCESS. */
1419 static char *
1420 make_fancy_name (tree expr)
1422 make_fancy_name_1 (expr);
1423 obstack_1grow (&name_obstack, '\0');
1424 return XOBFINISH (&name_obstack, char *);
1427 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1428 EXP_TYPE at the given OFFSET. If BASE is something for which
1429 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1430 to insert new statements either before or below the current one as specified
1431 by INSERT_AFTER. This function is not capable of handling bitfields.
1433 BASE must be either a declaration or a memory reference that has correct
1434 alignment ifformation embeded in it (e.g. a pre-existing one in SRA). */
1436 tree
1437 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1438 tree exp_type, gimple_stmt_iterator *gsi,
1439 bool insert_after)
1441 tree prev_base = base;
1442 tree off;
1443 HOST_WIDE_INT base_offset;
1444 unsigned HOST_WIDE_INT misalign;
1445 unsigned int align;
1447 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1448 get_object_alignment_1 (base, &align, &misalign);
1449 base = get_addr_base_and_unit_offset (base, &base_offset);
1451 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1452 offset such as array[var_index]. */
1453 if (!base)
1455 gimple stmt;
1456 tree tmp, addr;
1458 gcc_checking_assert (gsi);
1459 tmp = make_ssa_name (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1460 addr = build_fold_addr_expr (unshare_expr (prev_base));
1461 STRIP_USELESS_TYPE_CONVERSION (addr);
1462 stmt = gimple_build_assign (tmp, addr);
1463 gimple_set_location (stmt, loc);
1464 if (insert_after)
1465 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1466 else
1467 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1469 off = build_int_cst (reference_alias_ptr_type (prev_base),
1470 offset / BITS_PER_UNIT);
1471 base = tmp;
1473 else if (TREE_CODE (base) == MEM_REF)
1475 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1476 base_offset + offset / BITS_PER_UNIT);
1477 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1478 base = unshare_expr (TREE_OPERAND (base, 0));
1480 else
1482 off = build_int_cst (reference_alias_ptr_type (base),
1483 base_offset + offset / BITS_PER_UNIT);
1484 base = build_fold_addr_expr (unshare_expr (base));
1487 misalign = (misalign + offset) & (align - 1);
1488 if (misalign != 0)
1489 align = (misalign & -misalign);
1490 if (align < TYPE_ALIGN (exp_type))
1491 exp_type = build_aligned_type (exp_type, align);
1493 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1496 /* Construct a memory reference to a part of an aggregate BASE at the given
1497 OFFSET and of the same type as MODEL. In case this is a reference to a
1498 bit-field, the function will replicate the last component_ref of model's
1499 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1500 build_ref_for_offset. */
1502 static tree
1503 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1504 struct access *model, gimple_stmt_iterator *gsi,
1505 bool insert_after)
1507 if (TREE_CODE (model->expr) == COMPONENT_REF
1508 && DECL_BIT_FIELD (TREE_OPERAND (model->expr, 1)))
1510 /* This access represents a bit-field. */
1511 tree t, exp_type, fld = TREE_OPERAND (model->expr, 1);
1513 offset -= int_bit_position (fld);
1514 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1515 t = build_ref_for_offset (loc, base, offset, exp_type, gsi, insert_after);
1516 return fold_build3_loc (loc, COMPONENT_REF, TREE_TYPE (fld), t, fld,
1517 NULL_TREE);
1519 else
1520 return build_ref_for_offset (loc, base, offset, model->type,
1521 gsi, insert_after);
1524 /* Attempt to build a memory reference that we could but into a gimple
1525 debug_bind statement. Similar to build_ref_for_model but punts if it has to
1526 create statements and return s NULL instead. This function also ignores
1527 alignment issues and so its results should never end up in non-debug
1528 statements. */
1530 static tree
1531 build_debug_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1532 struct access *model)
1534 HOST_WIDE_INT base_offset;
1535 tree off;
1537 if (TREE_CODE (model->expr) == COMPONENT_REF
1538 && DECL_BIT_FIELD (TREE_OPERAND (model->expr, 1)))
1539 return NULL_TREE;
1541 base = get_addr_base_and_unit_offset (base, &base_offset);
1542 if (!base)
1543 return NULL_TREE;
1544 if (TREE_CODE (base) == MEM_REF)
1546 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1547 base_offset + offset / BITS_PER_UNIT);
1548 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1549 base = unshare_expr (TREE_OPERAND (base, 0));
1551 else
1553 off = build_int_cst (reference_alias_ptr_type (base),
1554 base_offset + offset / BITS_PER_UNIT);
1555 base = build_fold_addr_expr (unshare_expr (base));
1558 return fold_build2_loc (loc, MEM_REF, model->type, base, off);
1561 /* Construct a memory reference consisting of component_refs and array_refs to
1562 a part of an aggregate *RES (which is of type TYPE). The requested part
1563 should have type EXP_TYPE at be the given OFFSET. This function might not
1564 succeed, it returns true when it does and only then *RES points to something
1565 meaningful. This function should be used only to build expressions that we
1566 might need to present to user (e.g. in warnings). In all other situations,
1567 build_ref_for_model or build_ref_for_offset should be used instead. */
1569 static bool
1570 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1571 tree exp_type)
1573 while (1)
1575 tree fld;
1576 tree tr_size, index, minidx;
1577 HOST_WIDE_INT el_size;
1579 if (offset == 0 && exp_type
1580 && types_compatible_p (exp_type, type))
1581 return true;
1583 switch (TREE_CODE (type))
1585 case UNION_TYPE:
1586 case QUAL_UNION_TYPE:
1587 case RECORD_TYPE:
1588 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1590 HOST_WIDE_INT pos, size;
1591 tree tr_pos, expr, *expr_ptr;
1593 if (TREE_CODE (fld) != FIELD_DECL)
1594 continue;
1596 tr_pos = bit_position (fld);
1597 if (!tr_pos || !host_integerp (tr_pos, 1))
1598 continue;
1599 pos = TREE_INT_CST_LOW (tr_pos);
1600 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1601 tr_size = DECL_SIZE (fld);
1602 if (!tr_size || !host_integerp (tr_size, 1))
1603 continue;
1604 size = TREE_INT_CST_LOW (tr_size);
1605 if (size == 0)
1607 if (pos != offset)
1608 continue;
1610 else if (pos > offset || (pos + size) <= offset)
1611 continue;
1613 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1614 NULL_TREE);
1615 expr_ptr = &expr;
1616 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1617 offset - pos, exp_type))
1619 *res = expr;
1620 return true;
1623 return false;
1625 case ARRAY_TYPE:
1626 tr_size = TYPE_SIZE (TREE_TYPE (type));
1627 if (!tr_size || !host_integerp (tr_size, 1))
1628 return false;
1629 el_size = tree_low_cst (tr_size, 1);
1631 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1632 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1633 return false;
1634 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1635 if (!integer_zerop (minidx))
1636 index = int_const_binop (PLUS_EXPR, index, minidx);
1637 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1638 NULL_TREE, NULL_TREE);
1639 offset = offset % el_size;
1640 type = TREE_TYPE (type);
1641 break;
1643 default:
1644 if (offset != 0)
1645 return false;
1647 if (exp_type)
1648 return false;
1649 else
1650 return true;
1655 /* Return true iff TYPE is stdarg va_list type. */
1657 static inline bool
1658 is_va_list_type (tree type)
1660 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1663 /* Print message to dump file why a variable was rejected. */
1665 static void
1666 reject (tree var, const char *msg)
1668 if (dump_file && (dump_flags & TDF_DETAILS))
1670 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1671 print_generic_expr (dump_file, var, 0);
1672 fprintf (dump_file, "\n");
1676 /* Return true if VAR is a candidate for SRA. */
1678 static bool
1679 maybe_add_sra_candidate (tree var)
1681 tree type = TREE_TYPE (var);
1682 const char *msg;
1683 void **slot;
1685 if (!AGGREGATE_TYPE_P (type))
1687 reject (var, "not aggregate");
1688 return false;
1690 if (needs_to_live_in_memory (var))
1692 reject (var, "needs to live in memory");
1693 return false;
1695 if (TREE_THIS_VOLATILE (var))
1697 reject (var, "is volatile");
1698 return false;
1700 if (!COMPLETE_TYPE_P (type))
1702 reject (var, "has incomplete type");
1703 return false;
1705 if (!host_integerp (TYPE_SIZE (type), 1))
1707 reject (var, "type size not fixed");
1708 return false;
1710 if (tree_low_cst (TYPE_SIZE (type), 1) == 0)
1712 reject (var, "type size is zero");
1713 return false;
1715 if (type_internals_preclude_sra_p (type, &msg))
1717 reject (var, msg);
1718 return false;
1720 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1721 we also want to schedule it rather late. Thus we ignore it in
1722 the early pass. */
1723 (sra_mode == SRA_MODE_EARLY_INTRA
1724 && is_va_list_type (type)))
1726 reject (var, "is va_list");
1727 return false;
1730 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1731 slot = htab_find_slot_with_hash (candidates, var, DECL_UID (var), INSERT);
1732 *slot = (void *) var;
1734 if (dump_file && (dump_flags & TDF_DETAILS))
1736 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1737 print_generic_expr (dump_file, var, 0);
1738 fprintf (dump_file, "\n");
1741 return true;
1744 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1745 those with type which is suitable for scalarization. */
1747 static bool
1748 find_var_candidates (void)
1750 tree var, parm;
1751 unsigned int i;
1752 bool ret = false;
1754 for (parm = DECL_ARGUMENTS (current_function_decl);
1755 parm;
1756 parm = DECL_CHAIN (parm))
1757 ret |= maybe_add_sra_candidate (parm);
1759 FOR_EACH_LOCAL_DECL (cfun, i, var)
1761 if (TREE_CODE (var) != VAR_DECL)
1762 continue;
1764 ret |= maybe_add_sra_candidate (var);
1767 return ret;
1770 /* Sort all accesses for the given variable, check for partial overlaps and
1771 return NULL if there are any. If there are none, pick a representative for
1772 each combination of offset and size and create a linked list out of them.
1773 Return the pointer to the first representative and make sure it is the first
1774 one in the vector of accesses. */
1776 static struct access *
1777 sort_and_splice_var_accesses (tree var)
1779 int i, j, access_count;
1780 struct access *res, **prev_acc_ptr = &res;
1781 vec<access_p> *access_vec;
1782 bool first = true;
1783 HOST_WIDE_INT low = -1, high = 0;
1785 access_vec = get_base_access_vector (var);
1786 if (!access_vec)
1787 return NULL;
1788 access_count = access_vec->length ();
1790 /* Sort by <OFFSET, SIZE>. */
1791 access_vec->qsort (compare_access_positions);
1793 i = 0;
1794 while (i < access_count)
1796 struct access *access = (*access_vec)[i];
1797 bool grp_write = access->write;
1798 bool grp_read = !access->write;
1799 bool grp_scalar_write = access->write
1800 && is_gimple_reg_type (access->type);
1801 bool grp_scalar_read = !access->write
1802 && is_gimple_reg_type (access->type);
1803 bool grp_assignment_read = access->grp_assignment_read;
1804 bool grp_assignment_write = access->grp_assignment_write;
1805 bool multiple_scalar_reads = false;
1806 bool total_scalarization = access->grp_total_scalarization;
1807 bool grp_partial_lhs = access->grp_partial_lhs;
1808 bool first_scalar = is_gimple_reg_type (access->type);
1809 bool unscalarizable_region = access->grp_unscalarizable_region;
1811 if (first || access->offset >= high)
1813 first = false;
1814 low = access->offset;
1815 high = access->offset + access->size;
1817 else if (access->offset > low && access->offset + access->size > high)
1818 return NULL;
1819 else
1820 gcc_assert (access->offset >= low
1821 && access->offset + access->size <= high);
1823 j = i + 1;
1824 while (j < access_count)
1826 struct access *ac2 = (*access_vec)[j];
1827 if (ac2->offset != access->offset || ac2->size != access->size)
1828 break;
1829 if (ac2->write)
1831 grp_write = true;
1832 grp_scalar_write = (grp_scalar_write
1833 || is_gimple_reg_type (ac2->type));
1835 else
1837 grp_read = true;
1838 if (is_gimple_reg_type (ac2->type))
1840 if (grp_scalar_read)
1841 multiple_scalar_reads = true;
1842 else
1843 grp_scalar_read = true;
1846 grp_assignment_read |= ac2->grp_assignment_read;
1847 grp_assignment_write |= ac2->grp_assignment_write;
1848 grp_partial_lhs |= ac2->grp_partial_lhs;
1849 unscalarizable_region |= ac2->grp_unscalarizable_region;
1850 total_scalarization |= ac2->grp_total_scalarization;
1851 relink_to_new_repr (access, ac2);
1853 /* If there are both aggregate-type and scalar-type accesses with
1854 this combination of size and offset, the comparison function
1855 should have put the scalars first. */
1856 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1857 ac2->group_representative = access;
1858 j++;
1861 i = j;
1863 access->group_representative = access;
1864 access->grp_write = grp_write;
1865 access->grp_read = grp_read;
1866 access->grp_scalar_read = grp_scalar_read;
1867 access->grp_scalar_write = grp_scalar_write;
1868 access->grp_assignment_read = grp_assignment_read;
1869 access->grp_assignment_write = grp_assignment_write;
1870 access->grp_hint = multiple_scalar_reads || total_scalarization;
1871 access->grp_total_scalarization = total_scalarization;
1872 access->grp_partial_lhs = grp_partial_lhs;
1873 access->grp_unscalarizable_region = unscalarizable_region;
1874 if (access->first_link)
1875 add_access_to_work_queue (access);
1877 *prev_acc_ptr = access;
1878 prev_acc_ptr = &access->next_grp;
1881 gcc_assert (res == (*access_vec)[0]);
1882 return res;
1885 /* Create a variable for the given ACCESS which determines the type, name and a
1886 few other properties. Return the variable declaration and store it also to
1887 ACCESS->replacement. */
1889 static tree
1890 create_access_replacement (struct access *access)
1892 tree repl;
1894 if (access->grp_to_be_debug_replaced)
1896 repl = create_tmp_var_raw (access->type, NULL);
1897 DECL_CONTEXT (repl) = current_function_decl;
1899 else
1900 repl = create_tmp_var (access->type, "SR");
1901 if (TREE_CODE (access->type) == COMPLEX_TYPE
1902 || TREE_CODE (access->type) == VECTOR_TYPE)
1904 if (!access->grp_partial_lhs)
1905 DECL_GIMPLE_REG_P (repl) = 1;
1907 else if (access->grp_partial_lhs
1908 && is_gimple_reg_type (access->type))
1909 TREE_ADDRESSABLE (repl) = 1;
1911 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1912 DECL_ARTIFICIAL (repl) = 1;
1913 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1915 if (DECL_NAME (access->base)
1916 && !DECL_IGNORED_P (access->base)
1917 && !DECL_ARTIFICIAL (access->base))
1919 char *pretty_name = make_fancy_name (access->expr);
1920 tree debug_expr = unshare_expr (access->expr), d;
1921 bool fail = false;
1923 DECL_NAME (repl) = get_identifier (pretty_name);
1924 obstack_free (&name_obstack, pretty_name);
1926 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1927 as DECL_DEBUG_EXPR isn't considered when looking for still
1928 used SSA_NAMEs and thus they could be freed. All debug info
1929 generation cares is whether something is constant or variable
1930 and that get_ref_base_and_extent works properly on the
1931 expression. It cannot handle accesses at a non-constant offset
1932 though, so just give up in those cases. */
1933 for (d = debug_expr;
1934 !fail && (handled_component_p (d) || TREE_CODE (d) == MEM_REF);
1935 d = TREE_OPERAND (d, 0))
1936 switch (TREE_CODE (d))
1938 case ARRAY_REF:
1939 case ARRAY_RANGE_REF:
1940 if (TREE_OPERAND (d, 1)
1941 && TREE_CODE (TREE_OPERAND (d, 1)) != INTEGER_CST)
1942 fail = true;
1943 if (TREE_OPERAND (d, 3)
1944 && TREE_CODE (TREE_OPERAND (d, 3)) != INTEGER_CST)
1945 fail = true;
1946 /* FALLTHRU */
1947 case COMPONENT_REF:
1948 if (TREE_OPERAND (d, 2)
1949 && TREE_CODE (TREE_OPERAND (d, 2)) != INTEGER_CST)
1950 fail = true;
1951 break;
1952 case MEM_REF:
1953 if (TREE_CODE (TREE_OPERAND (d, 0)) != ADDR_EXPR)
1954 fail = true;
1955 else
1956 d = TREE_OPERAND (d, 0);
1957 break;
1958 default:
1959 break;
1961 if (!fail)
1963 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1964 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1966 if (access->grp_no_warning)
1967 TREE_NO_WARNING (repl) = 1;
1968 else
1969 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1971 else
1972 TREE_NO_WARNING (repl) = 1;
1974 if (dump_file)
1976 if (access->grp_to_be_debug_replaced)
1978 fprintf (dump_file, "Created a debug-only replacement for ");
1979 print_generic_expr (dump_file, access->base, 0);
1980 fprintf (dump_file, " offset: %u, size: %u\n",
1981 (unsigned) access->offset, (unsigned) access->size);
1983 else
1985 fprintf (dump_file, "Created a replacement for ");
1986 print_generic_expr (dump_file, access->base, 0);
1987 fprintf (dump_file, " offset: %u, size: %u: ",
1988 (unsigned) access->offset, (unsigned) access->size);
1989 print_generic_expr (dump_file, repl, 0);
1990 fprintf (dump_file, "\n");
1993 sra_stats.replacements++;
1995 return repl;
1998 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
2000 static inline tree
2001 get_access_replacement (struct access *access)
2003 if (!access->replacement_decl)
2004 access->replacement_decl = create_access_replacement (access);
2005 return access->replacement_decl;
2009 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
2010 linked list along the way. Stop when *ACCESS is NULL or the access pointed
2011 to it is not "within" the root. Return false iff some accesses partially
2012 overlap. */
2014 static bool
2015 build_access_subtree (struct access **access)
2017 struct access *root = *access, *last_child = NULL;
2018 HOST_WIDE_INT limit = root->offset + root->size;
2020 *access = (*access)->next_grp;
2021 while (*access && (*access)->offset + (*access)->size <= limit)
2023 if (!last_child)
2024 root->first_child = *access;
2025 else
2026 last_child->next_sibling = *access;
2027 last_child = *access;
2029 if (!build_access_subtree (access))
2030 return false;
2033 if (*access && (*access)->offset < limit)
2034 return false;
2036 return true;
2039 /* Build a tree of access representatives, ACCESS is the pointer to the first
2040 one, others are linked in a list by the next_grp field. Return false iff
2041 some accesses partially overlap. */
2043 static bool
2044 build_access_trees (struct access *access)
2046 while (access)
2048 struct access *root = access;
2050 if (!build_access_subtree (&access))
2051 return false;
2052 root->next_grp = access;
2054 return true;
2057 /* Return true if expr contains some ARRAY_REFs into a variable bounded
2058 array. */
2060 static bool
2061 expr_with_var_bounded_array_refs_p (tree expr)
2063 while (handled_component_p (expr))
2065 if (TREE_CODE (expr) == ARRAY_REF
2066 && !host_integerp (array_ref_low_bound (expr), 0))
2067 return true;
2068 expr = TREE_OPERAND (expr, 0);
2070 return false;
2073 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
2074 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
2075 sorts of access flags appropriately along the way, notably always set
2076 grp_read and grp_assign_read according to MARK_READ and grp_write when
2077 MARK_WRITE is true.
2079 Creating a replacement for a scalar access is considered beneficial if its
2080 grp_hint is set (this means we are either attempting total scalarization or
2081 there is more than one direct read access) or according to the following
2082 table:
2084 Access written to through a scalar type (once or more times)
2086 | Written to in an assignment statement
2088 | | Access read as scalar _once_
2089 | | |
2090 | | | Read in an assignment statement
2091 | | | |
2092 | | | | Scalarize Comment
2093 -----------------------------------------------------------------------------
2094 0 0 0 0 No access for the scalar
2095 0 0 0 1 No access for the scalar
2096 0 0 1 0 No Single read - won't help
2097 0 0 1 1 No The same case
2098 0 1 0 0 No access for the scalar
2099 0 1 0 1 No access for the scalar
2100 0 1 1 0 Yes s = *g; return s.i;
2101 0 1 1 1 Yes The same case as above
2102 1 0 0 0 No Won't help
2103 1 0 0 1 Yes s.i = 1; *g = s;
2104 1 0 1 0 Yes s.i = 5; g = s.i;
2105 1 0 1 1 Yes The same case as above
2106 1 1 0 0 No Won't help.
2107 1 1 0 1 Yes s.i = 1; *g = s;
2108 1 1 1 0 Yes s = *g; return s.i;
2109 1 1 1 1 Yes Any of the above yeses */
2111 static bool
2112 analyze_access_subtree (struct access *root, struct access *parent,
2113 bool allow_replacements)
2115 struct access *child;
2116 HOST_WIDE_INT limit = root->offset + root->size;
2117 HOST_WIDE_INT covered_to = root->offset;
2118 bool scalar = is_gimple_reg_type (root->type);
2119 bool hole = false, sth_created = false;
2121 if (parent)
2123 if (parent->grp_read)
2124 root->grp_read = 1;
2125 if (parent->grp_assignment_read)
2126 root->grp_assignment_read = 1;
2127 if (parent->grp_write)
2128 root->grp_write = 1;
2129 if (parent->grp_assignment_write)
2130 root->grp_assignment_write = 1;
2131 if (parent->grp_total_scalarization)
2132 root->grp_total_scalarization = 1;
2135 if (root->grp_unscalarizable_region)
2136 allow_replacements = false;
2138 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2139 allow_replacements = false;
2141 for (child = root->first_child; child; child = child->next_sibling)
2143 hole |= covered_to < child->offset;
2144 sth_created |= analyze_access_subtree (child, root,
2145 allow_replacements && !scalar);
2147 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2148 root->grp_total_scalarization &= child->grp_total_scalarization;
2149 if (child->grp_covered)
2150 covered_to += child->size;
2151 else
2152 hole = true;
2155 if (allow_replacements && scalar && !root->first_child
2156 && (root->grp_hint
2157 || ((root->grp_scalar_read || root->grp_assignment_read)
2158 && (root->grp_scalar_write || root->grp_assignment_write))))
2160 bool new_integer_type;
2161 /* Always create access replacements that cover the whole access.
2162 For integral types this means the precision has to match.
2163 Avoid assumptions based on the integral type kind, too. */
2164 if (INTEGRAL_TYPE_P (root->type)
2165 && (TREE_CODE (root->type) != INTEGER_TYPE
2166 || TYPE_PRECISION (root->type) != root->size)
2167 /* But leave bitfield accesses alone. */
2168 && (TREE_CODE (root->expr) != COMPONENT_REF
2169 || !DECL_BIT_FIELD (TREE_OPERAND (root->expr, 1))))
2171 tree rt = root->type;
2172 gcc_assert ((root->offset % BITS_PER_UNIT) == 0
2173 && (root->size % BITS_PER_UNIT) == 0);
2174 root->type = build_nonstandard_integer_type (root->size,
2175 TYPE_UNSIGNED (rt));
2176 root->expr = build_ref_for_offset (UNKNOWN_LOCATION,
2177 root->base, root->offset,
2178 root->type, NULL, false);
2179 new_integer_type = true;
2181 else
2182 new_integer_type = false;
2184 if (dump_file && (dump_flags & TDF_DETAILS))
2186 fprintf (dump_file, "Marking ");
2187 print_generic_expr (dump_file, root->base, 0);
2188 fprintf (dump_file, " offset: %u, size: %u ",
2189 (unsigned) root->offset, (unsigned) root->size);
2190 fprintf (dump_file, " to be replaced%s.\n",
2191 new_integer_type ? " with an integer": "");
2194 root->grp_to_be_replaced = 1;
2195 sth_created = true;
2196 hole = false;
2198 else
2200 if (MAY_HAVE_DEBUG_STMTS && allow_replacements
2201 && scalar && !root->first_child
2202 && (root->grp_scalar_write || root->grp_assignment_write))
2204 gcc_checking_assert (!root->grp_scalar_read
2205 && !root->grp_assignment_read);
2206 root->grp_to_be_debug_replaced = 1;
2207 if (dump_file && (dump_flags & TDF_DETAILS))
2209 fprintf (dump_file, "Marking ");
2210 print_generic_expr (dump_file, root->base, 0);
2211 fprintf (dump_file, " offset: %u, size: %u ",
2212 (unsigned) root->offset, (unsigned) root->size);
2213 fprintf (dump_file, " to be replaced with debug statements.\n");
2217 if (covered_to < limit)
2218 hole = true;
2219 if (scalar)
2220 root->grp_total_scalarization = 0;
2223 if (sth_created
2224 && (!hole || root->grp_total_scalarization))
2226 root->grp_covered = 1;
2227 return true;
2229 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2230 root->grp_unscalarized_data = 1; /* not covered and written to */
2231 if (sth_created)
2232 return true;
2233 return false;
2236 /* Analyze all access trees linked by next_grp by the means of
2237 analyze_access_subtree. */
2238 static bool
2239 analyze_access_trees (struct access *access)
2241 bool ret = false;
2243 while (access)
2245 if (analyze_access_subtree (access, NULL, true))
2246 ret = true;
2247 access = access->next_grp;
2250 return ret;
2253 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2254 SIZE would conflict with an already existing one. If exactly such a child
2255 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2257 static bool
2258 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2259 HOST_WIDE_INT size, struct access **exact_match)
2261 struct access *child;
2263 for (child = lacc->first_child; child; child = child->next_sibling)
2265 if (child->offset == norm_offset && child->size == size)
2267 *exact_match = child;
2268 return true;
2271 if (child->offset < norm_offset + size
2272 && child->offset + child->size > norm_offset)
2273 return true;
2276 return false;
2279 /* Create a new child access of PARENT, with all properties just like MODEL
2280 except for its offset and with its grp_write false and grp_read true.
2281 Return the new access or NULL if it cannot be created. Note that this access
2282 is created long after all splicing and sorting, it's not located in any
2283 access vector and is automatically a representative of its group. */
2285 static struct access *
2286 create_artificial_child_access (struct access *parent, struct access *model,
2287 HOST_WIDE_INT new_offset)
2289 struct access *access;
2290 struct access **child;
2291 tree expr = parent->base;
2293 gcc_assert (!model->grp_unscalarizable_region);
2295 access = (struct access *) pool_alloc (access_pool);
2296 memset (access, 0, sizeof (struct access));
2297 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2298 model->type))
2300 access->grp_no_warning = true;
2301 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2302 new_offset, model, NULL, false);
2305 access->base = parent->base;
2306 access->expr = expr;
2307 access->offset = new_offset;
2308 access->size = model->size;
2309 access->type = model->type;
2310 access->grp_write = true;
2311 access->grp_read = false;
2313 child = &parent->first_child;
2314 while (*child && (*child)->offset < new_offset)
2315 child = &(*child)->next_sibling;
2317 access->next_sibling = *child;
2318 *child = access;
2320 return access;
2324 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2325 true if any new subaccess was created. Additionally, if RACC is a scalar
2326 access but LACC is not, change the type of the latter, if possible. */
2328 static bool
2329 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2331 struct access *rchild;
2332 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2333 bool ret = false;
2335 if (is_gimple_reg_type (lacc->type)
2336 || lacc->grp_unscalarizable_region
2337 || racc->grp_unscalarizable_region)
2338 return false;
2340 if (is_gimple_reg_type (racc->type))
2342 if (!lacc->first_child && !racc->first_child)
2344 tree t = lacc->base;
2346 lacc->type = racc->type;
2347 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t),
2348 lacc->offset, racc->type))
2349 lacc->expr = t;
2350 else
2352 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2353 lacc->base, lacc->offset,
2354 racc, NULL, false);
2355 lacc->grp_no_warning = true;
2358 return false;
2361 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2363 struct access *new_acc = NULL;
2364 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2366 if (rchild->grp_unscalarizable_region)
2367 continue;
2369 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2370 &new_acc))
2372 if (new_acc)
2374 rchild->grp_hint = 1;
2375 new_acc->grp_hint |= new_acc->grp_read;
2376 if (rchild->first_child)
2377 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2379 continue;
2382 rchild->grp_hint = 1;
2383 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2384 if (new_acc)
2386 ret = true;
2387 if (racc->first_child)
2388 propagate_subaccesses_across_link (new_acc, rchild);
2392 return ret;
2395 /* Propagate all subaccesses across assignment links. */
2397 static void
2398 propagate_all_subaccesses (void)
2400 while (work_queue_head)
2402 struct access *racc = pop_access_from_work_queue ();
2403 struct assign_link *link;
2405 gcc_assert (racc->first_link);
2407 for (link = racc->first_link; link; link = link->next)
2409 struct access *lacc = link->lacc;
2411 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2412 continue;
2413 lacc = lacc->group_representative;
2414 if (propagate_subaccesses_across_link (lacc, racc)
2415 && lacc->first_link)
2416 add_access_to_work_queue (lacc);
2421 /* Go through all accesses collected throughout the (intraprocedural) analysis
2422 stage, exclude overlapping ones, identify representatives and build trees
2423 out of them, making decisions about scalarization on the way. Return true
2424 iff there are any to-be-scalarized variables after this stage. */
2426 static bool
2427 analyze_all_variable_accesses (void)
2429 int res = 0;
2430 bitmap tmp = BITMAP_ALLOC (NULL);
2431 bitmap_iterator bi;
2432 unsigned i, max_total_scalarization_size;
2434 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2435 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2437 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2438 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2439 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2441 tree var = candidate (i);
2443 if (TREE_CODE (var) == VAR_DECL
2444 && type_consists_of_records_p (TREE_TYPE (var)))
2446 if ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2447 <= max_total_scalarization_size)
2449 completely_scalarize_var (var);
2450 if (dump_file && (dump_flags & TDF_DETAILS))
2452 fprintf (dump_file, "Will attempt to totally scalarize ");
2453 print_generic_expr (dump_file, var, 0);
2454 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2457 else if (dump_file && (dump_flags & TDF_DETAILS))
2459 fprintf (dump_file, "Too big to totally scalarize: ");
2460 print_generic_expr (dump_file, var, 0);
2461 fprintf (dump_file, " (UID: %u)\n", DECL_UID (var));
2466 bitmap_copy (tmp, candidate_bitmap);
2467 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2469 tree var = candidate (i);
2470 struct access *access;
2472 access = sort_and_splice_var_accesses (var);
2473 if (!access || !build_access_trees (access))
2474 disqualify_candidate (var,
2475 "No or inhibitingly overlapping accesses.");
2478 propagate_all_subaccesses ();
2480 bitmap_copy (tmp, candidate_bitmap);
2481 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2483 tree var = candidate (i);
2484 struct access *access = get_first_repr_for_decl (var);
2486 if (analyze_access_trees (access))
2488 res++;
2489 if (dump_file && (dump_flags & TDF_DETAILS))
2491 fprintf (dump_file, "\nAccess trees for ");
2492 print_generic_expr (dump_file, var, 0);
2493 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2494 dump_access_tree (dump_file, access);
2495 fprintf (dump_file, "\n");
2498 else
2499 disqualify_candidate (var, "No scalar replacements to be created.");
2502 BITMAP_FREE (tmp);
2504 if (res)
2506 statistics_counter_event (cfun, "Scalarized aggregates", res);
2507 return true;
2509 else
2510 return false;
2513 /* Generate statements copying scalar replacements of accesses within a subtree
2514 into or out of AGG. ACCESS, all its children, siblings and their children
2515 are to be processed. AGG is an aggregate type expression (can be a
2516 declaration but does not have to be, it can for example also be a mem_ref or
2517 a series of handled components). TOP_OFFSET is the offset of the processed
2518 subtree which has to be subtracted from offsets of individual accesses to
2519 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2520 replacements in the interval <start_offset, start_offset + chunk_size>,
2521 otherwise copy all. GSI is a statement iterator used to place the new
2522 statements. WRITE should be true when the statements should write from AGG
2523 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2524 statements will be added after the current statement in GSI, they will be
2525 added before the statement otherwise. */
2527 static void
2528 generate_subtree_copies (struct access *access, tree agg,
2529 HOST_WIDE_INT top_offset,
2530 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2531 gimple_stmt_iterator *gsi, bool write,
2532 bool insert_after, location_t loc)
2536 if (chunk_size && access->offset >= start_offset + chunk_size)
2537 return;
2539 if (access->grp_to_be_replaced
2540 && (chunk_size == 0
2541 || access->offset + access->size > start_offset))
2543 tree expr, repl = get_access_replacement (access);
2544 gimple stmt;
2546 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2547 access, gsi, insert_after);
2549 if (write)
2551 if (access->grp_partial_lhs)
2552 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2553 !insert_after,
2554 insert_after ? GSI_NEW_STMT
2555 : GSI_SAME_STMT);
2556 stmt = gimple_build_assign (repl, expr);
2558 else
2560 TREE_NO_WARNING (repl) = 1;
2561 if (access->grp_partial_lhs)
2562 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2563 !insert_after,
2564 insert_after ? GSI_NEW_STMT
2565 : GSI_SAME_STMT);
2566 stmt = gimple_build_assign (expr, repl);
2568 gimple_set_location (stmt, loc);
2570 if (insert_after)
2571 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2572 else
2573 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2574 update_stmt (stmt);
2575 sra_stats.subtree_copies++;
2577 else if (write
2578 && access->grp_to_be_debug_replaced
2579 && (chunk_size == 0
2580 || access->offset + access->size > start_offset))
2582 gimple ds;
2583 tree drhs = build_debug_ref_for_model (loc, agg,
2584 access->offset - top_offset,
2585 access);
2586 ds = gimple_build_debug_bind (get_access_replacement (access),
2587 drhs, gsi_stmt (*gsi));
2588 if (insert_after)
2589 gsi_insert_after (gsi, ds, GSI_NEW_STMT);
2590 else
2591 gsi_insert_before (gsi, ds, GSI_SAME_STMT);
2594 if (access->first_child)
2595 generate_subtree_copies (access->first_child, agg, top_offset,
2596 start_offset, chunk_size, gsi,
2597 write, insert_after, loc);
2599 access = access->next_sibling;
2601 while (access);
2604 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2605 the root of the subtree to be processed. GSI is the statement iterator used
2606 for inserting statements which are added after the current statement if
2607 INSERT_AFTER is true or before it otherwise. */
2609 static void
2610 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2611 bool insert_after, location_t loc)
2614 struct access *child;
2616 if (access->grp_to_be_replaced)
2618 gimple stmt;
2620 stmt = gimple_build_assign (get_access_replacement (access),
2621 build_zero_cst (access->type));
2622 if (insert_after)
2623 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2624 else
2625 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2626 update_stmt (stmt);
2627 gimple_set_location (stmt, loc);
2629 else if (access->grp_to_be_debug_replaced)
2631 gimple ds = gimple_build_debug_bind (get_access_replacement (access),
2632 build_zero_cst (access->type),
2633 gsi_stmt (*gsi));
2634 if (insert_after)
2635 gsi_insert_after (gsi, ds, GSI_NEW_STMT);
2636 else
2637 gsi_insert_before (gsi, ds, GSI_SAME_STMT);
2640 for (child = access->first_child; child; child = child->next_sibling)
2641 init_subtree_with_zero (child, gsi, insert_after, loc);
2644 /* Search for an access representative for the given expression EXPR and
2645 return it or NULL if it cannot be found. */
2647 static struct access *
2648 get_access_for_expr (tree expr)
2650 HOST_WIDE_INT offset, size, max_size;
2651 tree base;
2653 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2654 a different size than the size of its argument and we need the latter
2655 one. */
2656 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2657 expr = TREE_OPERAND (expr, 0);
2659 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2660 if (max_size == -1 || !DECL_P (base))
2661 return NULL;
2663 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2664 return NULL;
2666 return get_var_base_offset_size_access (base, offset, max_size);
2669 /* Replace the expression EXPR with a scalar replacement if there is one and
2670 generate other statements to do type conversion or subtree copying if
2671 necessary. GSI is used to place newly created statements, WRITE is true if
2672 the expression is being written to (it is on a LHS of a statement or output
2673 in an assembly statement). */
2675 static bool
2676 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2678 location_t loc;
2679 struct access *access;
2680 tree type, bfr;
2682 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2684 bfr = *expr;
2685 expr = &TREE_OPERAND (*expr, 0);
2687 else
2688 bfr = NULL_TREE;
2690 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2691 expr = &TREE_OPERAND (*expr, 0);
2692 access = get_access_for_expr (*expr);
2693 if (!access)
2694 return false;
2695 type = TREE_TYPE (*expr);
2697 loc = gimple_location (gsi_stmt (*gsi));
2698 if (access->grp_to_be_replaced)
2700 tree repl = get_access_replacement (access);
2701 /* If we replace a non-register typed access simply use the original
2702 access expression to extract the scalar component afterwards.
2703 This happens if scalarizing a function return value or parameter
2704 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2705 gcc.c-torture/compile/20011217-1.c.
2707 We also want to use this when accessing a complex or vector which can
2708 be accessed as a different type too, potentially creating a need for
2709 type conversion (see PR42196) and when scalarized unions are involved
2710 in assembler statements (see PR42398). */
2711 if (!useless_type_conversion_p (type, access->type))
2713 tree ref;
2715 ref = build_ref_for_model (loc, access->base, access->offset, access,
2716 NULL, false);
2718 if (write)
2720 gimple stmt;
2722 if (access->grp_partial_lhs)
2723 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2724 false, GSI_NEW_STMT);
2725 stmt = gimple_build_assign (repl, ref);
2726 gimple_set_location (stmt, loc);
2727 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2729 else
2731 gimple stmt;
2733 if (access->grp_partial_lhs)
2734 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2735 true, GSI_SAME_STMT);
2736 stmt = gimple_build_assign (ref, repl);
2737 gimple_set_location (stmt, loc);
2738 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2741 else
2742 *expr = repl;
2743 sra_stats.exprs++;
2745 else if (write && access->grp_to_be_debug_replaced)
2747 gimple ds = gimple_build_debug_bind (get_access_replacement (access),
2748 NULL_TREE,
2749 gsi_stmt (*gsi));
2750 gsi_insert_after (gsi, ds, GSI_NEW_STMT);
2753 if (access->first_child)
2755 HOST_WIDE_INT start_offset, chunk_size;
2756 if (bfr
2757 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2758 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2760 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2761 start_offset = access->offset
2762 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2764 else
2765 start_offset = chunk_size = 0;
2767 generate_subtree_copies (access->first_child, access->base, 0,
2768 start_offset, chunk_size, gsi, write, write,
2769 loc);
2771 return true;
2774 /* Where scalar replacements of the RHS have been written to when a replacement
2775 of a LHS of an assigments cannot be direclty loaded from a replacement of
2776 the RHS. */
2777 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2778 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2779 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2781 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2782 base aggregate if there are unscalarized data or directly to LHS of the
2783 statement that is pointed to by GSI otherwise. */
2785 static enum unscalarized_data_handling
2786 handle_unscalarized_data_in_subtree (struct access *top_racc,
2787 gimple_stmt_iterator *gsi)
2789 if (top_racc->grp_unscalarized_data)
2791 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2792 gsi, false, false,
2793 gimple_location (gsi_stmt (*gsi)));
2794 return SRA_UDH_RIGHT;
2796 else
2798 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2799 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2800 0, 0, gsi, false, false,
2801 gimple_location (gsi_stmt (*gsi)));
2802 return SRA_UDH_LEFT;
2807 /* Try to generate statements to load all sub-replacements in an access subtree
2808 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2809 If that is not possible, refresh the TOP_RACC base aggregate and load the
2810 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2811 copied. NEW_GSI is stmt iterator used for statement insertions after the
2812 original assignment, OLD_GSI is used to insert statements before the
2813 assignment. *REFRESHED keeps the information whether we have needed to
2814 refresh replacements of the LHS and from which side of the assignments this
2815 takes place. */
2817 static void
2818 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2819 HOST_WIDE_INT left_offset,
2820 gimple_stmt_iterator *old_gsi,
2821 gimple_stmt_iterator *new_gsi,
2822 enum unscalarized_data_handling *refreshed)
2824 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2825 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2827 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2829 if (lacc->grp_to_be_replaced)
2831 struct access *racc;
2832 gimple stmt;
2833 tree rhs;
2835 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2836 if (racc && racc->grp_to_be_replaced)
2838 rhs = get_access_replacement (racc);
2839 if (!useless_type_conversion_p (lacc->type, racc->type))
2840 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2842 if (racc->grp_partial_lhs && lacc->grp_partial_lhs)
2843 rhs = force_gimple_operand_gsi (old_gsi, rhs, true, NULL_TREE,
2844 true, GSI_SAME_STMT);
2846 else
2848 /* No suitable access on the right hand side, need to load from
2849 the aggregate. See if we have to update it first... */
2850 if (*refreshed == SRA_UDH_NONE)
2851 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2852 old_gsi);
2854 if (*refreshed == SRA_UDH_LEFT)
2855 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2856 new_gsi, true);
2857 else
2858 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2859 new_gsi, true);
2860 if (lacc->grp_partial_lhs)
2861 rhs = force_gimple_operand_gsi (new_gsi, rhs, true, NULL_TREE,
2862 false, GSI_NEW_STMT);
2865 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2866 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2867 gimple_set_location (stmt, loc);
2868 update_stmt (stmt);
2869 sra_stats.subreplacements++;
2871 else
2873 if (*refreshed == SRA_UDH_NONE
2874 && lacc->grp_read && !lacc->grp_covered)
2875 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2876 old_gsi);
2877 if (lacc && lacc->grp_to_be_debug_replaced)
2879 gimple ds;
2880 tree drhs;
2881 struct access *racc = find_access_in_subtree (top_racc, offset,
2882 lacc->size);
2884 if (racc && racc->grp_to_be_replaced)
2885 drhs = get_access_replacement (racc);
2886 else if (*refreshed == SRA_UDH_LEFT)
2887 drhs = build_debug_ref_for_model (loc, lacc->base, lacc->offset,
2888 lacc);
2889 else if (*refreshed == SRA_UDH_RIGHT)
2890 drhs = build_debug_ref_for_model (loc, top_racc->base, offset,
2891 lacc);
2892 else
2893 drhs = NULL_TREE;
2894 ds = gimple_build_debug_bind (get_access_replacement (lacc),
2895 drhs, gsi_stmt (*old_gsi));
2896 gsi_insert_after (new_gsi, ds, GSI_NEW_STMT);
2900 if (lacc->first_child)
2901 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2902 old_gsi, new_gsi, refreshed);
2906 /* Result code for SRA assignment modification. */
2907 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2908 SRA_AM_MODIFIED, /* stmt changed but not
2909 removed */
2910 SRA_AM_REMOVED }; /* stmt eliminated */
2912 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2913 to the assignment and GSI is the statement iterator pointing at it. Returns
2914 the same values as sra_modify_assign. */
2916 static enum assignment_mod_result
2917 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2919 tree lhs = gimple_assign_lhs (*stmt);
2920 struct access *acc;
2921 location_t loc;
2923 acc = get_access_for_expr (lhs);
2924 if (!acc)
2925 return SRA_AM_NONE;
2927 if (gimple_clobber_p (*stmt))
2929 /* Remove clobbers of fully scalarized variables, otherwise
2930 do nothing. */
2931 if (acc->grp_covered)
2933 unlink_stmt_vdef (*stmt);
2934 gsi_remove (gsi, true);
2935 release_defs (*stmt);
2936 return SRA_AM_REMOVED;
2938 else
2939 return SRA_AM_NONE;
2942 loc = gimple_location (*stmt);
2943 if (vec_safe_length (CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2945 /* I have never seen this code path trigger but if it can happen the
2946 following should handle it gracefully. */
2947 if (access_has_children_p (acc))
2948 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2949 true, true, loc);
2950 return SRA_AM_MODIFIED;
2953 if (acc->grp_covered)
2955 init_subtree_with_zero (acc, gsi, false, loc);
2956 unlink_stmt_vdef (*stmt);
2957 gsi_remove (gsi, true);
2958 release_defs (*stmt);
2959 return SRA_AM_REMOVED;
2961 else
2963 init_subtree_with_zero (acc, gsi, true, loc);
2964 return SRA_AM_MODIFIED;
2968 /* Create and return a new suitable default definition SSA_NAME for RACC which
2969 is an access describing an uninitialized part of an aggregate that is being
2970 loaded. */
2972 static tree
2973 get_repl_default_def_ssa_name (struct access *racc)
2975 return get_or_create_ssa_default_def (cfun, get_access_replacement (racc));
2978 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2979 somewhere in it. */
2981 static inline bool
2982 contains_bitfld_comp_ref_p (const_tree ref)
2984 while (handled_component_p (ref))
2986 if (TREE_CODE (ref) == COMPONENT_REF
2987 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2988 return true;
2989 ref = TREE_OPERAND (ref, 0);
2992 return false;
2995 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2996 bit-field field declaration somewhere in it. */
2998 static inline bool
2999 contains_vce_or_bfcref_p (const_tree ref)
3001 while (handled_component_p (ref))
3003 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
3004 || (TREE_CODE (ref) == COMPONENT_REF
3005 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
3006 return true;
3007 ref = TREE_OPERAND (ref, 0);
3010 return false;
3013 /* Examine both sides of the assignment statement pointed to by STMT, replace
3014 them with a scalare replacement if there is one and generate copying of
3015 replacements if scalarized aggregates have been used in the assignment. GSI
3016 is used to hold generated statements for type conversions and subtree
3017 copying. */
3019 static enum assignment_mod_result
3020 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
3022 struct access *lacc, *racc;
3023 tree lhs, rhs;
3024 bool modify_this_stmt = false;
3025 bool force_gimple_rhs = false;
3026 location_t loc;
3027 gimple_stmt_iterator orig_gsi = *gsi;
3029 if (!gimple_assign_single_p (*stmt))
3030 return SRA_AM_NONE;
3031 lhs = gimple_assign_lhs (*stmt);
3032 rhs = gimple_assign_rhs1 (*stmt);
3034 if (TREE_CODE (rhs) == CONSTRUCTOR)
3035 return sra_modify_constructor_assign (stmt, gsi);
3037 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
3038 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
3039 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
3041 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
3042 gsi, false);
3043 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
3044 gsi, true);
3045 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3048 lacc = get_access_for_expr (lhs);
3049 racc = get_access_for_expr (rhs);
3050 if (!lacc && !racc)
3051 return SRA_AM_NONE;
3053 loc = gimple_location (*stmt);
3054 if (lacc && lacc->grp_to_be_replaced)
3056 lhs = get_access_replacement (lacc);
3057 gimple_assign_set_lhs (*stmt, lhs);
3058 modify_this_stmt = true;
3059 if (lacc->grp_partial_lhs)
3060 force_gimple_rhs = true;
3061 sra_stats.exprs++;
3064 if (racc && racc->grp_to_be_replaced)
3066 rhs = get_access_replacement (racc);
3067 modify_this_stmt = true;
3068 if (racc->grp_partial_lhs)
3069 force_gimple_rhs = true;
3070 sra_stats.exprs++;
3072 else if (racc
3073 && !racc->grp_unscalarized_data
3074 && TREE_CODE (lhs) == SSA_NAME
3075 && !access_has_replacements_p (racc))
3077 rhs = get_repl_default_def_ssa_name (racc);
3078 modify_this_stmt = true;
3079 sra_stats.exprs++;
3082 if (modify_this_stmt)
3084 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3086 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
3087 ??? This should move to fold_stmt which we simply should
3088 call after building a VIEW_CONVERT_EXPR here. */
3089 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
3090 && !contains_bitfld_comp_ref_p (lhs))
3092 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
3093 gimple_assign_set_lhs (*stmt, lhs);
3095 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
3096 && !contains_vce_or_bfcref_p (rhs))
3097 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
3099 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
3101 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
3102 rhs);
3103 if (is_gimple_reg_type (TREE_TYPE (lhs))
3104 && TREE_CODE (lhs) != SSA_NAME)
3105 force_gimple_rhs = true;
3110 if (lacc && lacc->grp_to_be_debug_replaced)
3112 gimple ds = gimple_build_debug_bind (get_access_replacement (lacc),
3113 unshare_expr (rhs), *stmt);
3114 gsi_insert_before (gsi, ds, GSI_SAME_STMT);
3117 /* From this point on, the function deals with assignments in between
3118 aggregates when at least one has scalar reductions of some of its
3119 components. There are three possible scenarios: Both the LHS and RHS have
3120 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
3122 In the first case, we would like to load the LHS components from RHS
3123 components whenever possible. If that is not possible, we would like to
3124 read it directly from the RHS (after updating it by storing in it its own
3125 components). If there are some necessary unscalarized data in the LHS,
3126 those will be loaded by the original assignment too. If neither of these
3127 cases happen, the original statement can be removed. Most of this is done
3128 by load_assign_lhs_subreplacements.
3130 In the second case, we would like to store all RHS scalarized components
3131 directly into LHS and if they cover the aggregate completely, remove the
3132 statement too. In the third case, we want the LHS components to be loaded
3133 directly from the RHS (DSE will remove the original statement if it
3134 becomes redundant).
3136 This is a bit complex but manageable when types match and when unions do
3137 not cause confusion in a way that we cannot really load a component of LHS
3138 from the RHS or vice versa (the access representing this level can have
3139 subaccesses that are accessible only through a different union field at a
3140 higher level - different from the one used in the examined expression).
3141 Unions are fun.
3143 Therefore, I specially handle a fourth case, happening when there is a
3144 specific type cast or it is impossible to locate a scalarized subaccess on
3145 the other side of the expression. If that happens, I simply "refresh" the
3146 RHS by storing in it is scalarized components leave the original statement
3147 there to do the copying and then load the scalar replacements of the LHS.
3148 This is what the first branch does. */
3150 if (modify_this_stmt
3151 || gimple_has_volatile_ops (*stmt)
3152 || contains_vce_or_bfcref_p (rhs)
3153 || contains_vce_or_bfcref_p (lhs))
3155 if (access_has_children_p (racc))
3156 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3157 gsi, false, false, loc);
3158 if (access_has_children_p (lacc))
3159 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
3160 gsi, true, true, loc);
3161 sra_stats.separate_lhs_rhs_handling++;
3163 /* This gimplification must be done after generate_subtree_copies,
3164 lest we insert the subtree copies in the middle of the gimplified
3165 sequence. */
3166 if (force_gimple_rhs)
3167 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
3168 true, GSI_SAME_STMT);
3169 if (gimple_assign_rhs1 (*stmt) != rhs)
3171 modify_this_stmt = true;
3172 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
3173 gcc_assert (*stmt == gsi_stmt (orig_gsi));
3176 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3178 else
3180 if (access_has_children_p (lacc)
3181 && access_has_children_p (racc)
3182 /* When an access represents an unscalarizable region, it usually
3183 represents accesses with variable offset and thus must not be used
3184 to generate new memory accesses. */
3185 && !lacc->grp_unscalarizable_region
3186 && !racc->grp_unscalarizable_region)
3188 gimple_stmt_iterator orig_gsi = *gsi;
3189 enum unscalarized_data_handling refreshed;
3191 if (lacc->grp_read && !lacc->grp_covered)
3192 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
3193 else
3194 refreshed = SRA_UDH_NONE;
3196 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
3197 &orig_gsi, gsi, &refreshed);
3198 if (refreshed != SRA_UDH_RIGHT)
3200 gsi_next (gsi);
3201 unlink_stmt_vdef (*stmt);
3202 gsi_remove (&orig_gsi, true);
3203 release_defs (*stmt);
3204 sra_stats.deleted++;
3205 return SRA_AM_REMOVED;
3208 else
3210 if (access_has_children_p (racc)
3211 && !racc->grp_unscalarized_data)
3213 if (dump_file)
3215 fprintf (dump_file, "Removing load: ");
3216 print_gimple_stmt (dump_file, *stmt, 0, 0);
3218 generate_subtree_copies (racc->first_child, lhs,
3219 racc->offset, 0, 0, gsi,
3220 false, false, loc);
3221 gcc_assert (*stmt == gsi_stmt (*gsi));
3222 unlink_stmt_vdef (*stmt);
3223 gsi_remove (gsi, true);
3224 release_defs (*stmt);
3225 sra_stats.deleted++;
3226 return SRA_AM_REMOVED;
3228 /* Restore the aggregate RHS from its components so the
3229 prevailing aggregate copy does the right thing. */
3230 if (access_has_children_p (racc))
3231 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
3232 gsi, false, false, loc);
3233 /* Re-load the components of the aggregate copy destination.
3234 But use the RHS aggregate to load from to expose more
3235 optimization opportunities. */
3236 if (access_has_children_p (lacc))
3237 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
3238 0, 0, gsi, true, true, loc);
3241 return SRA_AM_NONE;
3245 /* Traverse the function body and all modifications as decided in
3246 analyze_all_variable_accesses. Return true iff the CFG has been
3247 changed. */
3249 static bool
3250 sra_modify_function_body (void)
3252 bool cfg_changed = false;
3253 basic_block bb;
3255 FOR_EACH_BB (bb)
3257 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3258 while (!gsi_end_p (gsi))
3260 gimple stmt = gsi_stmt (gsi);
3261 enum assignment_mod_result assign_result;
3262 bool modified = false, deleted = false;
3263 tree *t;
3264 unsigned i;
3266 switch (gimple_code (stmt))
3268 case GIMPLE_RETURN:
3269 t = gimple_return_retval_ptr (stmt);
3270 if (*t != NULL_TREE)
3271 modified |= sra_modify_expr (t, &gsi, false);
3272 break;
3274 case GIMPLE_ASSIGN:
3275 assign_result = sra_modify_assign (&stmt, &gsi);
3276 modified |= assign_result == SRA_AM_MODIFIED;
3277 deleted = assign_result == SRA_AM_REMOVED;
3278 break;
3280 case GIMPLE_CALL:
3281 /* Operands must be processed before the lhs. */
3282 for (i = 0; i < gimple_call_num_args (stmt); i++)
3284 t = gimple_call_arg_ptr (stmt, i);
3285 modified |= sra_modify_expr (t, &gsi, false);
3288 if (gimple_call_lhs (stmt))
3290 t = gimple_call_lhs_ptr (stmt);
3291 modified |= sra_modify_expr (t, &gsi, true);
3293 break;
3295 case GIMPLE_ASM:
3296 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3298 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3299 modified |= sra_modify_expr (t, &gsi, false);
3301 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3303 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3304 modified |= sra_modify_expr (t, &gsi, true);
3306 break;
3308 default:
3309 break;
3312 if (modified)
3314 update_stmt (stmt);
3315 if (maybe_clean_eh_stmt (stmt)
3316 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3317 cfg_changed = true;
3319 if (!deleted)
3320 gsi_next (&gsi);
3324 return cfg_changed;
3327 /* Generate statements initializing scalar replacements of parts of function
3328 parameters. */
3330 static void
3331 initialize_parameter_reductions (void)
3333 gimple_stmt_iterator gsi;
3334 gimple_seq seq = NULL;
3335 tree parm;
3337 gsi = gsi_start (seq);
3338 for (parm = DECL_ARGUMENTS (current_function_decl);
3339 parm;
3340 parm = DECL_CHAIN (parm))
3342 vec<access_p> *access_vec;
3343 struct access *access;
3345 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3346 continue;
3347 access_vec = get_base_access_vector (parm);
3348 if (!access_vec)
3349 continue;
3351 for (access = (*access_vec)[0];
3352 access;
3353 access = access->next_grp)
3354 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3355 EXPR_LOCATION (parm));
3358 seq = gsi_seq (gsi);
3359 if (seq)
3360 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3363 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3364 it reveals there are components of some aggregates to be scalarized, it runs
3365 the required transformations. */
3366 static unsigned int
3367 perform_intra_sra (void)
3369 int ret = 0;
3370 sra_initialize ();
3372 if (!find_var_candidates ())
3373 goto out;
3375 if (!scan_function ())
3376 goto out;
3378 if (!analyze_all_variable_accesses ())
3379 goto out;
3381 if (sra_modify_function_body ())
3382 ret = TODO_update_ssa | TODO_cleanup_cfg;
3383 else
3384 ret = TODO_update_ssa;
3385 initialize_parameter_reductions ();
3387 statistics_counter_event (cfun, "Scalar replacements created",
3388 sra_stats.replacements);
3389 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3390 statistics_counter_event (cfun, "Subtree copy stmts",
3391 sra_stats.subtree_copies);
3392 statistics_counter_event (cfun, "Subreplacement stmts",
3393 sra_stats.subreplacements);
3394 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3395 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3396 sra_stats.separate_lhs_rhs_handling);
3398 out:
3399 sra_deinitialize ();
3400 return ret;
3403 /* Perform early intraprocedural SRA. */
3404 static unsigned int
3405 early_intra_sra (void)
3407 sra_mode = SRA_MODE_EARLY_INTRA;
3408 return perform_intra_sra ();
3411 /* Perform "late" intraprocedural SRA. */
3412 static unsigned int
3413 late_intra_sra (void)
3415 sra_mode = SRA_MODE_INTRA;
3416 return perform_intra_sra ();
3420 static bool
3421 gate_intra_sra (void)
3423 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3427 struct gimple_opt_pass pass_sra_early =
3430 GIMPLE_PASS,
3431 "esra", /* name */
3432 OPTGROUP_NONE, /* optinfo_flags */
3433 gate_intra_sra, /* gate */
3434 early_intra_sra, /* execute */
3435 NULL, /* sub */
3436 NULL, /* next */
3437 0, /* static_pass_number */
3438 TV_TREE_SRA, /* tv_id */
3439 PROP_cfg | PROP_ssa, /* properties_required */
3440 0, /* properties_provided */
3441 0, /* properties_destroyed */
3442 0, /* todo_flags_start */
3443 TODO_update_ssa
3444 | TODO_ggc_collect
3445 | TODO_verify_ssa /* todo_flags_finish */
3449 struct gimple_opt_pass pass_sra =
3452 GIMPLE_PASS,
3453 "sra", /* name */
3454 OPTGROUP_NONE, /* optinfo_flags */
3455 gate_intra_sra, /* gate */
3456 late_intra_sra, /* execute */
3457 NULL, /* sub */
3458 NULL, /* next */
3459 0, /* static_pass_number */
3460 TV_TREE_SRA, /* tv_id */
3461 PROP_cfg | PROP_ssa, /* properties_required */
3462 0, /* properties_provided */
3463 0, /* properties_destroyed */
3464 TODO_update_address_taken, /* todo_flags_start */
3465 TODO_update_ssa
3466 | TODO_ggc_collect
3467 | TODO_verify_ssa /* todo_flags_finish */
3472 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3473 parameter. */
3475 static bool
3476 is_unused_scalar_param (tree parm)
3478 tree name;
3479 return (is_gimple_reg (parm)
3480 && (!(name = ssa_default_def (cfun, parm))
3481 || has_zero_uses (name)));
3484 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3485 examine whether there are any direct or otherwise infeasible ones. If so,
3486 return true, otherwise return false. PARM must be a gimple register with a
3487 non-NULL default definition. */
3489 static bool
3490 ptr_parm_has_direct_uses (tree parm)
3492 imm_use_iterator ui;
3493 gimple stmt;
3494 tree name = ssa_default_def (cfun, parm);
3495 bool ret = false;
3497 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3499 int uses_ok = 0;
3500 use_operand_p use_p;
3502 if (is_gimple_debug (stmt))
3503 continue;
3505 /* Valid uses include dereferences on the lhs and the rhs. */
3506 if (gimple_has_lhs (stmt))
3508 tree lhs = gimple_get_lhs (stmt);
3509 while (handled_component_p (lhs))
3510 lhs = TREE_OPERAND (lhs, 0);
3511 if (TREE_CODE (lhs) == MEM_REF
3512 && TREE_OPERAND (lhs, 0) == name
3513 && integer_zerop (TREE_OPERAND (lhs, 1))
3514 && types_compatible_p (TREE_TYPE (lhs),
3515 TREE_TYPE (TREE_TYPE (name)))
3516 && !TREE_THIS_VOLATILE (lhs))
3517 uses_ok++;
3519 if (gimple_assign_single_p (stmt))
3521 tree rhs = gimple_assign_rhs1 (stmt);
3522 while (handled_component_p (rhs))
3523 rhs = TREE_OPERAND (rhs, 0);
3524 if (TREE_CODE (rhs) == MEM_REF
3525 && TREE_OPERAND (rhs, 0) == name
3526 && integer_zerop (TREE_OPERAND (rhs, 1))
3527 && types_compatible_p (TREE_TYPE (rhs),
3528 TREE_TYPE (TREE_TYPE (name)))
3529 && !TREE_THIS_VOLATILE (rhs))
3530 uses_ok++;
3532 else if (is_gimple_call (stmt))
3534 unsigned i;
3535 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3537 tree arg = gimple_call_arg (stmt, i);
3538 while (handled_component_p (arg))
3539 arg = TREE_OPERAND (arg, 0);
3540 if (TREE_CODE (arg) == MEM_REF
3541 && TREE_OPERAND (arg, 0) == name
3542 && integer_zerop (TREE_OPERAND (arg, 1))
3543 && types_compatible_p (TREE_TYPE (arg),
3544 TREE_TYPE (TREE_TYPE (name)))
3545 && !TREE_THIS_VOLATILE (arg))
3546 uses_ok++;
3550 /* If the number of valid uses does not match the number of
3551 uses in this stmt there is an unhandled use. */
3552 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3553 --uses_ok;
3555 if (uses_ok != 0)
3556 ret = true;
3558 if (ret)
3559 BREAK_FROM_IMM_USE_STMT (ui);
3562 return ret;
3565 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3566 them in candidate_bitmap. Note that these do not necessarily include
3567 parameter which are unused and thus can be removed. Return true iff any
3568 such candidate has been found. */
3570 static bool
3571 find_param_candidates (void)
3573 tree parm;
3574 int count = 0;
3575 bool ret = false;
3576 const char *msg;
3578 for (parm = DECL_ARGUMENTS (current_function_decl);
3579 parm;
3580 parm = DECL_CHAIN (parm))
3582 tree type = TREE_TYPE (parm);
3583 void **slot;
3585 count++;
3587 if (TREE_THIS_VOLATILE (parm)
3588 || TREE_ADDRESSABLE (parm)
3589 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3590 continue;
3592 if (is_unused_scalar_param (parm))
3594 ret = true;
3595 continue;
3598 if (POINTER_TYPE_P (type))
3600 type = TREE_TYPE (type);
3602 if (TREE_CODE (type) == FUNCTION_TYPE
3603 || TYPE_VOLATILE (type)
3604 || upc_shared_type_p (type)
3605 || (TREE_CODE (type) == ARRAY_TYPE
3606 && TYPE_NONALIASED_COMPONENT (type))
3607 || !is_gimple_reg (parm)
3608 || is_va_list_type (type)
3609 || ptr_parm_has_direct_uses (parm))
3610 continue;
3612 else if (!AGGREGATE_TYPE_P (type))
3613 continue;
3615 if (!COMPLETE_TYPE_P (type)
3616 || !host_integerp (TYPE_SIZE (type), 1)
3617 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3618 || (AGGREGATE_TYPE_P (type)
3619 && type_internals_preclude_sra_p (type, &msg)))
3620 continue;
3622 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3623 slot = htab_find_slot_with_hash (candidates, parm,
3624 DECL_UID (parm), INSERT);
3625 *slot = (void *) parm;
3627 ret = true;
3628 if (dump_file && (dump_flags & TDF_DETAILS))
3630 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3631 print_generic_expr (dump_file, parm, 0);
3632 fprintf (dump_file, "\n");
3636 func_param_count = count;
3637 return ret;
3640 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3641 maybe_modified. */
3643 static bool
3644 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3645 void *data)
3647 struct access *repr = (struct access *) data;
3649 repr->grp_maybe_modified = 1;
3650 return true;
3653 /* Analyze what representatives (in linked lists accessible from
3654 REPRESENTATIVES) can be modified by side effects of statements in the
3655 current function. */
3657 static void
3658 analyze_modified_params (vec<access_p> representatives)
3660 int i;
3662 for (i = 0; i < func_param_count; i++)
3664 struct access *repr;
3666 for (repr = representatives[i];
3667 repr;
3668 repr = repr->next_grp)
3670 struct access *access;
3671 bitmap visited;
3672 ao_ref ar;
3674 if (no_accesses_p (repr))
3675 continue;
3676 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3677 || repr->grp_maybe_modified)
3678 continue;
3680 ao_ref_init (&ar, repr->expr);
3681 visited = BITMAP_ALLOC (NULL);
3682 for (access = repr; access; access = access->next_sibling)
3684 /* All accesses are read ones, otherwise grp_maybe_modified would
3685 be trivially set. */
3686 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3687 mark_maybe_modified, repr, &visited);
3688 if (repr->grp_maybe_modified)
3689 break;
3691 BITMAP_FREE (visited);
3696 /* Propagate distances in bb_dereferences in the opposite direction than the
3697 control flow edges, in each step storing the maximum of the current value
3698 and the minimum of all successors. These steps are repeated until the table
3699 stabilizes. Note that BBs which might terminate the functions (according to
3700 final_bbs bitmap) never updated in this way. */
3702 static void
3703 propagate_dereference_distances (void)
3705 vec<basic_block> queue;
3706 basic_block bb;
3708 queue.create (last_basic_block_for_function (cfun));
3709 queue.quick_push (ENTRY_BLOCK_PTR);
3710 FOR_EACH_BB (bb)
3712 queue.quick_push (bb);
3713 bb->aux = bb;
3716 while (!queue.is_empty ())
3718 edge_iterator ei;
3719 edge e;
3720 bool change = false;
3721 int i;
3723 bb = queue.pop ();
3724 bb->aux = NULL;
3726 if (bitmap_bit_p (final_bbs, bb->index))
3727 continue;
3729 for (i = 0; i < func_param_count; i++)
3731 int idx = bb->index * func_param_count + i;
3732 bool first = true;
3733 HOST_WIDE_INT inh = 0;
3735 FOR_EACH_EDGE (e, ei, bb->succs)
3737 int succ_idx = e->dest->index * func_param_count + i;
3739 if (e->src == EXIT_BLOCK_PTR)
3740 continue;
3742 if (first)
3744 first = false;
3745 inh = bb_dereferences [succ_idx];
3747 else if (bb_dereferences [succ_idx] < inh)
3748 inh = bb_dereferences [succ_idx];
3751 if (!first && bb_dereferences[idx] < inh)
3753 bb_dereferences[idx] = inh;
3754 change = true;
3758 if (change && !bitmap_bit_p (final_bbs, bb->index))
3759 FOR_EACH_EDGE (e, ei, bb->preds)
3761 if (e->src->aux)
3762 continue;
3764 e->src->aux = e->src;
3765 queue.quick_push (e->src);
3769 queue.release ();
3772 /* Dump a dereferences TABLE with heading STR to file F. */
3774 static void
3775 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3777 basic_block bb;
3779 fprintf (dump_file, str);
3780 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3782 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3783 if (bb != EXIT_BLOCK_PTR)
3785 int i;
3786 for (i = 0; i < func_param_count; i++)
3788 int idx = bb->index * func_param_count + i;
3789 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3792 fprintf (f, "\n");
3794 fprintf (dump_file, "\n");
3797 /* Determine what (parts of) parameters passed by reference that are not
3798 assigned to are not certainly dereferenced in this function and thus the
3799 dereferencing cannot be safely moved to the caller without potentially
3800 introducing a segfault. Mark such REPRESENTATIVES as
3801 grp_not_necessarilly_dereferenced.
3803 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3804 part is calculated rather than simple booleans are calculated for each
3805 pointer parameter to handle cases when only a fraction of the whole
3806 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3807 an example).
3809 The maximum dereference distances for each pointer parameter and BB are
3810 already stored in bb_dereference. This routine simply propagates these
3811 values upwards by propagate_dereference_distances and then compares the
3812 distances of individual parameters in the ENTRY BB to the equivalent
3813 distances of each representative of a (fraction of a) parameter. */
3815 static void
3816 analyze_caller_dereference_legality (vec<access_p> representatives)
3818 int i;
3820 if (dump_file && (dump_flags & TDF_DETAILS))
3821 dump_dereferences_table (dump_file,
3822 "Dereference table before propagation:\n",
3823 bb_dereferences);
3825 propagate_dereference_distances ();
3827 if (dump_file && (dump_flags & TDF_DETAILS))
3828 dump_dereferences_table (dump_file,
3829 "Dereference table after propagation:\n",
3830 bb_dereferences);
3832 for (i = 0; i < func_param_count; i++)
3834 struct access *repr = representatives[i];
3835 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3837 if (!repr || no_accesses_p (repr))
3838 continue;
3842 if ((repr->offset + repr->size) > bb_dereferences[idx])
3843 repr->grp_not_necessarilly_dereferenced = 1;
3844 repr = repr->next_grp;
3846 while (repr);
3850 /* Return the representative access for the parameter declaration PARM if it is
3851 a scalar passed by reference which is not written to and the pointer value
3852 is not used directly. Thus, if it is legal to dereference it in the caller
3853 and we can rule out modifications through aliases, such parameter should be
3854 turned into one passed by value. Return NULL otherwise. */
3856 static struct access *
3857 unmodified_by_ref_scalar_representative (tree parm)
3859 int i, access_count;
3860 struct access *repr;
3861 vec<access_p> *access_vec;
3863 access_vec = get_base_access_vector (parm);
3864 gcc_assert (access_vec);
3865 repr = (*access_vec)[0];
3866 if (repr->write)
3867 return NULL;
3868 repr->group_representative = repr;
3870 access_count = access_vec->length ();
3871 for (i = 1; i < access_count; i++)
3873 struct access *access = (*access_vec)[i];
3874 if (access->write)
3875 return NULL;
3876 access->group_representative = repr;
3877 access->next_sibling = repr->next_sibling;
3878 repr->next_sibling = access;
3881 repr->grp_read = 1;
3882 repr->grp_scalar_ptr = 1;
3883 return repr;
3886 /* Return true iff this ACCESS precludes IPA-SRA of the parameter it is
3887 associated with. REQ_ALIGN is the minimum required alignment. */
3889 static bool
3890 access_precludes_ipa_sra_p (struct access *access, unsigned int req_align)
3892 unsigned int exp_align;
3893 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3894 is incompatible assign in a call statement (and possibly even in asm
3895 statements). This can be relaxed by using a new temporary but only for
3896 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3897 intraprocedural SRA we deal with this by keeping the old aggregate around,
3898 something we cannot do in IPA-SRA.) */
3899 if (access->write
3900 && (is_gimple_call (access->stmt)
3901 || gimple_code (access->stmt) == GIMPLE_ASM))
3902 return true;
3904 exp_align = get_object_alignment (access->expr);
3905 if (exp_align < req_align)
3906 return true;
3908 return false;
3912 /* Sort collected accesses for parameter PARM, identify representatives for
3913 each accessed region and link them together. Return NULL if there are
3914 different but overlapping accesses, return the special ptr value meaning
3915 there are no accesses for this parameter if that is the case and return the
3916 first representative otherwise. Set *RO_GRP if there is a group of accesses
3917 with only read (i.e. no write) accesses. */
3919 static struct access *
3920 splice_param_accesses (tree parm, bool *ro_grp)
3922 int i, j, access_count, group_count;
3923 int agg_size, total_size = 0;
3924 struct access *access, *res, **prev_acc_ptr = &res;
3925 vec<access_p> *access_vec;
3927 access_vec = get_base_access_vector (parm);
3928 if (!access_vec)
3929 return &no_accesses_representant;
3930 access_count = access_vec->length ();
3932 access_vec->qsort (compare_access_positions);
3934 i = 0;
3935 total_size = 0;
3936 group_count = 0;
3937 while (i < access_count)
3939 bool modification;
3940 tree a1_alias_type;
3941 access = (*access_vec)[i];
3942 modification = access->write;
3943 if (access_precludes_ipa_sra_p (access, TYPE_ALIGN (access->type)))
3944 return NULL;
3945 a1_alias_type = reference_alias_ptr_type (access->expr);
3947 /* Access is about to become group representative unless we find some
3948 nasty overlap which would preclude us from breaking this parameter
3949 apart. */
3951 j = i + 1;
3952 while (j < access_count)
3954 struct access *ac2 = (*access_vec)[j];
3955 if (ac2->offset != access->offset)
3957 /* All or nothing law for parameters. */
3958 if (access->offset + access->size > ac2->offset)
3959 return NULL;
3960 else
3961 break;
3963 else if (ac2->size != access->size)
3964 return NULL;
3966 if (access_precludes_ipa_sra_p (ac2, TYPE_ALIGN (access->type))
3967 || (ac2->type != access->type
3968 && (TREE_ADDRESSABLE (ac2->type)
3969 || TREE_ADDRESSABLE (access->type)))
3970 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3971 return NULL;
3973 modification |= ac2->write;
3974 ac2->group_representative = access;
3975 ac2->next_sibling = access->next_sibling;
3976 access->next_sibling = ac2;
3977 j++;
3980 group_count++;
3981 access->grp_maybe_modified = modification;
3982 if (!modification)
3983 *ro_grp = true;
3984 *prev_acc_ptr = access;
3985 prev_acc_ptr = &access->next_grp;
3986 total_size += access->size;
3987 i = j;
3990 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3991 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3992 else
3993 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3994 if (total_size >= agg_size)
3995 return NULL;
3997 gcc_assert (group_count > 0);
3998 return res;
4001 /* Decide whether parameters with representative accesses given by REPR should
4002 be reduced into components. */
4004 static int
4005 decide_one_param_reduction (struct access *repr)
4007 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
4008 bool by_ref;
4009 tree parm;
4011 parm = repr->base;
4012 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
4013 gcc_assert (cur_parm_size > 0);
4015 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4017 by_ref = true;
4018 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
4020 else
4022 by_ref = false;
4023 agg_size = cur_parm_size;
4026 if (dump_file)
4028 struct access *acc;
4029 fprintf (dump_file, "Evaluating PARAM group sizes for ");
4030 print_generic_expr (dump_file, parm, 0);
4031 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
4032 for (acc = repr; acc; acc = acc->next_grp)
4033 dump_access (dump_file, acc, true);
4036 total_size = 0;
4037 new_param_count = 0;
4039 for (; repr; repr = repr->next_grp)
4041 gcc_assert (parm == repr->base);
4043 /* Taking the address of a non-addressable field is verboten. */
4044 if (by_ref && repr->non_addressable)
4045 return 0;
4047 /* Do not decompose a non-BLKmode param in a way that would
4048 create BLKmode params. Especially for by-reference passing
4049 (thus, pointer-type param) this is hardly worthwhile. */
4050 if (DECL_MODE (parm) != BLKmode
4051 && TYPE_MODE (repr->type) == BLKmode)
4052 return 0;
4054 if (!by_ref || (!repr->grp_maybe_modified
4055 && !repr->grp_not_necessarilly_dereferenced))
4056 total_size += repr->size;
4057 else
4058 total_size += cur_parm_size;
4060 new_param_count++;
4063 gcc_assert (new_param_count > 0);
4065 if (optimize_function_for_size_p (cfun))
4066 parm_size_limit = cur_parm_size;
4067 else
4068 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
4069 * cur_parm_size);
4071 if (total_size < agg_size
4072 && total_size <= parm_size_limit)
4074 if (dump_file)
4075 fprintf (dump_file, " ....will be split into %i components\n",
4076 new_param_count);
4077 return new_param_count;
4079 else
4080 return 0;
4083 /* The order of the following enums is important, we need to do extra work for
4084 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
4085 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
4086 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
4088 /* Identify representatives of all accesses to all candidate parameters for
4089 IPA-SRA. Return result based on what representatives have been found. */
4091 static enum ipa_splicing_result
4092 splice_all_param_accesses (vec<access_p> &representatives)
4094 enum ipa_splicing_result result = NO_GOOD_ACCESS;
4095 tree parm;
4096 struct access *repr;
4098 representatives.create (func_param_count);
4100 for (parm = DECL_ARGUMENTS (current_function_decl);
4101 parm;
4102 parm = DECL_CHAIN (parm))
4104 if (is_unused_scalar_param (parm))
4106 representatives.quick_push (&no_accesses_representant);
4107 if (result == NO_GOOD_ACCESS)
4108 result = UNUSED_PARAMS;
4110 else if (POINTER_TYPE_P (TREE_TYPE (parm))
4111 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
4112 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4114 repr = unmodified_by_ref_scalar_representative (parm);
4115 representatives.quick_push (repr);
4116 if (repr)
4117 result = UNMODIF_BY_REF_ACCESSES;
4119 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
4121 bool ro_grp = false;
4122 repr = splice_param_accesses (parm, &ro_grp);
4123 representatives.quick_push (repr);
4125 if (repr && !no_accesses_p (repr))
4127 if (POINTER_TYPE_P (TREE_TYPE (parm)))
4129 if (ro_grp)
4130 result = UNMODIF_BY_REF_ACCESSES;
4131 else if (result < MODIF_BY_REF_ACCESSES)
4132 result = MODIF_BY_REF_ACCESSES;
4134 else if (result < BY_VAL_ACCESSES)
4135 result = BY_VAL_ACCESSES;
4137 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
4138 result = UNUSED_PARAMS;
4140 else
4141 representatives.quick_push (NULL);
4144 if (result == NO_GOOD_ACCESS)
4146 representatives.release ();
4147 return NO_GOOD_ACCESS;
4150 return result;
4153 /* Return the index of BASE in PARMS. Abort if it is not found. */
4155 static inline int
4156 get_param_index (tree base, vec<tree> parms)
4158 int i, len;
4160 len = parms.length ();
4161 for (i = 0; i < len; i++)
4162 if (parms[i] == base)
4163 return i;
4164 gcc_unreachable ();
4167 /* Convert the decisions made at the representative level into compact
4168 parameter adjustments. REPRESENTATIVES are pointers to first
4169 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
4170 final number of adjustments. */
4172 static ipa_parm_adjustment_vec
4173 turn_representatives_into_adjustments (vec<access_p> representatives,
4174 int adjustments_count)
4176 vec<tree> parms;
4177 ipa_parm_adjustment_vec adjustments;
4178 tree parm;
4179 int i;
4181 gcc_assert (adjustments_count > 0);
4182 parms = ipa_get_vector_of_formal_parms (current_function_decl);
4183 adjustments.create (adjustments_count);
4184 parm = DECL_ARGUMENTS (current_function_decl);
4185 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
4187 struct access *repr = representatives[i];
4189 if (!repr || no_accesses_p (repr))
4191 struct ipa_parm_adjustment adj;
4193 memset (&adj, 0, sizeof (adj));
4194 adj.base_index = get_param_index (parm, parms);
4195 adj.base = parm;
4196 if (!repr)
4197 adj.copy_param = 1;
4198 else
4199 adj.remove_param = 1;
4200 adjustments.quick_push (adj);
4202 else
4204 struct ipa_parm_adjustment adj;
4205 int index = get_param_index (parm, parms);
4207 for (; repr; repr = repr->next_grp)
4209 memset (&adj, 0, sizeof (adj));
4210 gcc_assert (repr->base == parm);
4211 adj.base_index = index;
4212 adj.base = repr->base;
4213 adj.type = repr->type;
4214 adj.alias_ptr_type = reference_alias_ptr_type (repr->expr);
4215 adj.offset = repr->offset;
4216 adj.by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
4217 && (repr->grp_maybe_modified
4218 || repr->grp_not_necessarilly_dereferenced));
4219 adjustments.quick_push (adj);
4223 parms.release ();
4224 return adjustments;
4227 /* Analyze the collected accesses and produce a plan what to do with the
4228 parameters in the form of adjustments, NULL meaning nothing. */
4230 static ipa_parm_adjustment_vec
4231 analyze_all_param_acesses (void)
4233 enum ipa_splicing_result repr_state;
4234 bool proceed = false;
4235 int i, adjustments_count = 0;
4236 vec<access_p> representatives;
4237 ipa_parm_adjustment_vec adjustments;
4239 repr_state = splice_all_param_accesses (representatives);
4240 if (repr_state == NO_GOOD_ACCESS)
4241 return ipa_parm_adjustment_vec();
4243 /* If there are any parameters passed by reference which are not modified
4244 directly, we need to check whether they can be modified indirectly. */
4245 if (repr_state == UNMODIF_BY_REF_ACCESSES)
4247 analyze_caller_dereference_legality (representatives);
4248 analyze_modified_params (representatives);
4251 for (i = 0; i < func_param_count; i++)
4253 struct access *repr = representatives[i];
4255 if (repr && !no_accesses_p (repr))
4257 if (repr->grp_scalar_ptr)
4259 adjustments_count++;
4260 if (repr->grp_not_necessarilly_dereferenced
4261 || repr->grp_maybe_modified)
4262 representatives[i] = NULL;
4263 else
4265 proceed = true;
4266 sra_stats.scalar_by_ref_to_by_val++;
4269 else
4271 int new_components = decide_one_param_reduction (repr);
4273 if (new_components == 0)
4275 representatives[i] = NULL;
4276 adjustments_count++;
4278 else
4280 adjustments_count += new_components;
4281 sra_stats.aggregate_params_reduced++;
4282 sra_stats.param_reductions_created += new_components;
4283 proceed = true;
4287 else
4289 if (no_accesses_p (repr))
4291 proceed = true;
4292 sra_stats.deleted_unused_parameters++;
4294 adjustments_count++;
4298 if (!proceed && dump_file)
4299 fprintf (dump_file, "NOT proceeding to change params.\n");
4301 if (proceed)
4302 adjustments = turn_representatives_into_adjustments (representatives,
4303 adjustments_count);
4304 else
4305 adjustments = ipa_parm_adjustment_vec();
4307 representatives.release ();
4308 return adjustments;
4311 /* If a parameter replacement identified by ADJ does not yet exist in the form
4312 of declaration, create it and record it, otherwise return the previously
4313 created one. */
4315 static tree
4316 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4318 tree repl;
4319 if (!adj->new_ssa_base)
4321 char *pretty_name = make_fancy_name (adj->base);
4323 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4324 DECL_NAME (repl) = get_identifier (pretty_name);
4325 obstack_free (&name_obstack, pretty_name);
4327 adj->new_ssa_base = repl;
4329 else
4330 repl = adj->new_ssa_base;
4331 return repl;
4334 /* Find the first adjustment for a particular parameter BASE in a vector of
4335 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4336 adjustment. */
4338 static struct ipa_parm_adjustment *
4339 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4341 int i, len;
4343 len = adjustments.length ();
4344 for (i = 0; i < len; i++)
4346 struct ipa_parm_adjustment *adj;
4348 adj = &adjustments[i];
4349 if (!adj->copy_param && adj->base == base)
4350 return adj;
4353 return NULL;
4356 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4357 removed because its value is not used, replace the SSA_NAME with a one
4358 relating to a created VAR_DECL together all of its uses and return true.
4359 ADJUSTMENTS is a pointer to an adjustments vector. */
4361 static bool
4362 replace_removed_params_ssa_names (gimple stmt,
4363 ipa_parm_adjustment_vec adjustments)
4365 struct ipa_parm_adjustment *adj;
4366 tree lhs, decl, repl, name;
4368 if (gimple_code (stmt) == GIMPLE_PHI)
4369 lhs = gimple_phi_result (stmt);
4370 else if (is_gimple_assign (stmt))
4371 lhs = gimple_assign_lhs (stmt);
4372 else if (is_gimple_call (stmt))
4373 lhs = gimple_call_lhs (stmt);
4374 else
4375 gcc_unreachable ();
4377 if (TREE_CODE (lhs) != SSA_NAME)
4378 return false;
4380 decl = SSA_NAME_VAR (lhs);
4381 if (decl == NULL_TREE
4382 || TREE_CODE (decl) != PARM_DECL)
4383 return false;
4385 adj = get_adjustment_for_base (adjustments, decl);
4386 if (!adj)
4387 return false;
4389 repl = get_replaced_param_substitute (adj);
4390 name = make_ssa_name (repl, stmt);
4392 if (dump_file)
4394 fprintf (dump_file, "replacing an SSA name of a removed param ");
4395 print_generic_expr (dump_file, lhs, 0);
4396 fprintf (dump_file, " with ");
4397 print_generic_expr (dump_file, name, 0);
4398 fprintf (dump_file, "\n");
4401 if (is_gimple_assign (stmt))
4402 gimple_assign_set_lhs (stmt, name);
4403 else if (is_gimple_call (stmt))
4404 gimple_call_set_lhs (stmt, name);
4405 else
4406 gimple_phi_set_result (stmt, name);
4408 replace_uses_by (lhs, name);
4409 release_ssa_name (lhs);
4410 return true;
4413 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4414 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4415 specifies whether the function should care about type incompatibility the
4416 current and new expressions. If it is false, the function will leave
4417 incompatibility issues to the caller. Return true iff the expression
4418 was modified. */
4420 static bool
4421 sra_ipa_modify_expr (tree *expr, bool convert,
4422 ipa_parm_adjustment_vec adjustments)
4424 int i, len;
4425 struct ipa_parm_adjustment *adj, *cand = NULL;
4426 HOST_WIDE_INT offset, size, max_size;
4427 tree base, src;
4429 len = adjustments.length ();
4431 if (TREE_CODE (*expr) == BIT_FIELD_REF
4432 || TREE_CODE (*expr) == IMAGPART_EXPR
4433 || TREE_CODE (*expr) == REALPART_EXPR)
4435 expr = &TREE_OPERAND (*expr, 0);
4436 convert = true;
4439 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4440 if (!base || size == -1 || max_size == -1)
4441 return false;
4443 if (TREE_CODE (base) == MEM_REF)
4445 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4446 base = TREE_OPERAND (base, 0);
4449 base = get_ssa_base_param (base);
4450 if (!base || TREE_CODE (base) != PARM_DECL)
4451 return false;
4453 for (i = 0; i < len; i++)
4455 adj = &adjustments[i];
4457 if (adj->base == base &&
4458 (adj->offset == offset || adj->remove_param))
4460 cand = adj;
4461 break;
4464 if (!cand || cand->copy_param || cand->remove_param)
4465 return false;
4467 if (cand->by_ref)
4468 src = build_simple_mem_ref (cand->reduction);
4469 else
4470 src = cand->reduction;
4472 if (dump_file && (dump_flags & TDF_DETAILS))
4474 fprintf (dump_file, "About to replace expr ");
4475 print_generic_expr (dump_file, *expr, 0);
4476 fprintf (dump_file, " with ");
4477 print_generic_expr (dump_file, src, 0);
4478 fprintf (dump_file, "\n");
4481 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4483 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4484 *expr = vce;
4486 else
4487 *expr = src;
4488 return true;
4491 /* If the statement pointed to by STMT_PTR contains any expressions that need
4492 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4493 potential type incompatibilities (GSI is used to accommodate conversion
4494 statements and must point to the statement). Return true iff the statement
4495 was modified. */
4497 static bool
4498 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4499 ipa_parm_adjustment_vec adjustments)
4501 gimple stmt = *stmt_ptr;
4502 tree *lhs_p, *rhs_p;
4503 bool any;
4505 if (!gimple_assign_single_p (stmt))
4506 return false;
4508 rhs_p = gimple_assign_rhs1_ptr (stmt);
4509 lhs_p = gimple_assign_lhs_ptr (stmt);
4511 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4512 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4513 if (any)
4515 tree new_rhs = NULL_TREE;
4517 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4519 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4521 /* V_C_Es of constructors can cause trouble (PR 42714). */
4522 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4523 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4524 else
4525 *rhs_p = build_constructor (TREE_TYPE (*lhs_p),
4526 NULL);
4528 else
4529 new_rhs = fold_build1_loc (gimple_location (stmt),
4530 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4531 *rhs_p);
4533 else if (REFERENCE_CLASS_P (*rhs_p)
4534 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4535 && !is_gimple_reg (*lhs_p))
4536 /* This can happen when an assignment in between two single field
4537 structures is turned into an assignment in between two pointers to
4538 scalars (PR 42237). */
4539 new_rhs = *rhs_p;
4541 if (new_rhs)
4543 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4544 true, GSI_SAME_STMT);
4546 gimple_assign_set_rhs_from_tree (gsi, tmp);
4549 return true;
4552 return false;
4555 /* Traverse the function body and all modifications as described in
4556 ADJUSTMENTS. Return true iff the CFG has been changed. */
4558 static bool
4559 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4561 bool cfg_changed = false;
4562 basic_block bb;
4564 FOR_EACH_BB (bb)
4566 gimple_stmt_iterator gsi;
4568 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4569 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4571 gsi = gsi_start_bb (bb);
4572 while (!gsi_end_p (gsi))
4574 gimple stmt = gsi_stmt (gsi);
4575 bool modified = false;
4576 tree *t;
4577 unsigned i;
4579 switch (gimple_code (stmt))
4581 case GIMPLE_RETURN:
4582 t = gimple_return_retval_ptr (stmt);
4583 if (*t != NULL_TREE)
4584 modified |= sra_ipa_modify_expr (t, true, adjustments);
4585 break;
4587 case GIMPLE_ASSIGN:
4588 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4589 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4590 break;
4592 case GIMPLE_CALL:
4593 /* Operands must be processed before the lhs. */
4594 for (i = 0; i < gimple_call_num_args (stmt); i++)
4596 t = gimple_call_arg_ptr (stmt, i);
4597 modified |= sra_ipa_modify_expr (t, true, adjustments);
4600 if (gimple_call_lhs (stmt))
4602 t = gimple_call_lhs_ptr (stmt);
4603 modified |= sra_ipa_modify_expr (t, false, adjustments);
4604 modified |= replace_removed_params_ssa_names (stmt,
4605 adjustments);
4607 break;
4609 case GIMPLE_ASM:
4610 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4612 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4613 modified |= sra_ipa_modify_expr (t, true, adjustments);
4615 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4617 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4618 modified |= sra_ipa_modify_expr (t, false, adjustments);
4620 break;
4622 default:
4623 break;
4626 if (modified)
4628 update_stmt (stmt);
4629 if (maybe_clean_eh_stmt (stmt)
4630 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4631 cfg_changed = true;
4633 gsi_next (&gsi);
4637 return cfg_changed;
4640 /* Call gimple_debug_bind_reset_value on all debug statements describing
4641 gimple register parameters that are being removed or replaced. */
4643 static void
4644 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4646 int i, len;
4647 gimple_stmt_iterator *gsip = NULL, gsi;
4649 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR))
4651 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
4652 gsip = &gsi;
4654 len = adjustments.length ();
4655 for (i = 0; i < len; i++)
4657 struct ipa_parm_adjustment *adj;
4658 imm_use_iterator ui;
4659 gimple stmt, def_temp;
4660 tree name, vexpr, copy = NULL_TREE;
4661 use_operand_p use_p;
4663 adj = &adjustments[i];
4664 if (adj->copy_param || !is_gimple_reg (adj->base))
4665 continue;
4666 name = ssa_default_def (cfun, adj->base);
4667 vexpr = NULL;
4668 if (name)
4669 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4671 /* All other users must have been removed by
4672 ipa_sra_modify_function_body. */
4673 gcc_assert (is_gimple_debug (stmt));
4674 if (vexpr == NULL && gsip != NULL)
4676 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4677 vexpr = make_node (DEBUG_EXPR_DECL);
4678 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4679 NULL);
4680 DECL_ARTIFICIAL (vexpr) = 1;
4681 TREE_TYPE (vexpr) = TREE_TYPE (name);
4682 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4683 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4685 if (vexpr)
4687 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4688 SET_USE (use_p, vexpr);
4690 else
4691 gimple_debug_bind_reset_value (stmt);
4692 update_stmt (stmt);
4694 /* Create a VAR_DECL for debug info purposes. */
4695 if (!DECL_IGNORED_P (adj->base))
4697 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4698 VAR_DECL, DECL_NAME (adj->base),
4699 TREE_TYPE (adj->base));
4700 if (DECL_PT_UID_SET_P (adj->base))
4701 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4702 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4703 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4704 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4705 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4706 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4707 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4708 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4709 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4710 SET_DECL_RTL (copy, 0);
4711 TREE_USED (copy) = 1;
4712 DECL_CONTEXT (copy) = current_function_decl;
4713 add_local_decl (cfun, copy);
4714 DECL_CHAIN (copy) =
4715 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4716 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4718 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4720 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4721 if (vexpr)
4722 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4723 else
4724 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4725 NULL);
4726 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4731 /* Return false iff all callers have at least as many actual arguments as there
4732 are formal parameters in the current function. */
4734 static bool
4735 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4736 void *data ATTRIBUTE_UNUSED)
4738 struct cgraph_edge *cs;
4739 for (cs = node->callers; cs; cs = cs->next_caller)
4740 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4741 return true;
4743 return false;
4746 /* Convert all callers of NODE. */
4748 static bool
4749 convert_callers_for_node (struct cgraph_node *node,
4750 void *data)
4752 ipa_parm_adjustment_vec *adjustments = (ipa_parm_adjustment_vec *) data;
4753 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4754 struct cgraph_edge *cs;
4756 for (cs = node->callers; cs; cs = cs->next_caller)
4758 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->symbol.decl));
4760 if (dump_file)
4761 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4762 cs->caller->uid, cs->callee->uid,
4763 xstrdup (cgraph_node_name (cs->caller)),
4764 xstrdup (cgraph_node_name (cs->callee)));
4766 ipa_modify_call_arguments (cs, cs->call_stmt, *adjustments);
4768 pop_cfun ();
4771 for (cs = node->callers; cs; cs = cs->next_caller)
4772 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4773 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->symbol.decl)))
4774 compute_inline_parameters (cs->caller, true);
4775 BITMAP_FREE (recomputed_callers);
4777 return true;
4780 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4782 static void
4783 convert_callers (struct cgraph_node *node, tree old_decl,
4784 ipa_parm_adjustment_vec adjustments)
4786 basic_block this_block;
4788 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4789 &adjustments, false);
4791 if (!encountered_recursive_call)
4792 return;
4794 FOR_EACH_BB (this_block)
4796 gimple_stmt_iterator gsi;
4798 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4800 gimple stmt = gsi_stmt (gsi);
4801 tree call_fndecl;
4802 if (gimple_code (stmt) != GIMPLE_CALL)
4803 continue;
4804 call_fndecl = gimple_call_fndecl (stmt);
4805 if (call_fndecl == old_decl)
4807 if (dump_file)
4808 fprintf (dump_file, "Adjusting recursive call");
4809 gimple_call_set_fndecl (stmt, node->symbol.decl);
4810 ipa_modify_call_arguments (NULL, stmt, adjustments);
4815 return;
4818 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4819 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4821 static bool
4822 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4824 struct cgraph_node *new_node;
4825 bool cfg_changed;
4826 vec<cgraph_edge_p> redirect_callers = collect_callers_of_node (node);
4828 rebuild_cgraph_edges ();
4829 free_dominance_info (CDI_DOMINATORS);
4830 pop_cfun ();
4832 new_node = cgraph_function_versioning (node, redirect_callers,
4833 NULL,
4834 NULL, false, NULL, NULL, "isra");
4835 redirect_callers.release ();
4837 push_cfun (DECL_STRUCT_FUNCTION (new_node->symbol.decl));
4838 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4839 cfg_changed = ipa_sra_modify_function_body (adjustments);
4840 sra_ipa_reset_debug_stmts (adjustments);
4841 convert_callers (new_node, node->symbol.decl, adjustments);
4842 cgraph_make_node_local (new_node);
4843 return cfg_changed;
4846 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4847 attributes, return true otherwise. NODE is the cgraph node of the current
4848 function. */
4850 static bool
4851 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4853 if (!cgraph_node_can_be_local_p (node))
4855 if (dump_file)
4856 fprintf (dump_file, "Function not local to this compilation unit.\n");
4857 return false;
4860 if (!node->local.can_change_signature)
4862 if (dump_file)
4863 fprintf (dump_file, "Function can not change signature.\n");
4864 return false;
4867 if (!tree_versionable_function_p (node->symbol.decl))
4869 if (dump_file)
4870 fprintf (dump_file, "Function is not versionable.\n");
4871 return false;
4874 if (DECL_VIRTUAL_P (current_function_decl))
4876 if (dump_file)
4877 fprintf (dump_file, "Function is a virtual method.\n");
4878 return false;
4881 if ((DECL_COMDAT (node->symbol.decl) || DECL_EXTERNAL (node->symbol.decl))
4882 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4884 if (dump_file)
4885 fprintf (dump_file, "Function too big to be made truly local.\n");
4886 return false;
4889 if (!node->callers)
4891 if (dump_file)
4892 fprintf (dump_file,
4893 "Function has no callers in this compilation unit.\n");
4894 return false;
4897 if (cfun->stdarg)
4899 if (dump_file)
4900 fprintf (dump_file, "Function uses stdarg. \n");
4901 return false;
4904 if (TYPE_ATTRIBUTES (TREE_TYPE (node->symbol.decl)))
4905 return false;
4907 return true;
4910 /* Perform early interprocedural SRA. */
4912 static unsigned int
4913 ipa_early_sra (void)
4915 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4916 ipa_parm_adjustment_vec adjustments;
4917 int ret = 0;
4919 if (!ipa_sra_preliminary_function_checks (node))
4920 return 0;
4922 sra_initialize ();
4923 sra_mode = SRA_MODE_EARLY_IPA;
4925 if (!find_param_candidates ())
4927 if (dump_file)
4928 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4929 goto simple_out;
4932 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4933 NULL, true))
4935 if (dump_file)
4936 fprintf (dump_file, "There are callers with insufficient number of "
4937 "arguments.\n");
4938 goto simple_out;
4941 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4942 func_param_count
4943 * last_basic_block_for_function (cfun));
4944 final_bbs = BITMAP_ALLOC (NULL);
4946 scan_function ();
4947 if (encountered_apply_args)
4949 if (dump_file)
4950 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4951 goto out;
4954 if (encountered_unchangable_recursive_call)
4956 if (dump_file)
4957 fprintf (dump_file, "Function calls itself with insufficient "
4958 "number of arguments.\n");
4959 goto out;
4962 adjustments = analyze_all_param_acesses ();
4963 if (!adjustments.exists ())
4964 goto out;
4965 if (dump_file)
4966 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4968 if (modify_function (node, adjustments))
4969 ret = TODO_update_ssa | TODO_cleanup_cfg;
4970 else
4971 ret = TODO_update_ssa;
4972 adjustments.release ();
4974 statistics_counter_event (cfun, "Unused parameters deleted",
4975 sra_stats.deleted_unused_parameters);
4976 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4977 sra_stats.scalar_by_ref_to_by_val);
4978 statistics_counter_event (cfun, "Aggregate parameters broken up",
4979 sra_stats.aggregate_params_reduced);
4980 statistics_counter_event (cfun, "Aggregate parameter components created",
4981 sra_stats.param_reductions_created);
4983 out:
4984 BITMAP_FREE (final_bbs);
4985 free (bb_dereferences);
4986 simple_out:
4987 sra_deinitialize ();
4988 return ret;
4991 /* Return if early ipa sra shall be performed. */
4992 static bool
4993 ipa_early_sra_gate (void)
4995 return flag_ipa_sra && dbg_cnt (eipa_sra);
4998 struct gimple_opt_pass pass_early_ipa_sra =
5001 GIMPLE_PASS,
5002 "eipa_sra", /* name */
5003 OPTGROUP_NONE, /* optinfo_flags */
5004 ipa_early_sra_gate, /* gate */
5005 ipa_early_sra, /* execute */
5006 NULL, /* sub */
5007 NULL, /* next */
5008 0, /* static_pass_number */
5009 TV_IPA_SRA, /* tv_id */
5010 0, /* properties_required */
5011 0, /* properties_provided */
5012 0, /* properties_destroyed */
5013 0, /* todo_flags_start */
5014 TODO_dump_symtab /* todo_flags_finish */