re PR bootstrap/51346 (LTO bootstrap failed with bootstrap-profiled)
[official-gcc.git] / gcc / tree-sra.c
blob472ad5d6f5d0c9092b3f0fcdea2c3e7eae2249d6
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 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 "ipa-prop.h"
84 #include "tree-pretty-print.h"
85 #include "statistics.h"
86 #include "tree-dump.h"
87 #include "timevar.h"
88 #include "params.h"
89 #include "target.h"
90 #include "flags.h"
91 #include "dbgcnt.h"
92 #include "tree-inline.h"
93 #include "gimple-pretty-print.h"
94 #include "ipa-inline.h"
96 /* Enumeration of all aggregate reductions we can do. */
97 enum sra_mode { SRA_MODE_EARLY_IPA, /* early call regularization */
98 SRA_MODE_EARLY_INTRA, /* early intraprocedural SRA */
99 SRA_MODE_INTRA }; /* late intraprocedural SRA */
101 /* Global variable describing which aggregate reduction we are performing at
102 the moment. */
103 static enum sra_mode sra_mode;
105 struct assign_link;
107 /* ACCESS represents each access to an aggregate variable (as a whole or a
108 part). It can also represent a group of accesses that refer to exactly the
109 same fragment of an aggregate (i.e. those that have exactly the same offset
110 and size). Such representatives for a single aggregate, once determined,
111 are linked in a linked list and have the group fields set.
113 Moreover, when doing intraprocedural SRA, a tree is built from those
114 representatives (by the means of first_child and next_sibling pointers), in
115 which all items in a subtree are "within" the root, i.e. their offset is
116 greater or equal to offset of the root and offset+size is smaller or equal
117 to offset+size of the root. Children of an access are sorted by offset.
119 Note that accesses to parts of vector and complex number types always
120 represented by an access to the whole complex number or a vector. It is a
121 duty of the modifying functions to replace them appropriately. */
123 struct access
125 /* Values returned by `get_ref_base_and_extent' for each component reference
126 If EXPR isn't a component reference just set `BASE = EXPR', `OFFSET = 0',
127 `SIZE = TREE_SIZE (TREE_TYPE (expr))'. */
128 HOST_WIDE_INT offset;
129 HOST_WIDE_INT size;
130 tree base;
132 /* Expression. It is context dependent so do not use it to create new
133 expressions to access the original aggregate. See PR 42154 for a
134 testcase. */
135 tree expr;
136 /* Type. */
137 tree type;
139 /* The statement this access belongs to. */
140 gimple stmt;
142 /* Next group representative for this aggregate. */
143 struct access *next_grp;
145 /* Pointer to the group representative. Pointer to itself if the struct is
146 the representative. */
147 struct access *group_representative;
149 /* If this access has any children (in terms of the definition above), this
150 points to the first one. */
151 struct access *first_child;
153 /* In intraprocedural SRA, pointer to the next sibling in the access tree as
154 described above. In IPA-SRA this is a pointer to the next access
155 belonging to the same group (having the same representative). */
156 struct access *next_sibling;
158 /* Pointers to the first and last element in the linked list of assign
159 links. */
160 struct assign_link *first_link, *last_link;
162 /* Pointer to the next access in the work queue. */
163 struct access *next_queued;
165 /* Replacement variable for this access "region." Never to be accessed
166 directly, always only by the means of get_access_replacement() and only
167 when grp_to_be_replaced flag is set. */
168 tree replacement_decl;
170 /* Is this particular access write access? */
171 unsigned write : 1;
173 /* Is this access an access to a non-addressable field? */
174 unsigned non_addressable : 1;
176 /* Is this access currently in the work queue? */
177 unsigned grp_queued : 1;
179 /* Does this group contain a write access? This flag is propagated down the
180 access tree. */
181 unsigned grp_write : 1;
183 /* Does this group contain a read access? This flag is propagated down the
184 access tree. */
185 unsigned grp_read : 1;
187 /* Does this group contain a read access that comes from an assignment
188 statement? This flag is propagated down the access tree. */
189 unsigned grp_assignment_read : 1;
191 /* Does this group contain a write access that comes from an assignment
192 statement? This flag is propagated down the access tree. */
193 unsigned grp_assignment_write : 1;
195 /* Does this group contain a read access through a scalar type? This flag is
196 not propagated in the access tree in any direction. */
197 unsigned grp_scalar_read : 1;
199 /* Does this group contain a write access through a scalar type? This flag
200 is not propagated in the access tree in any direction. */
201 unsigned grp_scalar_write : 1;
203 /* Is this access an artificial one created to scalarize some record
204 entirely? */
205 unsigned grp_total_scalarization : 1;
207 /* Other passes of the analysis use this bit to make function
208 analyze_access_subtree create scalar replacements for this group if
209 possible. */
210 unsigned grp_hint : 1;
212 /* Is the subtree rooted in this access fully covered by scalar
213 replacements? */
214 unsigned grp_covered : 1;
216 /* If set to true, this access and all below it in an access tree must not be
217 scalarized. */
218 unsigned grp_unscalarizable_region : 1;
220 /* Whether data have been written to parts of the aggregate covered by this
221 access which is not to be scalarized. This flag is propagated up in the
222 access tree. */
223 unsigned grp_unscalarized_data : 1;
225 /* Does this access and/or group contain a write access through a
226 BIT_FIELD_REF? */
227 unsigned grp_partial_lhs : 1;
229 /* Set when a scalar replacement should be created for this variable. We do
230 the decision and creation at different places because create_tmp_var
231 cannot be called from within FOR_EACH_REFERENCED_VAR. */
232 unsigned grp_to_be_replaced : 1;
234 /* Should TREE_NO_WARNING of a replacement be set? */
235 unsigned grp_no_warning : 1;
237 /* Is it possible that the group refers to data which might be (directly or
238 otherwise) modified? */
239 unsigned grp_maybe_modified : 1;
241 /* Set when this is a representative of a pointer to scalar (i.e. by
242 reference) parameter which we consider for turning into a plain scalar
243 (i.e. a by value parameter). */
244 unsigned grp_scalar_ptr : 1;
246 /* Set when we discover that this pointer is not safe to dereference in the
247 caller. */
248 unsigned grp_not_necessarilly_dereferenced : 1;
251 typedef struct access *access_p;
253 DEF_VEC_P (access_p);
254 DEF_VEC_ALLOC_P (access_p, heap);
256 /* Alloc pool for allocating access structures. */
257 static alloc_pool access_pool;
259 /* A structure linking lhs and rhs accesses from an aggregate assignment. They
260 are used to propagate subaccesses from rhs to lhs as long as they don't
261 conflict with what is already there. */
262 struct assign_link
264 struct access *lacc, *racc;
265 struct assign_link *next;
268 /* Alloc pool for allocating assign link structures. */
269 static alloc_pool link_pool;
271 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
272 static struct pointer_map_t *base_access_vec;
274 /* Bitmap of candidates. */
275 static bitmap candidate_bitmap;
277 /* Bitmap of candidates which we should try to entirely scalarize away and
278 those which cannot be (because they are and need be used as a whole). */
279 static bitmap should_scalarize_away_bitmap, cannot_scalarize_away_bitmap;
281 /* Obstack for creation of fancy names. */
282 static struct obstack name_obstack;
284 /* Head of a linked list of accesses that need to have its subaccesses
285 propagated to their assignment counterparts. */
286 static struct access *work_queue_head;
288 /* Number of parameters of the analyzed function when doing early ipa SRA. */
289 static int func_param_count;
291 /* scan_function sets the following to true if it encounters a call to
292 __builtin_apply_args. */
293 static bool encountered_apply_args;
295 /* Set by scan_function when it finds a recursive call. */
296 static bool encountered_recursive_call;
298 /* Set by scan_function when it finds a recursive call with less actual
299 arguments than formal parameters.. */
300 static bool encountered_unchangable_recursive_call;
302 /* This is a table in which for each basic block and parameter there is a
303 distance (offset + size) in that parameter which is dereferenced and
304 accessed in that BB. */
305 static HOST_WIDE_INT *bb_dereferences;
306 /* Bitmap of BBs that can cause the function to "stop" progressing by
307 returning, throwing externally, looping infinitely or calling a function
308 which might abort etc.. */
309 static bitmap final_bbs;
311 /* Representative of no accesses at all. */
312 static struct access no_accesses_representant;
314 /* Predicate to test the special value. */
316 static inline bool
317 no_accesses_p (struct access *access)
319 return access == &no_accesses_representant;
322 /* Dump contents of ACCESS to file F in a human friendly way. If GRP is true,
323 representative fields are dumped, otherwise those which only describe the
324 individual access are. */
326 static struct
328 /* Number of processed aggregates is readily available in
329 analyze_all_variable_accesses and so is not stored here. */
331 /* Number of created scalar replacements. */
332 int replacements;
334 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
335 expression. */
336 int exprs;
338 /* Number of statements created by generate_subtree_copies. */
339 int subtree_copies;
341 /* Number of statements created by load_assign_lhs_subreplacements. */
342 int subreplacements;
344 /* Number of times sra_modify_assign has deleted a statement. */
345 int deleted;
347 /* Number of times sra_modify_assign has to deal with subaccesses of LHS and
348 RHS reparately due to type conversions or nonexistent matching
349 references. */
350 int separate_lhs_rhs_handling;
352 /* Number of parameters that were removed because they were unused. */
353 int deleted_unused_parameters;
355 /* Number of scalars passed as parameters by reference that have been
356 converted to be passed by value. */
357 int scalar_by_ref_to_by_val;
359 /* Number of aggregate parameters that were replaced by one or more of their
360 components. */
361 int aggregate_params_reduced;
363 /* Numbber of components created when splitting aggregate parameters. */
364 int param_reductions_created;
365 } sra_stats;
367 static void
368 dump_access (FILE *f, struct access *access, bool grp)
370 fprintf (f, "access { ");
371 fprintf (f, "base = (%d)'", DECL_UID (access->base));
372 print_generic_expr (f, access->base, 0);
373 fprintf (f, "', offset = " HOST_WIDE_INT_PRINT_DEC, access->offset);
374 fprintf (f, ", size = " HOST_WIDE_INT_PRINT_DEC, access->size);
375 fprintf (f, ", expr = ");
376 print_generic_expr (f, access->expr, 0);
377 fprintf (f, ", type = ");
378 print_generic_expr (f, access->type, 0);
379 if (grp)
380 fprintf (f, ", grp_read = %d, grp_write = %d, grp_assignment_read = %d, "
381 "grp_assignment_write = %d, grp_scalar_read = %d, "
382 "grp_scalar_write = %d, grp_total_scalarization = %d, "
383 "grp_hint = %d, grp_covered = %d, "
384 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
385 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
386 "grp_maybe_modified = %d, "
387 "grp_not_necessarilly_dereferenced = %d\n",
388 access->grp_read, access->grp_write, access->grp_assignment_read,
389 access->grp_assignment_write, access->grp_scalar_read,
390 access->grp_scalar_write, access->grp_total_scalarization,
391 access->grp_hint, access->grp_covered,
392 access->grp_unscalarizable_region, access->grp_unscalarized_data,
393 access->grp_partial_lhs, access->grp_to_be_replaced,
394 access->grp_maybe_modified,
395 access->grp_not_necessarilly_dereferenced);
396 else
397 fprintf (f, ", write = %d, grp_total_scalarization = %d, "
398 "grp_partial_lhs = %d\n",
399 access->write, access->grp_total_scalarization,
400 access->grp_partial_lhs);
403 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
405 static void
406 dump_access_tree_1 (FILE *f, struct access *access, int level)
410 int i;
412 for (i = 0; i < level; i++)
413 fputs ("* ", dump_file);
415 dump_access (f, access, true);
417 if (access->first_child)
418 dump_access_tree_1 (f, access->first_child, level + 1);
420 access = access->next_sibling;
422 while (access);
425 /* Dump all access trees for a variable, given the pointer to the first root in
426 ACCESS. */
428 static void
429 dump_access_tree (FILE *f, struct access *access)
431 for (; access; access = access->next_grp)
432 dump_access_tree_1 (f, access, 0);
435 /* Return true iff ACC is non-NULL and has subaccesses. */
437 static inline bool
438 access_has_children_p (struct access *acc)
440 return acc && acc->first_child;
443 /* Return a vector of pointers to accesses for the variable given in BASE or
444 NULL if there is none. */
446 static VEC (access_p, heap) *
447 get_base_access_vector (tree base)
449 void **slot;
451 slot = pointer_map_contains (base_access_vec, base);
452 if (!slot)
453 return NULL;
454 else
455 return *(VEC (access_p, heap) **) slot;
458 /* Find an access with required OFFSET and SIZE in a subtree of accesses rooted
459 in ACCESS. Return NULL if it cannot be found. */
461 static struct access *
462 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
463 HOST_WIDE_INT size)
465 while (access && (access->offset != offset || access->size != size))
467 struct access *child = access->first_child;
469 while (child && (child->offset + child->size <= offset))
470 child = child->next_sibling;
471 access = child;
474 return access;
477 /* Return the first group representative for DECL or NULL if none exists. */
479 static struct access *
480 get_first_repr_for_decl (tree base)
482 VEC (access_p, heap) *access_vec;
484 access_vec = get_base_access_vector (base);
485 if (!access_vec)
486 return NULL;
488 return VEC_index (access_p, access_vec, 0);
491 /* Find an access representative for the variable BASE and given OFFSET and
492 SIZE. Requires that access trees have already been built. Return NULL if
493 it cannot be found. */
495 static struct access *
496 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
497 HOST_WIDE_INT size)
499 struct access *access;
501 access = get_first_repr_for_decl (base);
502 while (access && (access->offset + access->size <= offset))
503 access = access->next_grp;
504 if (!access)
505 return NULL;
507 return find_access_in_subtree (access, offset, size);
510 /* Add LINK to the linked list of assign links of RACC. */
511 static void
512 add_link_to_rhs (struct access *racc, struct assign_link *link)
514 gcc_assert (link->racc == racc);
516 if (!racc->first_link)
518 gcc_assert (!racc->last_link);
519 racc->first_link = link;
521 else
522 racc->last_link->next = link;
524 racc->last_link = link;
525 link->next = NULL;
528 /* Move all link structures in their linked list in OLD_RACC to the linked list
529 in NEW_RACC. */
530 static void
531 relink_to_new_repr (struct access *new_racc, struct access *old_racc)
533 if (!old_racc->first_link)
535 gcc_assert (!old_racc->last_link);
536 return;
539 if (new_racc->first_link)
541 gcc_assert (!new_racc->last_link->next);
542 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
544 new_racc->last_link->next = old_racc->first_link;
545 new_racc->last_link = old_racc->last_link;
547 else
549 gcc_assert (!new_racc->last_link);
551 new_racc->first_link = old_racc->first_link;
552 new_racc->last_link = old_racc->last_link;
554 old_racc->first_link = old_racc->last_link = NULL;
557 /* Add ACCESS to the work queue (which is actually a stack). */
559 static void
560 add_access_to_work_queue (struct access *access)
562 if (!access->grp_queued)
564 gcc_assert (!access->next_queued);
565 access->next_queued = work_queue_head;
566 access->grp_queued = 1;
567 work_queue_head = access;
571 /* Pop an access from the work queue, and return it, assuming there is one. */
573 static struct access *
574 pop_access_from_work_queue (void)
576 struct access *access = work_queue_head;
578 work_queue_head = access->next_queued;
579 access->next_queued = NULL;
580 access->grp_queued = 0;
581 return access;
585 /* Allocate necessary structures. */
587 static void
588 sra_initialize (void)
590 candidate_bitmap = BITMAP_ALLOC (NULL);
591 should_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
592 cannot_scalarize_away_bitmap = BITMAP_ALLOC (NULL);
593 gcc_obstack_init (&name_obstack);
594 access_pool = create_alloc_pool ("SRA accesses", sizeof (struct access), 16);
595 link_pool = create_alloc_pool ("SRA links", sizeof (struct assign_link), 16);
596 base_access_vec = pointer_map_create ();
597 memset (&sra_stats, 0, sizeof (sra_stats));
598 encountered_apply_args = false;
599 encountered_recursive_call = false;
600 encountered_unchangable_recursive_call = false;
603 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
605 static bool
606 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
607 void *data ATTRIBUTE_UNUSED)
609 VEC (access_p, heap) *access_vec;
610 access_vec = (VEC (access_p, heap) *) *value;
611 VEC_free (access_p, heap, access_vec);
613 return true;
616 /* Deallocate all general structures. */
618 static void
619 sra_deinitialize (void)
621 BITMAP_FREE (candidate_bitmap);
622 BITMAP_FREE (should_scalarize_away_bitmap);
623 BITMAP_FREE (cannot_scalarize_away_bitmap);
624 free_alloc_pool (access_pool);
625 free_alloc_pool (link_pool);
626 obstack_free (&name_obstack, NULL);
628 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
629 pointer_map_destroy (base_access_vec);
632 /* Remove DECL from candidates for SRA and write REASON to the dump file if
633 there is one. */
634 static void
635 disqualify_candidate (tree decl, const char *reason)
637 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
639 if (dump_file && (dump_flags & TDF_DETAILS))
641 fprintf (dump_file, "! Disqualifying ");
642 print_generic_expr (dump_file, decl, 0);
643 fprintf (dump_file, " - %s\n", reason);
647 /* Return true iff the type contains a field or an element which does not allow
648 scalarization. */
650 static bool
651 type_internals_preclude_sra_p (tree type, const char **msg)
653 tree fld;
654 tree et;
656 switch (TREE_CODE (type))
658 case RECORD_TYPE:
659 case UNION_TYPE:
660 case QUAL_UNION_TYPE:
661 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
662 if (TREE_CODE (fld) == FIELD_DECL)
664 tree ft = TREE_TYPE (fld);
666 if (TREE_THIS_VOLATILE (fld))
668 *msg = "volatile structure field";
669 return true;
671 if (!DECL_FIELD_OFFSET (fld))
673 *msg = "no structure field offset";
674 return true;
676 if (!DECL_SIZE (fld))
678 *msg = "zero structure field size";
679 return true;
681 if (!host_integerp (DECL_FIELD_OFFSET (fld), 1))
683 *msg = "structure field offset not fixed";
684 return true;
686 if (!host_integerp (DECL_SIZE (fld), 1))
688 *msg = "structure field size not fixed";
689 return true;
691 if (AGGREGATE_TYPE_P (ft)
692 && int_bit_position (fld) % BITS_PER_UNIT != 0)
694 *msg = "structure field is bit field";
695 return true;
698 if (AGGREGATE_TYPE_P (ft) && type_internals_preclude_sra_p (ft, msg))
699 return true;
702 return false;
704 case ARRAY_TYPE:
705 et = TREE_TYPE (type);
707 if (TYPE_VOLATILE (et))
709 *msg = "element type is volatile";
710 return true;
713 if (AGGREGATE_TYPE_P (et) && type_internals_preclude_sra_p (et, msg))
714 return true;
716 return false;
718 default:
719 return false;
723 /* If T is an SSA_NAME, return NULL if it is not a default def or return its
724 base variable if it is. Return T if it is not an SSA_NAME. */
726 static tree
727 get_ssa_base_param (tree t)
729 if (TREE_CODE (t) == SSA_NAME)
731 if (SSA_NAME_IS_DEFAULT_DEF (t))
732 return SSA_NAME_VAR (t);
733 else
734 return NULL_TREE;
736 return t;
739 /* Mark a dereference of BASE of distance DIST in a basic block tht STMT
740 belongs to, unless the BB has already been marked as a potentially
741 final. */
743 static void
744 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
746 basic_block bb = gimple_bb (stmt);
747 int idx, parm_index = 0;
748 tree parm;
750 if (bitmap_bit_p (final_bbs, bb->index))
751 return;
753 for (parm = DECL_ARGUMENTS (current_function_decl);
754 parm && parm != base;
755 parm = DECL_CHAIN (parm))
756 parm_index++;
758 gcc_assert (parm_index < func_param_count);
760 idx = bb->index * func_param_count + parm_index;
761 if (bb_dereferences[idx] < dist)
762 bb_dereferences[idx] = dist;
765 /* Allocate an access structure for BASE, OFFSET and SIZE, clear it, fill in
766 the three fields. Also add it to the vector of accesses corresponding to
767 the base. Finally, return the new access. */
769 static struct access *
770 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
772 VEC (access_p, heap) *vec;
773 struct access *access;
774 void **slot;
776 access = (struct access *) pool_alloc (access_pool);
777 memset (access, 0, sizeof (struct access));
778 access->base = base;
779 access->offset = offset;
780 access->size = size;
782 slot = pointer_map_contains (base_access_vec, base);
783 if (slot)
784 vec = (VEC (access_p, heap) *) *slot;
785 else
786 vec = VEC_alloc (access_p, heap, 32);
788 VEC_safe_push (access_p, heap, vec, access);
790 *((struct VEC (access_p,heap) **)
791 pointer_map_insert (base_access_vec, base)) = vec;
793 return access;
796 /* Create and insert access for EXPR. Return created access, or NULL if it is
797 not possible. */
799 static struct access *
800 create_access (tree expr, gimple stmt, bool write)
802 struct access *access;
803 HOST_WIDE_INT offset, size, max_size;
804 tree base = expr;
805 bool ptr, unscalarizable_region = false;
807 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
809 if (sra_mode == SRA_MODE_EARLY_IPA
810 && TREE_CODE (base) == MEM_REF)
812 base = get_ssa_base_param (TREE_OPERAND (base, 0));
813 if (!base)
814 return NULL;
815 ptr = true;
817 else
818 ptr = false;
820 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
821 return NULL;
823 if (sra_mode == SRA_MODE_EARLY_IPA)
825 if (size < 0 || size != max_size)
827 disqualify_candidate (base, "Encountered a variable sized access.");
828 return NULL;
830 if (TREE_CODE (expr) == COMPONENT_REF
831 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
833 disqualify_candidate (base, "Encountered a bit-field access.");
834 return NULL;
836 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
838 if (ptr)
839 mark_parm_dereference (base, offset + size, stmt);
841 else
843 if (size != max_size)
845 size = max_size;
846 unscalarizable_region = true;
848 if (size < 0)
850 disqualify_candidate (base, "Encountered an unconstrained access.");
851 return NULL;
855 access = create_access_1 (base, offset, size);
856 access->expr = expr;
857 access->type = TREE_TYPE (expr);
858 access->write = write;
859 access->grp_unscalarizable_region = unscalarizable_region;
860 access->stmt = stmt;
862 if (TREE_CODE (expr) == COMPONENT_REF
863 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
864 access->non_addressable = 1;
866 return access;
870 /* Return true iff TYPE is a RECORD_TYPE with fields that are either of gimple
871 register types or (recursively) records with only these two kinds of fields.
872 It also returns false if any of these records contains a bit-field. */
874 static bool
875 type_consists_of_records_p (tree type)
877 tree fld;
879 if (TREE_CODE (type) != RECORD_TYPE)
880 return false;
882 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
883 if (TREE_CODE (fld) == FIELD_DECL)
885 tree ft = TREE_TYPE (fld);
887 if (DECL_BIT_FIELD (fld))
888 return false;
890 if (!is_gimple_reg_type (ft)
891 && !type_consists_of_records_p (ft))
892 return false;
895 return true;
898 /* Create total_scalarization accesses for all scalar type fields in DECL that
899 must be of a RECORD_TYPE conforming to type_consists_of_records_p. BASE
900 must be the top-most VAR_DECL representing the variable, OFFSET must be the
901 offset of DECL within BASE. REF must be the memory reference expression for
902 the given decl. */
904 static void
905 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
906 tree ref)
908 tree fld, decl_type = TREE_TYPE (decl);
910 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
911 if (TREE_CODE (fld) == FIELD_DECL)
913 HOST_WIDE_INT pos = offset + int_bit_position (fld);
914 tree ft = TREE_TYPE (fld);
915 tree nref = build3 (COMPONENT_REF, TREE_TYPE (fld), ref, fld,
916 NULL_TREE);
918 if (is_gimple_reg_type (ft))
920 struct access *access;
921 HOST_WIDE_INT size;
923 size = tree_low_cst (DECL_SIZE (fld), 1);
924 access = create_access_1 (base, pos, size);
925 access->expr = nref;
926 access->type = ft;
927 access->grp_total_scalarization = 1;
928 /* Accesses for intraprocedural SRA can have their stmt NULL. */
930 else
931 completely_scalarize_record (base, fld, pos, nref);
935 /* Create total_scalarization accesses for all scalar type fields in VAR and
936 for VAR a a whole. VAR must be of a RECORD_TYPE conforming to
937 type_consists_of_records_p. */
939 static void
940 completely_scalarize_var (tree var)
942 HOST_WIDE_INT size = tree_low_cst (DECL_SIZE (var), 1);
943 struct access *access;
945 access = create_access_1 (var, 0, size);
946 access->expr = var;
947 access->type = TREE_TYPE (var);
948 access->grp_total_scalarization = 1;
950 completely_scalarize_record (var, var, 0, var);
953 /* Search the given tree for a declaration by skipping handled components and
954 exclude it from the candidates. */
956 static void
957 disqualify_base_of_expr (tree t, const char *reason)
959 t = get_base_address (t);
960 if (sra_mode == SRA_MODE_EARLY_IPA
961 && TREE_CODE (t) == MEM_REF)
962 t = get_ssa_base_param (TREE_OPERAND (t, 0));
964 if (t && DECL_P (t))
965 disqualify_candidate (t, reason);
968 /* Scan expression EXPR and create access structures for all accesses to
969 candidates for scalarization. Return the created access or NULL if none is
970 created. */
972 static struct access *
973 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
975 struct access *ret = NULL;
976 bool partial_ref;
978 if (TREE_CODE (expr) == BIT_FIELD_REF
979 || TREE_CODE (expr) == IMAGPART_EXPR
980 || TREE_CODE (expr) == REALPART_EXPR)
982 expr = TREE_OPERAND (expr, 0);
983 partial_ref = true;
985 else
986 partial_ref = false;
988 /* We need to dive through V_C_Es in order to get the size of its parameter
989 and not the result type. Ada produces such statements. We are also
990 capable of handling the topmost V_C_E but not any of those buried in other
991 handled components. */
992 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
993 expr = TREE_OPERAND (expr, 0);
995 if (contains_view_convert_expr_p (expr))
997 disqualify_base_of_expr (expr, "V_C_E under a different handled "
998 "component.");
999 return NULL;
1002 switch (TREE_CODE (expr))
1004 case MEM_REF:
1005 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
1006 && sra_mode != SRA_MODE_EARLY_IPA)
1007 return NULL;
1008 /* fall through */
1009 case VAR_DECL:
1010 case PARM_DECL:
1011 case RESULT_DECL:
1012 case COMPONENT_REF:
1013 case ARRAY_REF:
1014 case ARRAY_RANGE_REF:
1015 ret = create_access (expr, stmt, write);
1016 break;
1018 default:
1019 break;
1022 if (write && partial_ref && ret)
1023 ret->grp_partial_lhs = 1;
1025 return ret;
1028 /* Scan expression EXPR and create access structures for all accesses to
1029 candidates for scalarization. Return true if any access has been inserted.
1030 STMT must be the statement from which the expression is taken, WRITE must be
1031 true if the expression is a store and false otherwise. */
1033 static bool
1034 build_access_from_expr (tree expr, gimple stmt, bool write)
1036 struct access *access;
1038 access = build_access_from_expr_1 (expr, stmt, write);
1039 if (access)
1041 /* This means the aggregate is accesses as a whole in a way other than an
1042 assign statement and thus cannot be removed even if we had a scalar
1043 replacement for everything. */
1044 if (cannot_scalarize_away_bitmap)
1045 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1046 return true;
1048 return false;
1051 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1052 modes in which it matters, return true iff they have been disqualified. RHS
1053 may be NULL, in that case ignore it. If we scalarize an aggregate in
1054 intra-SRA we may need to add statements after each statement. This is not
1055 possible if a statement unconditionally has to end the basic block. */
1056 static bool
1057 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1059 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1060 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1062 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1063 if (rhs)
1064 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1065 return true;
1067 return false;
1070 /* Return true iff type of EXP is not sufficiently aligned. */
1072 static bool
1073 tree_non_mode_aligned_mem_p (tree exp)
1075 enum machine_mode mode = TYPE_MODE (TREE_TYPE (exp));
1076 unsigned int align;
1078 if (TREE_CODE (exp) == VIEW_CONVERT_EXPR)
1079 exp = TREE_OPERAND (exp, 0);
1081 if (TREE_CODE (exp) == SSA_NAME
1082 || TREE_CODE (exp) == MEM_REF
1083 || mode == BLKmode
1084 || is_gimple_min_invariant (exp)
1085 || !STRICT_ALIGNMENT)
1086 return false;
1088 align = get_object_alignment (exp);
1089 if (GET_MODE_ALIGNMENT (mode) > align)
1090 return true;
1092 return false;
1095 /* Scan expressions occuring in STMT, create access structures for all accesses
1096 to candidates for scalarization and remove those candidates which occur in
1097 statements or expressions that prevent them from being split apart. Return
1098 true if any access has been inserted. */
1100 static bool
1101 build_accesses_from_assign (gimple stmt)
1103 tree lhs, rhs;
1104 struct access *lacc, *racc;
1106 if (!gimple_assign_single_p (stmt)
1107 /* Scope clobbers don't influence scalarization. */
1108 || gimple_clobber_p (stmt))
1109 return false;
1111 lhs = gimple_assign_lhs (stmt);
1112 rhs = gimple_assign_rhs1 (stmt);
1114 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1115 return false;
1117 racc = build_access_from_expr_1 (rhs, stmt, false);
1118 lacc = build_access_from_expr_1 (lhs, stmt, true);
1120 if (lacc)
1122 lacc->grp_assignment_write = 1;
1123 lacc->grp_unscalarizable_region |= tree_non_mode_aligned_mem_p (rhs);
1126 if (racc)
1128 racc->grp_assignment_read = 1;
1129 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1130 && !is_gimple_reg_type (racc->type))
1131 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1132 racc->grp_unscalarizable_region |= tree_non_mode_aligned_mem_p (lhs);
1135 if (lacc && racc
1136 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1137 && !lacc->grp_unscalarizable_region
1138 && !racc->grp_unscalarizable_region
1139 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1140 /* FIXME: Turn the following line into an assert after PR 40058 is
1141 fixed. */
1142 && lacc->size == racc->size
1143 && useless_type_conversion_p (lacc->type, racc->type))
1145 struct assign_link *link;
1147 link = (struct assign_link *) pool_alloc (link_pool);
1148 memset (link, 0, sizeof (struct assign_link));
1150 link->lacc = lacc;
1151 link->racc = racc;
1153 add_link_to_rhs (racc, link);
1156 return lacc || racc;
1159 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1160 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1162 static bool
1163 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1164 void *data ATTRIBUTE_UNUSED)
1166 op = get_base_address (op);
1167 if (op
1168 && DECL_P (op))
1169 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1171 return false;
1174 /* Return true iff callsite CALL has at least as many actual arguments as there
1175 are formal parameters of the function currently processed by IPA-SRA. */
1177 static inline bool
1178 callsite_has_enough_arguments_p (gimple call)
1180 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1183 /* Scan function and look for interesting expressions and create access
1184 structures for them. Return true iff any access is created. */
1186 static bool
1187 scan_function (void)
1189 basic_block bb;
1190 bool ret = false;
1192 FOR_EACH_BB (bb)
1194 gimple_stmt_iterator gsi;
1195 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1197 gimple stmt = gsi_stmt (gsi);
1198 tree t;
1199 unsigned i;
1201 if (final_bbs && stmt_can_throw_external (stmt))
1202 bitmap_set_bit (final_bbs, bb->index);
1203 switch (gimple_code (stmt))
1205 case GIMPLE_RETURN:
1206 t = gimple_return_retval (stmt);
1207 if (t != NULL_TREE)
1208 ret |= build_access_from_expr (t, stmt, false);
1209 if (final_bbs)
1210 bitmap_set_bit (final_bbs, bb->index);
1211 break;
1213 case GIMPLE_ASSIGN:
1214 ret |= build_accesses_from_assign (stmt);
1215 break;
1217 case GIMPLE_CALL:
1218 for (i = 0; i < gimple_call_num_args (stmt); i++)
1219 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1220 stmt, false);
1222 if (sra_mode == SRA_MODE_EARLY_IPA)
1224 tree dest = gimple_call_fndecl (stmt);
1225 int flags = gimple_call_flags (stmt);
1227 if (dest)
1229 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1230 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1231 encountered_apply_args = true;
1232 if (cgraph_get_node (dest)
1233 == cgraph_get_node (current_function_decl))
1235 encountered_recursive_call = true;
1236 if (!callsite_has_enough_arguments_p (stmt))
1237 encountered_unchangable_recursive_call = true;
1241 if (final_bbs
1242 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1243 bitmap_set_bit (final_bbs, bb->index);
1246 t = gimple_call_lhs (stmt);
1247 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1248 ret |= build_access_from_expr (t, stmt, true);
1249 break;
1251 case GIMPLE_ASM:
1252 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1253 asm_visit_addr);
1254 if (final_bbs)
1255 bitmap_set_bit (final_bbs, bb->index);
1257 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1259 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1260 ret |= build_access_from_expr (t, stmt, false);
1262 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1264 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1265 ret |= build_access_from_expr (t, stmt, true);
1267 break;
1269 default:
1270 break;
1275 return ret;
1278 /* Helper of QSORT function. There are pointers to accesses in the array. An
1279 access is considered smaller than another if it has smaller offset or if the
1280 offsets are the same but is size is bigger. */
1282 static int
1283 compare_access_positions (const void *a, const void *b)
1285 const access_p *fp1 = (const access_p *) a;
1286 const access_p *fp2 = (const access_p *) b;
1287 const access_p f1 = *fp1;
1288 const access_p f2 = *fp2;
1290 if (f1->offset != f2->offset)
1291 return f1->offset < f2->offset ? -1 : 1;
1293 if (f1->size == f2->size)
1295 if (f1->type == f2->type)
1296 return 0;
1297 /* Put any non-aggregate type before any aggregate type. */
1298 else if (!is_gimple_reg_type (f1->type)
1299 && is_gimple_reg_type (f2->type))
1300 return 1;
1301 else if (is_gimple_reg_type (f1->type)
1302 && !is_gimple_reg_type (f2->type))
1303 return -1;
1304 /* Put any complex or vector type before any other scalar type. */
1305 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1306 && TREE_CODE (f1->type) != VECTOR_TYPE
1307 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1308 || TREE_CODE (f2->type) == VECTOR_TYPE))
1309 return 1;
1310 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1311 || TREE_CODE (f1->type) == VECTOR_TYPE)
1312 && TREE_CODE (f2->type) != COMPLEX_TYPE
1313 && TREE_CODE (f2->type) != VECTOR_TYPE)
1314 return -1;
1315 /* Put the integral type with the bigger precision first. */
1316 else if (INTEGRAL_TYPE_P (f1->type)
1317 && INTEGRAL_TYPE_P (f2->type))
1318 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1319 /* Put any integral type with non-full precision last. */
1320 else if (INTEGRAL_TYPE_P (f1->type)
1321 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1322 != TYPE_PRECISION (f1->type)))
1323 return 1;
1324 else if (INTEGRAL_TYPE_P (f2->type)
1325 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1326 != TYPE_PRECISION (f2->type)))
1327 return -1;
1328 /* Stabilize the sort. */
1329 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1332 /* We want the bigger accesses first, thus the opposite operator in the next
1333 line: */
1334 return f1->size > f2->size ? -1 : 1;
1338 /* Append a name of the declaration to the name obstack. A helper function for
1339 make_fancy_name. */
1341 static void
1342 make_fancy_decl_name (tree decl)
1344 char buffer[32];
1346 tree name = DECL_NAME (decl);
1347 if (name)
1348 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1349 IDENTIFIER_LENGTH (name));
1350 else
1352 sprintf (buffer, "D%u", DECL_UID (decl));
1353 obstack_grow (&name_obstack, buffer, strlen (buffer));
1357 /* Helper for make_fancy_name. */
1359 static void
1360 make_fancy_name_1 (tree expr)
1362 char buffer[32];
1363 tree index;
1365 if (DECL_P (expr))
1367 make_fancy_decl_name (expr);
1368 return;
1371 switch (TREE_CODE (expr))
1373 case COMPONENT_REF:
1374 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1375 obstack_1grow (&name_obstack, '$');
1376 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1377 break;
1379 case ARRAY_REF:
1380 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1381 obstack_1grow (&name_obstack, '$');
1382 /* Arrays with only one element may not have a constant as their
1383 index. */
1384 index = TREE_OPERAND (expr, 1);
1385 if (TREE_CODE (index) != INTEGER_CST)
1386 break;
1387 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1388 obstack_grow (&name_obstack, buffer, strlen (buffer));
1389 break;
1391 case ADDR_EXPR:
1392 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1393 break;
1395 case MEM_REF:
1396 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1397 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1399 obstack_1grow (&name_obstack, '$');
1400 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1401 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1402 obstack_grow (&name_obstack, buffer, strlen (buffer));
1404 break;
1406 case BIT_FIELD_REF:
1407 case REALPART_EXPR:
1408 case IMAGPART_EXPR:
1409 gcc_unreachable (); /* we treat these as scalars. */
1410 break;
1411 default:
1412 break;
1416 /* Create a human readable name for replacement variable of ACCESS. */
1418 static char *
1419 make_fancy_name (tree expr)
1421 make_fancy_name_1 (expr);
1422 obstack_1grow (&name_obstack, '\0');
1423 return XOBFINISH (&name_obstack, char *);
1426 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1427 EXP_TYPE at the given OFFSET. If BASE is something for which
1428 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1429 to insert new statements either before or below the current one as specified
1430 by INSERT_AFTER. This function is not capable of handling bitfields. */
1432 tree
1433 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1434 tree exp_type, gimple_stmt_iterator *gsi,
1435 bool insert_after)
1437 tree prev_base = base;
1438 tree off;
1439 HOST_WIDE_INT base_offset;
1441 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1443 base = get_addr_base_and_unit_offset (base, &base_offset);
1445 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1446 offset such as array[var_index]. */
1447 if (!base)
1449 gimple stmt;
1450 tree tmp, addr;
1452 gcc_checking_assert (gsi);
1453 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1454 add_referenced_var (tmp);
1455 tmp = make_ssa_name (tmp, NULL);
1456 addr = build_fold_addr_expr (unshare_expr (prev_base));
1457 STRIP_USELESS_TYPE_CONVERSION (addr);
1458 stmt = gimple_build_assign (tmp, addr);
1459 gimple_set_location (stmt, loc);
1460 SSA_NAME_DEF_STMT (tmp) = stmt;
1461 if (insert_after)
1462 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1463 else
1464 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1465 update_stmt (stmt);
1467 off = build_int_cst (reference_alias_ptr_type (prev_base),
1468 offset / BITS_PER_UNIT);
1469 base = tmp;
1471 else if (TREE_CODE (base) == MEM_REF)
1473 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1474 base_offset + offset / BITS_PER_UNIT);
1475 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1476 base = unshare_expr (TREE_OPERAND (base, 0));
1478 else
1480 off = build_int_cst (reference_alias_ptr_type (base),
1481 base_offset + offset / BITS_PER_UNIT);
1482 base = build_fold_addr_expr (unshare_expr (base));
1485 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1488 /* Construct a memory reference to a part of an aggregate BASE at the given
1489 OFFSET and of the same type as MODEL. In case this is a reference to a
1490 component, the function will replicate the last COMPONENT_REF of model's
1491 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1492 build_ref_for_offset. */
1494 static tree
1495 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1496 struct access *model, gimple_stmt_iterator *gsi,
1497 bool insert_after)
1499 if (TREE_CODE (model->expr) == COMPONENT_REF)
1501 tree t, exp_type, fld = TREE_OPERAND (model->expr, 1);
1502 tree cr_offset = component_ref_field_offset (model->expr);
1504 gcc_assert (cr_offset && host_integerp (cr_offset, 1));
1505 offset -= TREE_INT_CST_LOW (cr_offset) * BITS_PER_UNIT;
1506 offset -= TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (fld));
1507 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1508 t = build_ref_for_offset (loc, base, offset, exp_type, gsi, insert_after);
1509 return fold_build3_loc (loc, COMPONENT_REF, TREE_TYPE (fld), t, fld,
1510 TREE_OPERAND (model->expr, 2));
1512 else
1513 return build_ref_for_offset (loc, base, offset, model->type,
1514 gsi, insert_after);
1517 /* Construct a memory reference consisting of component_refs and array_refs to
1518 a part of an aggregate *RES (which is of type TYPE). The requested part
1519 should have type EXP_TYPE at be the given OFFSET. This function might not
1520 succeed, it returns true when it does and only then *RES points to something
1521 meaningful. This function should be used only to build expressions that we
1522 might need to present to user (e.g. in warnings). In all other situations,
1523 build_ref_for_model or build_ref_for_offset should be used instead. */
1525 static bool
1526 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1527 tree exp_type)
1529 while (1)
1531 tree fld;
1532 tree tr_size, index, minidx;
1533 HOST_WIDE_INT el_size;
1535 if (offset == 0 && exp_type
1536 && types_compatible_p (exp_type, type))
1537 return true;
1539 switch (TREE_CODE (type))
1541 case UNION_TYPE:
1542 case QUAL_UNION_TYPE:
1543 case RECORD_TYPE:
1544 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1546 HOST_WIDE_INT pos, size;
1547 tree expr, *expr_ptr;
1549 if (TREE_CODE (fld) != FIELD_DECL)
1550 continue;
1552 pos = int_bit_position (fld);
1553 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1554 tr_size = DECL_SIZE (fld);
1555 if (!tr_size || !host_integerp (tr_size, 1))
1556 continue;
1557 size = tree_low_cst (tr_size, 1);
1558 if (size == 0)
1560 if (pos != offset)
1561 continue;
1563 else if (pos > offset || (pos + size) <= offset)
1564 continue;
1566 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1567 NULL_TREE);
1568 expr_ptr = &expr;
1569 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1570 offset - pos, exp_type))
1572 *res = expr;
1573 return true;
1576 return false;
1578 case ARRAY_TYPE:
1579 tr_size = TYPE_SIZE (TREE_TYPE (type));
1580 if (!tr_size || !host_integerp (tr_size, 1))
1581 return false;
1582 el_size = tree_low_cst (tr_size, 1);
1584 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1585 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1586 return false;
1587 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1588 if (!integer_zerop (minidx))
1589 index = int_const_binop (PLUS_EXPR, index, minidx);
1590 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1591 NULL_TREE, NULL_TREE);
1592 offset = offset % el_size;
1593 type = TREE_TYPE (type);
1594 break;
1596 default:
1597 if (offset != 0)
1598 return false;
1600 if (exp_type)
1601 return false;
1602 else
1603 return true;
1608 /* Return true iff TYPE is stdarg va_list type. */
1610 static inline bool
1611 is_va_list_type (tree type)
1613 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1616 /* Print message to dump file why a variable was rejected. */
1618 static void
1619 reject (tree var, const char *msg)
1621 if (dump_file && (dump_flags & TDF_DETAILS))
1623 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1624 print_generic_expr (dump_file, var, 0);
1625 fprintf (dump_file, "\n");
1629 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1630 those with type which is suitable for scalarization. */
1632 static bool
1633 find_var_candidates (void)
1635 tree var, type;
1636 referenced_var_iterator rvi;
1637 bool ret = false;
1638 const char *msg;
1640 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
1642 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1643 continue;
1644 type = TREE_TYPE (var);
1646 if (!AGGREGATE_TYPE_P (type))
1648 reject (var, "not aggregate");
1649 continue;
1651 if (needs_to_live_in_memory (var))
1653 reject (var, "needs to live in memory");
1654 continue;
1656 if (TREE_THIS_VOLATILE (var))
1658 reject (var, "is volatile");
1659 continue;
1661 if (!COMPLETE_TYPE_P (type))
1663 reject (var, "has incomplete type");
1664 continue;
1666 if (!host_integerp (TYPE_SIZE (type), 1))
1668 reject (var, "type size not fixed");
1669 continue;
1671 if (tree_low_cst (TYPE_SIZE (type), 1) == 0)
1673 reject (var, "type size is zero");
1674 continue;
1676 if (type_internals_preclude_sra_p (type, &msg))
1678 reject (var, msg);
1679 continue;
1681 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1682 we also want to schedule it rather late. Thus we ignore it in
1683 the early pass. */
1684 (sra_mode == SRA_MODE_EARLY_INTRA
1685 && is_va_list_type (type)))
1687 reject (var, "is va_list");
1688 continue;
1691 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1693 if (dump_file && (dump_flags & TDF_DETAILS))
1695 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1696 print_generic_expr (dump_file, var, 0);
1697 fprintf (dump_file, "\n");
1699 ret = true;
1702 return ret;
1705 /* Sort all accesses for the given variable, check for partial overlaps and
1706 return NULL if there are any. If there are none, pick a representative for
1707 each combination of offset and size and create a linked list out of them.
1708 Return the pointer to the first representative and make sure it is the first
1709 one in the vector of accesses. */
1711 static struct access *
1712 sort_and_splice_var_accesses (tree var)
1714 int i, j, access_count;
1715 struct access *res, **prev_acc_ptr = &res;
1716 VEC (access_p, heap) *access_vec;
1717 bool first = true;
1718 HOST_WIDE_INT low = -1, high = 0;
1720 access_vec = get_base_access_vector (var);
1721 if (!access_vec)
1722 return NULL;
1723 access_count = VEC_length (access_p, access_vec);
1725 /* Sort by <OFFSET, SIZE>. */
1726 VEC_qsort (access_p, access_vec, compare_access_positions);
1728 i = 0;
1729 while (i < access_count)
1731 struct access *access = VEC_index (access_p, access_vec, i);
1732 bool grp_write = access->write;
1733 bool grp_read = !access->write;
1734 bool grp_scalar_write = access->write
1735 && is_gimple_reg_type (access->type);
1736 bool grp_scalar_read = !access->write
1737 && is_gimple_reg_type (access->type);
1738 bool grp_assignment_read = access->grp_assignment_read;
1739 bool grp_assignment_write = access->grp_assignment_write;
1740 bool multiple_scalar_reads = false;
1741 bool total_scalarization = access->grp_total_scalarization;
1742 bool grp_partial_lhs = access->grp_partial_lhs;
1743 bool first_scalar = is_gimple_reg_type (access->type);
1744 bool unscalarizable_region = access->grp_unscalarizable_region;
1746 if (first || access->offset >= high)
1748 first = false;
1749 low = access->offset;
1750 high = access->offset + access->size;
1752 else if (access->offset > low && access->offset + access->size > high)
1753 return NULL;
1754 else
1755 gcc_assert (access->offset >= low
1756 && access->offset + access->size <= high);
1758 j = i + 1;
1759 while (j < access_count)
1761 struct access *ac2 = VEC_index (access_p, access_vec, j);
1762 if (ac2->offset != access->offset || ac2->size != access->size)
1763 break;
1764 if (ac2->write)
1766 grp_write = true;
1767 grp_scalar_write = (grp_scalar_write
1768 || is_gimple_reg_type (ac2->type));
1770 else
1772 grp_read = true;
1773 if (is_gimple_reg_type (ac2->type))
1775 if (grp_scalar_read)
1776 multiple_scalar_reads = true;
1777 else
1778 grp_scalar_read = true;
1781 grp_assignment_read |= ac2->grp_assignment_read;
1782 grp_assignment_write |= ac2->grp_assignment_write;
1783 grp_partial_lhs |= ac2->grp_partial_lhs;
1784 unscalarizable_region |= ac2->grp_unscalarizable_region;
1785 total_scalarization |= ac2->grp_total_scalarization;
1786 relink_to_new_repr (access, ac2);
1788 /* If there are both aggregate-type and scalar-type accesses with
1789 this combination of size and offset, the comparison function
1790 should have put the scalars first. */
1791 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1792 ac2->group_representative = access;
1793 j++;
1796 i = j;
1798 access->group_representative = access;
1799 access->grp_write = grp_write;
1800 access->grp_read = grp_read;
1801 access->grp_scalar_read = grp_scalar_read;
1802 access->grp_scalar_write = grp_scalar_write;
1803 access->grp_assignment_read = grp_assignment_read;
1804 access->grp_assignment_write = grp_assignment_write;
1805 access->grp_hint = multiple_scalar_reads || total_scalarization;
1806 access->grp_total_scalarization = total_scalarization;
1807 access->grp_partial_lhs = grp_partial_lhs;
1808 access->grp_unscalarizable_region = unscalarizable_region;
1809 if (access->first_link)
1810 add_access_to_work_queue (access);
1812 *prev_acc_ptr = access;
1813 prev_acc_ptr = &access->next_grp;
1816 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1817 return res;
1820 /* Create a variable for the given ACCESS which determines the type, name and a
1821 few other properties. Return the variable declaration and store it also to
1822 ACCESS->replacement. */
1824 static tree
1825 create_access_replacement (struct access *access, bool rename)
1827 tree repl;
1829 repl = create_tmp_var (access->type, "SR");
1830 add_referenced_var (repl);
1831 if (rename)
1832 mark_sym_for_renaming (repl);
1834 if (!access->grp_partial_lhs
1835 && (TREE_CODE (access->type) == COMPLEX_TYPE
1836 || TREE_CODE (access->type) == VECTOR_TYPE))
1837 DECL_GIMPLE_REG_P (repl) = 1;
1839 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1840 DECL_ARTIFICIAL (repl) = 1;
1841 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1843 if (DECL_NAME (access->base)
1844 && !DECL_IGNORED_P (access->base)
1845 && !DECL_ARTIFICIAL (access->base))
1847 char *pretty_name = make_fancy_name (access->expr);
1848 tree debug_expr = unshare_expr (access->expr), d;
1850 DECL_NAME (repl) = get_identifier (pretty_name);
1851 obstack_free (&name_obstack, pretty_name);
1853 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1854 as DECL_DEBUG_EXPR isn't considered when looking for still
1855 used SSA_NAMEs and thus they could be freed. All debug info
1856 generation cares is whether something is constant or variable
1857 and that get_ref_base_and_extent works properly on the
1858 expression. */
1859 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1860 switch (TREE_CODE (d))
1862 case ARRAY_REF:
1863 case ARRAY_RANGE_REF:
1864 if (TREE_OPERAND (d, 1)
1865 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1866 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1867 if (TREE_OPERAND (d, 3)
1868 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1869 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1870 /* FALLTHRU */
1871 case COMPONENT_REF:
1872 if (TREE_OPERAND (d, 2)
1873 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1874 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1875 break;
1876 default:
1877 break;
1879 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1880 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1881 if (access->grp_no_warning)
1882 TREE_NO_WARNING (repl) = 1;
1883 else
1884 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1886 else
1887 TREE_NO_WARNING (repl) = 1;
1889 if (dump_file)
1891 fprintf (dump_file, "Created a replacement for ");
1892 print_generic_expr (dump_file, access->base, 0);
1893 fprintf (dump_file, " offset: %u, size: %u: ",
1894 (unsigned) access->offset, (unsigned) access->size);
1895 print_generic_expr (dump_file, repl, 0);
1896 fprintf (dump_file, "\n");
1898 sra_stats.replacements++;
1900 return repl;
1903 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1905 static inline tree
1906 get_access_replacement (struct access *access)
1908 gcc_assert (access->grp_to_be_replaced);
1910 if (!access->replacement_decl)
1911 access->replacement_decl = create_access_replacement (access, true);
1912 return access->replacement_decl;
1915 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1916 not mark it for renaming. */
1918 static inline tree
1919 get_unrenamed_access_replacement (struct access *access)
1921 gcc_assert (!access->grp_to_be_replaced);
1923 if (!access->replacement_decl)
1924 access->replacement_decl = create_access_replacement (access, false);
1925 return access->replacement_decl;
1929 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1930 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1931 to it is not "within" the root. Return false iff some accesses partially
1932 overlap. */
1934 static bool
1935 build_access_subtree (struct access **access)
1937 struct access *root = *access, *last_child = NULL;
1938 HOST_WIDE_INT limit = root->offset + root->size;
1940 *access = (*access)->next_grp;
1941 while (*access && (*access)->offset + (*access)->size <= limit)
1943 if (!last_child)
1944 root->first_child = *access;
1945 else
1946 last_child->next_sibling = *access;
1947 last_child = *access;
1949 if (!build_access_subtree (access))
1950 return false;
1953 if (*access && (*access)->offset < limit)
1954 return false;
1956 return true;
1959 /* Build a tree of access representatives, ACCESS is the pointer to the first
1960 one, others are linked in a list by the next_grp field. Return false iff
1961 some accesses partially overlap. */
1963 static bool
1964 build_access_trees (struct access *access)
1966 while (access)
1968 struct access *root = access;
1970 if (!build_access_subtree (&access))
1971 return false;
1972 root->next_grp = access;
1974 return true;
1977 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1978 array. */
1980 static bool
1981 expr_with_var_bounded_array_refs_p (tree expr)
1983 while (handled_component_p (expr))
1985 if (TREE_CODE (expr) == ARRAY_REF
1986 && !host_integerp (array_ref_low_bound (expr), 0))
1987 return true;
1988 expr = TREE_OPERAND (expr, 0);
1990 return false;
1993 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1994 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1995 sorts of access flags appropriately along the way, notably always set
1996 grp_read and grp_assign_read according to MARK_READ and grp_write when
1997 MARK_WRITE is true.
1999 Creating a replacement for a scalar access is considered beneficial if its
2000 grp_hint is set (this means we are either attempting total scalarization or
2001 there is more than one direct read access) or according to the following
2002 table:
2004 Access written to through a scalar type (once or more times)
2006 | Written to in an assignment statement
2008 | | Access read as scalar _once_
2009 | | |
2010 | | | Read in an assignment statement
2011 | | | |
2012 | | | | Scalarize Comment
2013 -----------------------------------------------------------------------------
2014 0 0 0 0 No access for the scalar
2015 0 0 0 1 No access for the scalar
2016 0 0 1 0 No Single read - won't help
2017 0 0 1 1 No The same case
2018 0 1 0 0 No access for the scalar
2019 0 1 0 1 No access for the scalar
2020 0 1 1 0 Yes s = *g; return s.i;
2021 0 1 1 1 Yes The same case as above
2022 1 0 0 0 No Won't help
2023 1 0 0 1 Yes s.i = 1; *g = s;
2024 1 0 1 0 Yes s.i = 5; g = s.i;
2025 1 0 1 1 Yes The same case as above
2026 1 1 0 0 No Won't help.
2027 1 1 0 1 Yes s.i = 1; *g = s;
2028 1 1 1 0 Yes s = *g; return s.i;
2029 1 1 1 1 Yes Any of the above yeses */
2031 static bool
2032 analyze_access_subtree (struct access *root, struct access *parent,
2033 bool allow_replacements)
2035 struct access *child;
2036 HOST_WIDE_INT limit = root->offset + root->size;
2037 HOST_WIDE_INT covered_to = root->offset;
2038 bool scalar = is_gimple_reg_type (root->type);
2039 bool hole = false, sth_created = false;
2041 if (parent)
2043 if (parent->grp_read)
2044 root->grp_read = 1;
2045 if (parent->grp_assignment_read)
2046 root->grp_assignment_read = 1;
2047 if (parent->grp_write)
2048 root->grp_write = 1;
2049 if (parent->grp_assignment_write)
2050 root->grp_assignment_write = 1;
2051 if (parent->grp_total_scalarization)
2052 root->grp_total_scalarization = 1;
2055 if (root->grp_unscalarizable_region)
2056 allow_replacements = false;
2058 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2059 allow_replacements = false;
2061 for (child = root->first_child; child; child = child->next_sibling)
2063 hole |= covered_to < child->offset;
2064 sth_created |= analyze_access_subtree (child, root,
2065 allow_replacements && !scalar);
2067 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2068 root->grp_total_scalarization &= child->grp_total_scalarization;
2069 if (child->grp_covered)
2070 covered_to += child->size;
2071 else
2072 hole = true;
2075 if (allow_replacements && scalar && !root->first_child
2076 && (root->grp_hint
2077 || ((root->grp_scalar_read || root->grp_assignment_read)
2078 && (root->grp_scalar_write || root->grp_assignment_write))))
2080 bool new_integer_type;
2081 if (TREE_CODE (root->type) == ENUMERAL_TYPE)
2083 tree rt = root->type;
2084 root->type = build_nonstandard_integer_type (TYPE_PRECISION (rt),
2085 TYPE_UNSIGNED (rt));
2086 new_integer_type = true;
2088 else
2089 new_integer_type = false;
2091 if (dump_file && (dump_flags & TDF_DETAILS))
2093 fprintf (dump_file, "Marking ");
2094 print_generic_expr (dump_file, root->base, 0);
2095 fprintf (dump_file, " offset: %u, size: %u ",
2096 (unsigned) root->offset, (unsigned) root->size);
2097 fprintf (dump_file, " to be replaced%s.\n",
2098 new_integer_type ? " with an integer": "");
2101 root->grp_to_be_replaced = 1;
2102 sth_created = true;
2103 hole = false;
2105 else
2107 if (covered_to < limit)
2108 hole = true;
2109 if (scalar)
2110 root->grp_total_scalarization = 0;
2113 if (sth_created
2114 && (!hole || root->grp_total_scalarization))
2116 root->grp_covered = 1;
2117 return true;
2119 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2120 root->grp_unscalarized_data = 1; /* not covered and written to */
2121 if (sth_created)
2122 return true;
2123 return false;
2126 /* Analyze all access trees linked by next_grp by the means of
2127 analyze_access_subtree. */
2128 static bool
2129 analyze_access_trees (struct access *access)
2131 bool ret = false;
2133 while (access)
2135 if (analyze_access_subtree (access, NULL, true))
2136 ret = true;
2137 access = access->next_grp;
2140 return ret;
2143 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2144 SIZE would conflict with an already existing one. If exactly such a child
2145 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2147 static bool
2148 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2149 HOST_WIDE_INT size, struct access **exact_match)
2151 struct access *child;
2153 for (child = lacc->first_child; child; child = child->next_sibling)
2155 if (child->offset == norm_offset && child->size == size)
2157 *exact_match = child;
2158 return true;
2161 if (child->offset < norm_offset + size
2162 && child->offset + child->size > norm_offset)
2163 return true;
2166 return false;
2169 /* Create a new child access of PARENT, with all properties just like MODEL
2170 except for its offset and with its grp_write false and grp_read true.
2171 Return the new access or NULL if it cannot be created. Note that this access
2172 is created long after all splicing and sorting, it's not located in any
2173 access vector and is automatically a representative of its group. */
2175 static struct access *
2176 create_artificial_child_access (struct access *parent, struct access *model,
2177 HOST_WIDE_INT new_offset)
2179 struct access *access;
2180 struct access **child;
2181 tree expr = parent->base;
2183 gcc_assert (!model->grp_unscalarizable_region);
2185 access = (struct access *) pool_alloc (access_pool);
2186 memset (access, 0, sizeof (struct access));
2187 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2188 model->type))
2190 access->grp_no_warning = true;
2191 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2192 new_offset, model, NULL, false);
2195 access->base = parent->base;
2196 access->expr = expr;
2197 access->offset = new_offset;
2198 access->size = model->size;
2199 access->type = model->type;
2200 access->grp_write = true;
2201 access->grp_read = false;
2203 child = &parent->first_child;
2204 while (*child && (*child)->offset < new_offset)
2205 child = &(*child)->next_sibling;
2207 access->next_sibling = *child;
2208 *child = access;
2210 return access;
2214 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2215 true if any new subaccess was created. Additionally, if RACC is a scalar
2216 access but LACC is not, change the type of the latter, if possible. */
2218 static bool
2219 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2221 struct access *rchild;
2222 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2223 bool ret = false;
2225 if (is_gimple_reg_type (lacc->type)
2226 || lacc->grp_unscalarizable_region
2227 || racc->grp_unscalarizable_region)
2228 return false;
2230 if (!lacc->first_child && !racc->first_child
2231 && is_gimple_reg_type (racc->type))
2233 tree t = lacc->base;
2235 lacc->type = racc->type;
2236 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t), lacc->offset,
2237 racc->type))
2238 lacc->expr = t;
2239 else
2241 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2242 lacc->base, lacc->offset,
2243 racc, NULL, false);
2244 lacc->grp_no_warning = true;
2246 return false;
2249 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2251 struct access *new_acc = NULL;
2252 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2254 if (rchild->grp_unscalarizable_region)
2255 continue;
2257 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2258 &new_acc))
2260 if (new_acc)
2262 rchild->grp_hint = 1;
2263 new_acc->grp_hint |= new_acc->grp_read;
2264 if (rchild->first_child)
2265 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2267 continue;
2270 rchild->grp_hint = 1;
2271 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2272 if (new_acc)
2274 ret = true;
2275 if (racc->first_child)
2276 propagate_subaccesses_across_link (new_acc, rchild);
2280 return ret;
2283 /* Propagate all subaccesses across assignment links. */
2285 static void
2286 propagate_all_subaccesses (void)
2288 while (work_queue_head)
2290 struct access *racc = pop_access_from_work_queue ();
2291 struct assign_link *link;
2293 gcc_assert (racc->first_link);
2295 for (link = racc->first_link; link; link = link->next)
2297 struct access *lacc = link->lacc;
2299 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2300 continue;
2301 lacc = lacc->group_representative;
2302 if (propagate_subaccesses_across_link (lacc, racc)
2303 && lacc->first_link)
2304 add_access_to_work_queue (lacc);
2309 /* Go through all accesses collected throughout the (intraprocedural) analysis
2310 stage, exclude overlapping ones, identify representatives and build trees
2311 out of them, making decisions about scalarization on the way. Return true
2312 iff there are any to-be-scalarized variables after this stage. */
2314 static bool
2315 analyze_all_variable_accesses (void)
2317 int res = 0;
2318 bitmap tmp = BITMAP_ALLOC (NULL);
2319 bitmap_iterator bi;
2320 unsigned i, max_total_scalarization_size;
2322 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2323 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2325 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2326 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2327 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2329 tree var = referenced_var (i);
2331 if (TREE_CODE (var) == VAR_DECL
2332 && type_consists_of_records_p (TREE_TYPE (var)))
2334 if ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2335 <= max_total_scalarization_size)
2337 completely_scalarize_var (var);
2338 if (dump_file && (dump_flags & TDF_DETAILS))
2340 fprintf (dump_file, "Will attempt to totally scalarize ");
2341 print_generic_expr (dump_file, var, 0);
2342 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2345 else if (dump_file && (dump_flags & TDF_DETAILS))
2347 fprintf (dump_file, "Too big to totally scalarize: ");
2348 print_generic_expr (dump_file, var, 0);
2349 fprintf (dump_file, " (UID: %u)\n", DECL_UID (var));
2354 bitmap_copy (tmp, candidate_bitmap);
2355 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2357 tree var = referenced_var (i);
2358 struct access *access;
2360 access = sort_and_splice_var_accesses (var);
2361 if (!access || !build_access_trees (access))
2362 disqualify_candidate (var,
2363 "No or inhibitingly overlapping accesses.");
2366 propagate_all_subaccesses ();
2368 bitmap_copy (tmp, candidate_bitmap);
2369 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2371 tree var = referenced_var (i);
2372 struct access *access = get_first_repr_for_decl (var);
2374 if (analyze_access_trees (access))
2376 res++;
2377 if (dump_file && (dump_flags & TDF_DETAILS))
2379 fprintf (dump_file, "\nAccess trees for ");
2380 print_generic_expr (dump_file, var, 0);
2381 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2382 dump_access_tree (dump_file, access);
2383 fprintf (dump_file, "\n");
2386 else
2387 disqualify_candidate (var, "No scalar replacements to be created.");
2390 BITMAP_FREE (tmp);
2392 if (res)
2394 statistics_counter_event (cfun, "Scalarized aggregates", res);
2395 return true;
2397 else
2398 return false;
2401 /* Generate statements copying scalar replacements of accesses within a subtree
2402 into or out of AGG. ACCESS, all its children, siblings and their children
2403 are to be processed. AGG is an aggregate type expression (can be a
2404 declaration but does not have to be, it can for example also be a mem_ref or
2405 a series of handled components). TOP_OFFSET is the offset of the processed
2406 subtree which has to be subtracted from offsets of individual accesses to
2407 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2408 replacements in the interval <start_offset, start_offset + chunk_size>,
2409 otherwise copy all. GSI is a statement iterator used to place the new
2410 statements. WRITE should be true when the statements should write from AGG
2411 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2412 statements will be added after the current statement in GSI, they will be
2413 added before the statement otherwise. */
2415 static void
2416 generate_subtree_copies (struct access *access, tree agg,
2417 HOST_WIDE_INT top_offset,
2418 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2419 gimple_stmt_iterator *gsi, bool write,
2420 bool insert_after, location_t loc)
2424 if (chunk_size && access->offset >= start_offset + chunk_size)
2425 return;
2427 if (access->grp_to_be_replaced
2428 && (chunk_size == 0
2429 || access->offset + access->size > start_offset))
2431 tree expr, repl = get_access_replacement (access);
2432 gimple stmt;
2434 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2435 access, gsi, insert_after);
2437 if (write)
2439 if (access->grp_partial_lhs)
2440 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2441 !insert_after,
2442 insert_after ? GSI_NEW_STMT
2443 : GSI_SAME_STMT);
2444 stmt = gimple_build_assign (repl, expr);
2446 else
2448 TREE_NO_WARNING (repl) = 1;
2449 if (access->grp_partial_lhs)
2450 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2451 !insert_after,
2452 insert_after ? GSI_NEW_STMT
2453 : GSI_SAME_STMT);
2454 stmt = gimple_build_assign (expr, repl);
2456 gimple_set_location (stmt, loc);
2458 if (insert_after)
2459 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2460 else
2461 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2462 update_stmt (stmt);
2463 sra_stats.subtree_copies++;
2466 if (access->first_child)
2467 generate_subtree_copies (access->first_child, agg, top_offset,
2468 start_offset, chunk_size, gsi,
2469 write, insert_after, loc);
2471 access = access->next_sibling;
2473 while (access);
2476 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2477 the root of the subtree to be processed. GSI is the statement iterator used
2478 for inserting statements which are added after the current statement if
2479 INSERT_AFTER is true or before it otherwise. */
2481 static void
2482 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2483 bool insert_after, location_t loc)
2486 struct access *child;
2488 if (access->grp_to_be_replaced)
2490 gimple stmt;
2492 stmt = gimple_build_assign (get_access_replacement (access),
2493 build_zero_cst (access->type));
2494 if (insert_after)
2495 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2496 else
2497 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2498 update_stmt (stmt);
2499 gimple_set_location (stmt, loc);
2502 for (child = access->first_child; child; child = child->next_sibling)
2503 init_subtree_with_zero (child, gsi, insert_after, loc);
2506 /* Search for an access representative for the given expression EXPR and
2507 return it or NULL if it cannot be found. */
2509 static struct access *
2510 get_access_for_expr (tree expr)
2512 HOST_WIDE_INT offset, size, max_size;
2513 tree base;
2515 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2516 a different size than the size of its argument and we need the latter
2517 one. */
2518 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2519 expr = TREE_OPERAND (expr, 0);
2521 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2522 if (max_size == -1 || !DECL_P (base))
2523 return NULL;
2525 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2526 return NULL;
2528 return get_var_base_offset_size_access (base, offset, max_size);
2531 /* Replace the expression EXPR with a scalar replacement if there is one and
2532 generate other statements to do type conversion or subtree copying if
2533 necessary. GSI is used to place newly created statements, WRITE is true if
2534 the expression is being written to (it is on a LHS of a statement or output
2535 in an assembly statement). */
2537 static bool
2538 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2540 location_t loc;
2541 struct access *access;
2542 tree type, bfr;
2544 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2546 bfr = *expr;
2547 expr = &TREE_OPERAND (*expr, 0);
2549 else
2550 bfr = NULL_TREE;
2552 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2553 expr = &TREE_OPERAND (*expr, 0);
2554 access = get_access_for_expr (*expr);
2555 if (!access)
2556 return false;
2557 type = TREE_TYPE (*expr);
2559 loc = gimple_location (gsi_stmt (*gsi));
2560 if (access->grp_to_be_replaced)
2562 tree repl = get_access_replacement (access);
2563 /* If we replace a non-register typed access simply use the original
2564 access expression to extract the scalar component afterwards.
2565 This happens if scalarizing a function return value or parameter
2566 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2567 gcc.c-torture/compile/20011217-1.c.
2569 We also want to use this when accessing a complex or vector which can
2570 be accessed as a different type too, potentially creating a need for
2571 type conversion (see PR42196) and when scalarized unions are involved
2572 in assembler statements (see PR42398). */
2573 if (!useless_type_conversion_p (type, access->type))
2575 tree ref;
2577 ref = build_ref_for_model (loc, access->base, access->offset, access,
2578 NULL, false);
2580 if (write)
2582 gimple stmt;
2584 if (access->grp_partial_lhs)
2585 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2586 false, GSI_NEW_STMT);
2587 stmt = gimple_build_assign (repl, ref);
2588 gimple_set_location (stmt, loc);
2589 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2591 else
2593 gimple stmt;
2595 if (access->grp_partial_lhs)
2596 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2597 true, GSI_SAME_STMT);
2598 stmt = gimple_build_assign (ref, repl);
2599 gimple_set_location (stmt, loc);
2600 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2603 else
2604 *expr = repl;
2605 sra_stats.exprs++;
2608 if (access->first_child)
2610 HOST_WIDE_INT start_offset, chunk_size;
2611 if (bfr
2612 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2613 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2615 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2616 start_offset = access->offset
2617 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2619 else
2620 start_offset = chunk_size = 0;
2622 generate_subtree_copies (access->first_child, access->base, 0,
2623 start_offset, chunk_size, gsi, write, write,
2624 loc);
2626 return true;
2629 /* Where scalar replacements of the RHS have been written to when a replacement
2630 of a LHS of an assigments cannot be direclty loaded from a replacement of
2631 the RHS. */
2632 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2633 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2634 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2636 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2637 base aggregate if there are unscalarized data or directly to LHS of the
2638 statement that is pointed to by GSI otherwise. */
2640 static enum unscalarized_data_handling
2641 handle_unscalarized_data_in_subtree (struct access *top_racc,
2642 gimple_stmt_iterator *gsi)
2644 if (top_racc->grp_unscalarized_data)
2646 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2647 gsi, false, false,
2648 gimple_location (gsi_stmt (*gsi)));
2649 return SRA_UDH_RIGHT;
2651 else
2653 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2654 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2655 0, 0, gsi, false, false,
2656 gimple_location (gsi_stmt (*gsi)));
2657 return SRA_UDH_LEFT;
2662 /* Try to generate statements to load all sub-replacements in an access subtree
2663 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2664 If that is not possible, refresh the TOP_RACC base aggregate and load the
2665 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2666 copied. NEW_GSI is stmt iterator used for statement insertions after the
2667 original assignment, OLD_GSI is used to insert statements before the
2668 assignment. *REFRESHED keeps the information whether we have needed to
2669 refresh replacements of the LHS and from which side of the assignments this
2670 takes place. */
2672 static void
2673 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2674 HOST_WIDE_INT left_offset,
2675 gimple_stmt_iterator *old_gsi,
2676 gimple_stmt_iterator *new_gsi,
2677 enum unscalarized_data_handling *refreshed)
2679 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2680 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2682 if (lacc->grp_to_be_replaced)
2684 struct access *racc;
2685 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2686 gimple stmt;
2687 tree rhs;
2689 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2690 if (racc && racc->grp_to_be_replaced)
2692 rhs = get_access_replacement (racc);
2693 if (!useless_type_conversion_p (lacc->type, racc->type))
2694 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2696 else
2698 /* No suitable access on the right hand side, need to load from
2699 the aggregate. See if we have to update it first... */
2700 if (*refreshed == SRA_UDH_NONE)
2701 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2702 old_gsi);
2704 if (*refreshed == SRA_UDH_LEFT)
2705 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2706 new_gsi, true);
2707 else
2708 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2709 new_gsi, true);
2712 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2713 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2714 gimple_set_location (stmt, loc);
2715 update_stmt (stmt);
2716 sra_stats.subreplacements++;
2718 else if (*refreshed == SRA_UDH_NONE
2719 && lacc->grp_read && !lacc->grp_covered)
2720 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2721 old_gsi);
2723 if (lacc->first_child)
2724 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2725 old_gsi, new_gsi, refreshed);
2729 /* Result code for SRA assignment modification. */
2730 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2731 SRA_AM_MODIFIED, /* stmt changed but not
2732 removed */
2733 SRA_AM_REMOVED }; /* stmt eliminated */
2735 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2736 to the assignment and GSI is the statement iterator pointing at it. Returns
2737 the same values as sra_modify_assign. */
2739 static enum assignment_mod_result
2740 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2742 tree lhs = gimple_assign_lhs (*stmt);
2743 struct access *acc;
2744 location_t loc;
2746 acc = get_access_for_expr (lhs);
2747 if (!acc)
2748 return SRA_AM_NONE;
2750 loc = gimple_location (*stmt);
2751 if (VEC_length (constructor_elt,
2752 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2754 /* I have never seen this code path trigger but if it can happen the
2755 following should handle it gracefully. */
2756 if (access_has_children_p (acc))
2757 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2758 true, true, loc);
2759 return SRA_AM_MODIFIED;
2762 if (acc->grp_covered)
2764 init_subtree_with_zero (acc, gsi, false, loc);
2765 unlink_stmt_vdef (*stmt);
2766 gsi_remove (gsi, true);
2767 return SRA_AM_REMOVED;
2769 else
2771 init_subtree_with_zero (acc, gsi, true, loc);
2772 return SRA_AM_MODIFIED;
2776 /* Create and return a new suitable default definition SSA_NAME for RACC which
2777 is an access describing an uninitialized part of an aggregate that is being
2778 loaded. */
2780 static tree
2781 get_repl_default_def_ssa_name (struct access *racc)
2783 tree repl, decl;
2785 decl = get_unrenamed_access_replacement (racc);
2787 repl = gimple_default_def (cfun, decl);
2788 if (!repl)
2790 repl = make_ssa_name (decl, gimple_build_nop ());
2791 set_default_def (decl, repl);
2794 return repl;
2797 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2798 somewhere in it. */
2800 static inline bool
2801 contains_bitfld_comp_ref_p (const_tree ref)
2803 while (handled_component_p (ref))
2805 if (TREE_CODE (ref) == COMPONENT_REF
2806 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2807 return true;
2808 ref = TREE_OPERAND (ref, 0);
2811 return false;
2814 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2815 bit-field field declaration somewhere in it. */
2817 static inline bool
2818 contains_vce_or_bfcref_p (const_tree ref)
2820 while (handled_component_p (ref))
2822 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2823 || (TREE_CODE (ref) == COMPONENT_REF
2824 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2825 return true;
2826 ref = TREE_OPERAND (ref, 0);
2829 return false;
2832 /* Examine both sides of the assignment statement pointed to by STMT, replace
2833 them with a scalare replacement if there is one and generate copying of
2834 replacements if scalarized aggregates have been used in the assignment. GSI
2835 is used to hold generated statements for type conversions and subtree
2836 copying. */
2838 static enum assignment_mod_result
2839 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2841 struct access *lacc, *racc;
2842 tree lhs, rhs;
2843 bool modify_this_stmt = false;
2844 bool force_gimple_rhs = false;
2845 location_t loc;
2846 gimple_stmt_iterator orig_gsi = *gsi;
2848 if (!gimple_assign_single_p (*stmt))
2849 return SRA_AM_NONE;
2850 lhs = gimple_assign_lhs (*stmt);
2851 rhs = gimple_assign_rhs1 (*stmt);
2853 if (TREE_CODE (rhs) == CONSTRUCTOR)
2854 return sra_modify_constructor_assign (stmt, gsi);
2856 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2857 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2858 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2860 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2861 gsi, false);
2862 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2863 gsi, true);
2864 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2867 lacc = get_access_for_expr (lhs);
2868 racc = get_access_for_expr (rhs);
2869 if (!lacc && !racc)
2870 return SRA_AM_NONE;
2872 loc = gimple_location (*stmt);
2873 if (lacc && lacc->grp_to_be_replaced)
2875 lhs = get_access_replacement (lacc);
2876 gimple_assign_set_lhs (*stmt, lhs);
2877 modify_this_stmt = true;
2878 if (lacc->grp_partial_lhs)
2879 force_gimple_rhs = true;
2880 sra_stats.exprs++;
2883 if (racc && racc->grp_to_be_replaced)
2885 rhs = get_access_replacement (racc);
2886 modify_this_stmt = true;
2887 if (racc->grp_partial_lhs)
2888 force_gimple_rhs = true;
2889 sra_stats.exprs++;
2892 if (modify_this_stmt)
2894 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2896 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2897 ??? This should move to fold_stmt which we simply should
2898 call after building a VIEW_CONVERT_EXPR here. */
2899 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2900 && !contains_bitfld_comp_ref_p (lhs)
2901 && !access_has_children_p (lacc))
2903 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
2904 gimple_assign_set_lhs (*stmt, lhs);
2906 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2907 && !contains_vce_or_bfcref_p (rhs)
2908 && !access_has_children_p (racc))
2909 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
2911 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2913 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
2914 rhs);
2915 if (is_gimple_reg_type (TREE_TYPE (lhs))
2916 && TREE_CODE (lhs) != SSA_NAME)
2917 force_gimple_rhs = true;
2922 /* From this point on, the function deals with assignments in between
2923 aggregates when at least one has scalar reductions of some of its
2924 components. There are three possible scenarios: Both the LHS and RHS have
2925 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2927 In the first case, we would like to load the LHS components from RHS
2928 components whenever possible. If that is not possible, we would like to
2929 read it directly from the RHS (after updating it by storing in it its own
2930 components). If there are some necessary unscalarized data in the LHS,
2931 those will be loaded by the original assignment too. If neither of these
2932 cases happen, the original statement can be removed. Most of this is done
2933 by load_assign_lhs_subreplacements.
2935 In the second case, we would like to store all RHS scalarized components
2936 directly into LHS and if they cover the aggregate completely, remove the
2937 statement too. In the third case, we want the LHS components to be loaded
2938 directly from the RHS (DSE will remove the original statement if it
2939 becomes redundant).
2941 This is a bit complex but manageable when types match and when unions do
2942 not cause confusion in a way that we cannot really load a component of LHS
2943 from the RHS or vice versa (the access representing this level can have
2944 subaccesses that are accessible only through a different union field at a
2945 higher level - different from the one used in the examined expression).
2946 Unions are fun.
2948 Therefore, I specially handle a fourth case, happening when there is a
2949 specific type cast or it is impossible to locate a scalarized subaccess on
2950 the other side of the expression. If that happens, I simply "refresh" the
2951 RHS by storing in it is scalarized components leave the original statement
2952 there to do the copying and then load the scalar replacements of the LHS.
2953 This is what the first branch does. */
2955 if (modify_this_stmt
2956 || gimple_has_volatile_ops (*stmt)
2957 || contains_vce_or_bfcref_p (rhs)
2958 || contains_vce_or_bfcref_p (lhs))
2960 if (access_has_children_p (racc))
2961 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2962 gsi, false, false, loc);
2963 if (access_has_children_p (lacc))
2964 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2965 gsi, true, true, loc);
2966 sra_stats.separate_lhs_rhs_handling++;
2968 else
2970 if (access_has_children_p (lacc) && access_has_children_p (racc))
2972 gimple_stmt_iterator orig_gsi = *gsi;
2973 enum unscalarized_data_handling refreshed;
2975 if (lacc->grp_read && !lacc->grp_covered)
2976 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
2977 else
2978 refreshed = SRA_UDH_NONE;
2980 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
2981 &orig_gsi, gsi, &refreshed);
2982 if (refreshed != SRA_UDH_RIGHT)
2984 gsi_next (gsi);
2985 unlink_stmt_vdef (*stmt);
2986 gsi_remove (&orig_gsi, true);
2987 sra_stats.deleted++;
2988 return SRA_AM_REMOVED;
2991 else
2993 if (racc)
2995 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2997 if (dump_file)
2999 fprintf (dump_file, "Removing load: ");
3000 print_gimple_stmt (dump_file, *stmt, 0, 0);
3003 if (TREE_CODE (lhs) == SSA_NAME)
3005 rhs = get_repl_default_def_ssa_name (racc);
3006 if (!useless_type_conversion_p (TREE_TYPE (lhs),
3007 TREE_TYPE (rhs)))
3008 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR,
3009 TREE_TYPE (lhs), rhs);
3011 else
3013 if (racc->first_child)
3014 generate_subtree_copies (racc->first_child, lhs,
3015 racc->offset, 0, 0, gsi,
3016 false, false, loc);
3018 gcc_assert (*stmt == gsi_stmt (*gsi));
3019 unlink_stmt_vdef (*stmt);
3020 gsi_remove (gsi, true);
3021 sra_stats.deleted++;
3022 return SRA_AM_REMOVED;
3025 else if (racc->first_child)
3026 generate_subtree_copies (racc->first_child, lhs, racc->offset,
3027 0, 0, gsi, false, true, loc);
3029 if (access_has_children_p (lacc))
3030 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
3031 0, 0, gsi, true, true, loc);
3035 /* This gimplification must be done after generate_subtree_copies, lest we
3036 insert the subtree copies in the middle of the gimplified sequence. */
3037 if (force_gimple_rhs)
3038 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
3039 true, GSI_SAME_STMT);
3040 if (gimple_assign_rhs1 (*stmt) != rhs)
3042 modify_this_stmt = true;
3043 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
3044 gcc_assert (*stmt == gsi_stmt (orig_gsi));
3047 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
3050 /* Traverse the function body and all modifications as decided in
3051 analyze_all_variable_accesses. Return true iff the CFG has been
3052 changed. */
3054 static bool
3055 sra_modify_function_body (void)
3057 bool cfg_changed = false;
3058 basic_block bb;
3060 FOR_EACH_BB (bb)
3062 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3063 while (!gsi_end_p (gsi))
3065 gimple stmt = gsi_stmt (gsi);
3066 enum assignment_mod_result assign_result;
3067 bool modified = false, deleted = false;
3068 tree *t;
3069 unsigned i;
3071 switch (gimple_code (stmt))
3073 case GIMPLE_RETURN:
3074 t = gimple_return_retval_ptr (stmt);
3075 if (*t != NULL_TREE)
3076 modified |= sra_modify_expr (t, &gsi, false);
3077 break;
3079 case GIMPLE_ASSIGN:
3080 assign_result = sra_modify_assign (&stmt, &gsi);
3081 modified |= assign_result == SRA_AM_MODIFIED;
3082 deleted = assign_result == SRA_AM_REMOVED;
3083 break;
3085 case GIMPLE_CALL:
3086 /* Operands must be processed before the lhs. */
3087 for (i = 0; i < gimple_call_num_args (stmt); i++)
3089 t = gimple_call_arg_ptr (stmt, i);
3090 modified |= sra_modify_expr (t, &gsi, false);
3093 if (gimple_call_lhs (stmt))
3095 t = gimple_call_lhs_ptr (stmt);
3096 modified |= sra_modify_expr (t, &gsi, true);
3098 break;
3100 case GIMPLE_ASM:
3101 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3103 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3104 modified |= sra_modify_expr (t, &gsi, false);
3106 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3108 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3109 modified |= sra_modify_expr (t, &gsi, true);
3111 break;
3113 default:
3114 break;
3117 if (modified)
3119 update_stmt (stmt);
3120 if (maybe_clean_eh_stmt (stmt)
3121 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3122 cfg_changed = true;
3124 if (!deleted)
3125 gsi_next (&gsi);
3129 return cfg_changed;
3132 /* Generate statements initializing scalar replacements of parts of function
3133 parameters. */
3135 static void
3136 initialize_parameter_reductions (void)
3138 gimple_stmt_iterator gsi;
3139 gimple_seq seq = NULL;
3140 tree parm;
3142 for (parm = DECL_ARGUMENTS (current_function_decl);
3143 parm;
3144 parm = DECL_CHAIN (parm))
3146 VEC (access_p, heap) *access_vec;
3147 struct access *access;
3149 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3150 continue;
3151 access_vec = get_base_access_vector (parm);
3152 if (!access_vec)
3153 continue;
3155 if (!seq)
3157 seq = gimple_seq_alloc ();
3158 gsi = gsi_start (seq);
3161 for (access = VEC_index (access_p, access_vec, 0);
3162 access;
3163 access = access->next_grp)
3164 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3165 EXPR_LOCATION (parm));
3168 if (seq)
3169 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3172 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3173 it reveals there are components of some aggregates to be scalarized, it runs
3174 the required transformations. */
3175 static unsigned int
3176 perform_intra_sra (void)
3178 int ret = 0;
3179 sra_initialize ();
3181 if (!find_var_candidates ())
3182 goto out;
3184 if (!scan_function ())
3185 goto out;
3187 if (!analyze_all_variable_accesses ())
3188 goto out;
3190 if (sra_modify_function_body ())
3191 ret = TODO_update_ssa | TODO_cleanup_cfg;
3192 else
3193 ret = TODO_update_ssa;
3194 initialize_parameter_reductions ();
3196 statistics_counter_event (cfun, "Scalar replacements created",
3197 sra_stats.replacements);
3198 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3199 statistics_counter_event (cfun, "Subtree copy stmts",
3200 sra_stats.subtree_copies);
3201 statistics_counter_event (cfun, "Subreplacement stmts",
3202 sra_stats.subreplacements);
3203 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3204 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3205 sra_stats.separate_lhs_rhs_handling);
3207 out:
3208 sra_deinitialize ();
3209 return ret;
3212 /* Perform early intraprocedural SRA. */
3213 static unsigned int
3214 early_intra_sra (void)
3216 sra_mode = SRA_MODE_EARLY_INTRA;
3217 return perform_intra_sra ();
3220 /* Perform "late" intraprocedural SRA. */
3221 static unsigned int
3222 late_intra_sra (void)
3224 sra_mode = SRA_MODE_INTRA;
3225 return perform_intra_sra ();
3229 static bool
3230 gate_intra_sra (void)
3232 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3236 struct gimple_opt_pass pass_sra_early =
3239 GIMPLE_PASS,
3240 "esra", /* name */
3241 gate_intra_sra, /* gate */
3242 early_intra_sra, /* execute */
3243 NULL, /* sub */
3244 NULL, /* next */
3245 0, /* static_pass_number */
3246 TV_TREE_SRA, /* tv_id */
3247 PROP_cfg | PROP_ssa, /* properties_required */
3248 0, /* properties_provided */
3249 0, /* properties_destroyed */
3250 0, /* todo_flags_start */
3251 TODO_update_ssa
3252 | TODO_ggc_collect
3253 | TODO_verify_ssa /* todo_flags_finish */
3257 struct gimple_opt_pass pass_sra =
3260 GIMPLE_PASS,
3261 "sra", /* name */
3262 gate_intra_sra, /* gate */
3263 late_intra_sra, /* execute */
3264 NULL, /* sub */
3265 NULL, /* next */
3266 0, /* static_pass_number */
3267 TV_TREE_SRA, /* tv_id */
3268 PROP_cfg | PROP_ssa, /* properties_required */
3269 0, /* properties_provided */
3270 0, /* properties_destroyed */
3271 TODO_update_address_taken, /* todo_flags_start */
3272 TODO_update_ssa
3273 | TODO_ggc_collect
3274 | TODO_verify_ssa /* todo_flags_finish */
3279 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3280 parameter. */
3282 static bool
3283 is_unused_scalar_param (tree parm)
3285 tree name;
3286 return (is_gimple_reg (parm)
3287 && (!(name = gimple_default_def (cfun, parm))
3288 || has_zero_uses (name)));
3291 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3292 examine whether there are any direct or otherwise infeasible ones. If so,
3293 return true, otherwise return false. PARM must be a gimple register with a
3294 non-NULL default definition. */
3296 static bool
3297 ptr_parm_has_direct_uses (tree parm)
3299 imm_use_iterator ui;
3300 gimple stmt;
3301 tree name = gimple_default_def (cfun, parm);
3302 bool ret = false;
3304 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3306 int uses_ok = 0;
3307 use_operand_p use_p;
3309 if (is_gimple_debug (stmt))
3310 continue;
3312 /* Valid uses include dereferences on the lhs and the rhs. */
3313 if (gimple_has_lhs (stmt))
3315 tree lhs = gimple_get_lhs (stmt);
3316 while (handled_component_p (lhs))
3317 lhs = TREE_OPERAND (lhs, 0);
3318 if (TREE_CODE (lhs) == MEM_REF
3319 && TREE_OPERAND (lhs, 0) == name
3320 && integer_zerop (TREE_OPERAND (lhs, 1))
3321 && types_compatible_p (TREE_TYPE (lhs),
3322 TREE_TYPE (TREE_TYPE (name)))
3323 && !TREE_THIS_VOLATILE (lhs))
3324 uses_ok++;
3326 if (gimple_assign_single_p (stmt))
3328 tree rhs = gimple_assign_rhs1 (stmt);
3329 while (handled_component_p (rhs))
3330 rhs = TREE_OPERAND (rhs, 0);
3331 if (TREE_CODE (rhs) == MEM_REF
3332 && TREE_OPERAND (rhs, 0) == name
3333 && integer_zerop (TREE_OPERAND (rhs, 1))
3334 && types_compatible_p (TREE_TYPE (rhs),
3335 TREE_TYPE (TREE_TYPE (name)))
3336 && !TREE_THIS_VOLATILE (rhs))
3337 uses_ok++;
3339 else if (is_gimple_call (stmt))
3341 unsigned i;
3342 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3344 tree arg = gimple_call_arg (stmt, i);
3345 while (handled_component_p (arg))
3346 arg = TREE_OPERAND (arg, 0);
3347 if (TREE_CODE (arg) == MEM_REF
3348 && TREE_OPERAND (arg, 0) == name
3349 && integer_zerop (TREE_OPERAND (arg, 1))
3350 && types_compatible_p (TREE_TYPE (arg),
3351 TREE_TYPE (TREE_TYPE (name)))
3352 && !TREE_THIS_VOLATILE (arg))
3353 uses_ok++;
3357 /* If the number of valid uses does not match the number of
3358 uses in this stmt there is an unhandled use. */
3359 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3360 --uses_ok;
3362 if (uses_ok != 0)
3363 ret = true;
3365 if (ret)
3366 BREAK_FROM_IMM_USE_STMT (ui);
3369 return ret;
3372 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3373 them in candidate_bitmap. Note that these do not necessarily include
3374 parameter which are unused and thus can be removed. Return true iff any
3375 such candidate has been found. */
3377 static bool
3378 find_param_candidates (void)
3380 tree parm;
3381 int count = 0;
3382 bool ret = false;
3383 const char *msg;
3385 for (parm = DECL_ARGUMENTS (current_function_decl);
3386 parm;
3387 parm = DECL_CHAIN (parm))
3389 tree type = TREE_TYPE (parm);
3391 count++;
3393 if (TREE_THIS_VOLATILE (parm)
3394 || TREE_ADDRESSABLE (parm)
3395 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3396 continue;
3398 if (is_unused_scalar_param (parm))
3400 ret = true;
3401 continue;
3404 if (POINTER_TYPE_P (type))
3406 type = TREE_TYPE (type);
3408 if (TREE_CODE (type) == FUNCTION_TYPE
3409 || TYPE_VOLATILE (type)
3410 || (TREE_CODE (type) == ARRAY_TYPE
3411 && TYPE_NONALIASED_COMPONENT (type))
3412 || !is_gimple_reg (parm)
3413 || is_va_list_type (type)
3414 || ptr_parm_has_direct_uses (parm))
3415 continue;
3417 else if (!AGGREGATE_TYPE_P (type))
3418 continue;
3420 if (!COMPLETE_TYPE_P (type)
3421 || !host_integerp (TYPE_SIZE (type), 1)
3422 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3423 || (AGGREGATE_TYPE_P (type)
3424 && type_internals_preclude_sra_p (type, &msg)))
3425 continue;
3427 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3428 ret = true;
3429 if (dump_file && (dump_flags & TDF_DETAILS))
3431 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3432 print_generic_expr (dump_file, parm, 0);
3433 fprintf (dump_file, "\n");
3437 func_param_count = count;
3438 return ret;
3441 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3442 maybe_modified. */
3444 static bool
3445 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3446 void *data)
3448 struct access *repr = (struct access *) data;
3450 repr->grp_maybe_modified = 1;
3451 return true;
3454 /* Analyze what representatives (in linked lists accessible from
3455 REPRESENTATIVES) can be modified by side effects of statements in the
3456 current function. */
3458 static void
3459 analyze_modified_params (VEC (access_p, heap) *representatives)
3461 int i;
3463 for (i = 0; i < func_param_count; i++)
3465 struct access *repr;
3467 for (repr = VEC_index (access_p, representatives, i);
3468 repr;
3469 repr = repr->next_grp)
3471 struct access *access;
3472 bitmap visited;
3473 ao_ref ar;
3475 if (no_accesses_p (repr))
3476 continue;
3477 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3478 || repr->grp_maybe_modified)
3479 continue;
3481 ao_ref_init (&ar, repr->expr);
3482 visited = BITMAP_ALLOC (NULL);
3483 for (access = repr; access; access = access->next_sibling)
3485 /* All accesses are read ones, otherwise grp_maybe_modified would
3486 be trivially set. */
3487 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3488 mark_maybe_modified, repr, &visited);
3489 if (repr->grp_maybe_modified)
3490 break;
3492 BITMAP_FREE (visited);
3497 /* Propagate distances in bb_dereferences in the opposite direction than the
3498 control flow edges, in each step storing the maximum of the current value
3499 and the minimum of all successors. These steps are repeated until the table
3500 stabilizes. Note that BBs which might terminate the functions (according to
3501 final_bbs bitmap) never updated in this way. */
3503 static void
3504 propagate_dereference_distances (void)
3506 VEC (basic_block, heap) *queue;
3507 basic_block bb;
3509 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3510 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3511 FOR_EACH_BB (bb)
3513 VEC_quick_push (basic_block, queue, bb);
3514 bb->aux = bb;
3517 while (!VEC_empty (basic_block, queue))
3519 edge_iterator ei;
3520 edge e;
3521 bool change = false;
3522 int i;
3524 bb = VEC_pop (basic_block, queue);
3525 bb->aux = NULL;
3527 if (bitmap_bit_p (final_bbs, bb->index))
3528 continue;
3530 for (i = 0; i < func_param_count; i++)
3532 int idx = bb->index * func_param_count + i;
3533 bool first = true;
3534 HOST_WIDE_INT inh = 0;
3536 FOR_EACH_EDGE (e, ei, bb->succs)
3538 int succ_idx = e->dest->index * func_param_count + i;
3540 if (e->src == EXIT_BLOCK_PTR)
3541 continue;
3543 if (first)
3545 first = false;
3546 inh = bb_dereferences [succ_idx];
3548 else if (bb_dereferences [succ_idx] < inh)
3549 inh = bb_dereferences [succ_idx];
3552 if (!first && bb_dereferences[idx] < inh)
3554 bb_dereferences[idx] = inh;
3555 change = true;
3559 if (change && !bitmap_bit_p (final_bbs, bb->index))
3560 FOR_EACH_EDGE (e, ei, bb->preds)
3562 if (e->src->aux)
3563 continue;
3565 e->src->aux = e->src;
3566 VEC_quick_push (basic_block, queue, e->src);
3570 VEC_free (basic_block, heap, queue);
3573 /* Dump a dereferences TABLE with heading STR to file F. */
3575 static void
3576 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3578 basic_block bb;
3580 fprintf (dump_file, str);
3581 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3583 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3584 if (bb != EXIT_BLOCK_PTR)
3586 int i;
3587 for (i = 0; i < func_param_count; i++)
3589 int idx = bb->index * func_param_count + i;
3590 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3593 fprintf (f, "\n");
3595 fprintf (dump_file, "\n");
3598 /* Determine what (parts of) parameters passed by reference that are not
3599 assigned to are not certainly dereferenced in this function and thus the
3600 dereferencing cannot be safely moved to the caller without potentially
3601 introducing a segfault. Mark such REPRESENTATIVES as
3602 grp_not_necessarilly_dereferenced.
3604 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3605 part is calculated rather than simple booleans are calculated for each
3606 pointer parameter to handle cases when only a fraction of the whole
3607 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3608 an example).
3610 The maximum dereference distances for each pointer parameter and BB are
3611 already stored in bb_dereference. This routine simply propagates these
3612 values upwards by propagate_dereference_distances and then compares the
3613 distances of individual parameters in the ENTRY BB to the equivalent
3614 distances of each representative of a (fraction of a) parameter. */
3616 static void
3617 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3619 int i;
3621 if (dump_file && (dump_flags & TDF_DETAILS))
3622 dump_dereferences_table (dump_file,
3623 "Dereference table before propagation:\n",
3624 bb_dereferences);
3626 propagate_dereference_distances ();
3628 if (dump_file && (dump_flags & TDF_DETAILS))
3629 dump_dereferences_table (dump_file,
3630 "Dereference table after propagation:\n",
3631 bb_dereferences);
3633 for (i = 0; i < func_param_count; i++)
3635 struct access *repr = VEC_index (access_p, representatives, i);
3636 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3638 if (!repr || no_accesses_p (repr))
3639 continue;
3643 if ((repr->offset + repr->size) > bb_dereferences[idx])
3644 repr->grp_not_necessarilly_dereferenced = 1;
3645 repr = repr->next_grp;
3647 while (repr);
3651 /* Return the representative access for the parameter declaration PARM if it is
3652 a scalar passed by reference which is not written to and the pointer value
3653 is not used directly. Thus, if it is legal to dereference it in the caller
3654 and we can rule out modifications through aliases, such parameter should be
3655 turned into one passed by value. Return NULL otherwise. */
3657 static struct access *
3658 unmodified_by_ref_scalar_representative (tree parm)
3660 int i, access_count;
3661 struct access *repr;
3662 VEC (access_p, heap) *access_vec;
3664 access_vec = get_base_access_vector (parm);
3665 gcc_assert (access_vec);
3666 repr = VEC_index (access_p, access_vec, 0);
3667 if (repr->write)
3668 return NULL;
3669 repr->group_representative = repr;
3671 access_count = VEC_length (access_p, access_vec);
3672 for (i = 1; i < access_count; i++)
3674 struct access *access = VEC_index (access_p, access_vec, i);
3675 if (access->write)
3676 return NULL;
3677 access->group_representative = repr;
3678 access->next_sibling = repr->next_sibling;
3679 repr->next_sibling = access;
3682 repr->grp_read = 1;
3683 repr->grp_scalar_ptr = 1;
3684 return repr;
3687 /* Return true iff this access precludes IPA-SRA of the parameter it is
3688 associated with. */
3690 static bool
3691 access_precludes_ipa_sra_p (struct access *access)
3693 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3694 is incompatible assign in a call statement (and possibly even in asm
3695 statements). This can be relaxed by using a new temporary but only for
3696 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3697 intraprocedural SRA we deal with this by keeping the old aggregate around,
3698 something we cannot do in IPA-SRA.) */
3699 if (access->write
3700 && (is_gimple_call (access->stmt)
3701 || gimple_code (access->stmt) == GIMPLE_ASM))
3702 return true;
3704 if (tree_non_mode_aligned_mem_p (access->expr))
3705 return true;
3707 return false;
3711 /* Sort collected accesses for parameter PARM, identify representatives for
3712 each accessed region and link them together. Return NULL if there are
3713 different but overlapping accesses, return the special ptr value meaning
3714 there are no accesses for this parameter if that is the case and return the
3715 first representative otherwise. Set *RO_GRP if there is a group of accesses
3716 with only read (i.e. no write) accesses. */
3718 static struct access *
3719 splice_param_accesses (tree parm, bool *ro_grp)
3721 int i, j, access_count, group_count;
3722 int agg_size, total_size = 0;
3723 struct access *access, *res, **prev_acc_ptr = &res;
3724 VEC (access_p, heap) *access_vec;
3726 access_vec = get_base_access_vector (parm);
3727 if (!access_vec)
3728 return &no_accesses_representant;
3729 access_count = VEC_length (access_p, access_vec);
3731 VEC_qsort (access_p, access_vec, compare_access_positions);
3733 i = 0;
3734 total_size = 0;
3735 group_count = 0;
3736 while (i < access_count)
3738 bool modification;
3739 tree a1_alias_type;
3740 access = VEC_index (access_p, access_vec, i);
3741 modification = access->write;
3742 if (access_precludes_ipa_sra_p (access))
3743 return NULL;
3744 a1_alias_type = reference_alias_ptr_type (access->expr);
3746 /* Access is about to become group representative unless we find some
3747 nasty overlap which would preclude us from breaking this parameter
3748 apart. */
3750 j = i + 1;
3751 while (j < access_count)
3753 struct access *ac2 = VEC_index (access_p, access_vec, j);
3754 if (ac2->offset != access->offset)
3756 /* All or nothing law for parameters. */
3757 if (access->offset + access->size > ac2->offset)
3758 return NULL;
3759 else
3760 break;
3762 else if (ac2->size != access->size)
3763 return NULL;
3765 if (access_precludes_ipa_sra_p (ac2)
3766 || (ac2->type != access->type
3767 && (TREE_ADDRESSABLE (ac2->type)
3768 || TREE_ADDRESSABLE (access->type)))
3769 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3770 return NULL;
3772 modification |= ac2->write;
3773 ac2->group_representative = access;
3774 ac2->next_sibling = access->next_sibling;
3775 access->next_sibling = ac2;
3776 j++;
3779 group_count++;
3780 access->grp_maybe_modified = modification;
3781 if (!modification)
3782 *ro_grp = true;
3783 *prev_acc_ptr = access;
3784 prev_acc_ptr = &access->next_grp;
3785 total_size += access->size;
3786 i = j;
3789 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3790 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3791 else
3792 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3793 if (total_size >= agg_size)
3794 return NULL;
3796 gcc_assert (group_count > 0);
3797 return res;
3800 /* Decide whether parameters with representative accesses given by REPR should
3801 be reduced into components. */
3803 static int
3804 decide_one_param_reduction (struct access *repr)
3806 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3807 bool by_ref;
3808 tree parm;
3810 parm = repr->base;
3811 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3812 gcc_assert (cur_parm_size > 0);
3814 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3816 by_ref = true;
3817 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3819 else
3821 by_ref = false;
3822 agg_size = cur_parm_size;
3825 if (dump_file)
3827 struct access *acc;
3828 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3829 print_generic_expr (dump_file, parm, 0);
3830 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3831 for (acc = repr; acc; acc = acc->next_grp)
3832 dump_access (dump_file, acc, true);
3835 total_size = 0;
3836 new_param_count = 0;
3838 for (; repr; repr = repr->next_grp)
3840 gcc_assert (parm == repr->base);
3842 /* Taking the address of a non-addressable field is verboten. */
3843 if (by_ref && repr->non_addressable)
3844 return 0;
3846 if (!by_ref || (!repr->grp_maybe_modified
3847 && !repr->grp_not_necessarilly_dereferenced))
3848 total_size += repr->size;
3849 else
3850 total_size += cur_parm_size;
3852 new_param_count++;
3855 gcc_assert (new_param_count > 0);
3857 if (optimize_function_for_size_p (cfun))
3858 parm_size_limit = cur_parm_size;
3859 else
3860 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3861 * cur_parm_size);
3863 if (total_size < agg_size
3864 && total_size <= parm_size_limit)
3866 if (dump_file)
3867 fprintf (dump_file, " ....will be split into %i components\n",
3868 new_param_count);
3869 return new_param_count;
3871 else
3872 return 0;
3875 /* The order of the following enums is important, we need to do extra work for
3876 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3877 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3878 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3880 /* Identify representatives of all accesses to all candidate parameters for
3881 IPA-SRA. Return result based on what representatives have been found. */
3883 static enum ipa_splicing_result
3884 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3886 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3887 tree parm;
3888 struct access *repr;
3890 *representatives = VEC_alloc (access_p, heap, func_param_count);
3892 for (parm = DECL_ARGUMENTS (current_function_decl);
3893 parm;
3894 parm = DECL_CHAIN (parm))
3896 if (is_unused_scalar_param (parm))
3898 VEC_quick_push (access_p, *representatives,
3899 &no_accesses_representant);
3900 if (result == NO_GOOD_ACCESS)
3901 result = UNUSED_PARAMS;
3903 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3904 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3905 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3907 repr = unmodified_by_ref_scalar_representative (parm);
3908 VEC_quick_push (access_p, *representatives, repr);
3909 if (repr)
3910 result = UNMODIF_BY_REF_ACCESSES;
3912 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3914 bool ro_grp = false;
3915 repr = splice_param_accesses (parm, &ro_grp);
3916 VEC_quick_push (access_p, *representatives, repr);
3918 if (repr && !no_accesses_p (repr))
3920 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3922 if (ro_grp)
3923 result = UNMODIF_BY_REF_ACCESSES;
3924 else if (result < MODIF_BY_REF_ACCESSES)
3925 result = MODIF_BY_REF_ACCESSES;
3927 else if (result < BY_VAL_ACCESSES)
3928 result = BY_VAL_ACCESSES;
3930 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3931 result = UNUSED_PARAMS;
3933 else
3934 VEC_quick_push (access_p, *representatives, NULL);
3937 if (result == NO_GOOD_ACCESS)
3939 VEC_free (access_p, heap, *representatives);
3940 *representatives = NULL;
3941 return NO_GOOD_ACCESS;
3944 return result;
3947 /* Return the index of BASE in PARMS. Abort if it is not found. */
3949 static inline int
3950 get_param_index (tree base, VEC(tree, heap) *parms)
3952 int i, len;
3954 len = VEC_length (tree, parms);
3955 for (i = 0; i < len; i++)
3956 if (VEC_index (tree, parms, i) == base)
3957 return i;
3958 gcc_unreachable ();
3961 /* Convert the decisions made at the representative level into compact
3962 parameter adjustments. REPRESENTATIVES are pointers to first
3963 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3964 final number of adjustments. */
3966 static ipa_parm_adjustment_vec
3967 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3968 int adjustments_count)
3970 VEC (tree, heap) *parms;
3971 ipa_parm_adjustment_vec adjustments;
3972 tree parm;
3973 int i;
3975 gcc_assert (adjustments_count > 0);
3976 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3977 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3978 parm = DECL_ARGUMENTS (current_function_decl);
3979 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
3981 struct access *repr = VEC_index (access_p, representatives, i);
3983 if (!repr || no_accesses_p (repr))
3985 struct ipa_parm_adjustment *adj;
3987 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3988 memset (adj, 0, sizeof (*adj));
3989 adj->base_index = get_param_index (parm, parms);
3990 adj->base = parm;
3991 if (!repr)
3992 adj->copy_param = 1;
3993 else
3994 adj->remove_param = 1;
3996 else
3998 struct ipa_parm_adjustment *adj;
3999 int index = get_param_index (parm, parms);
4001 for (; repr; repr = repr->next_grp)
4003 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
4004 memset (adj, 0, sizeof (*adj));
4005 gcc_assert (repr->base == parm);
4006 adj->base_index = index;
4007 adj->base = repr->base;
4008 adj->type = repr->type;
4009 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
4010 adj->offset = repr->offset;
4011 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
4012 && (repr->grp_maybe_modified
4013 || repr->grp_not_necessarilly_dereferenced));
4018 VEC_free (tree, heap, parms);
4019 return adjustments;
4022 /* Analyze the collected accesses and produce a plan what to do with the
4023 parameters in the form of adjustments, NULL meaning nothing. */
4025 static ipa_parm_adjustment_vec
4026 analyze_all_param_acesses (void)
4028 enum ipa_splicing_result repr_state;
4029 bool proceed = false;
4030 int i, adjustments_count = 0;
4031 VEC (access_p, heap) *representatives;
4032 ipa_parm_adjustment_vec adjustments;
4034 repr_state = splice_all_param_accesses (&representatives);
4035 if (repr_state == NO_GOOD_ACCESS)
4036 return NULL;
4038 /* If there are any parameters passed by reference which are not modified
4039 directly, we need to check whether they can be modified indirectly. */
4040 if (repr_state == UNMODIF_BY_REF_ACCESSES)
4042 analyze_caller_dereference_legality (representatives);
4043 analyze_modified_params (representatives);
4046 for (i = 0; i < func_param_count; i++)
4048 struct access *repr = VEC_index (access_p, representatives, i);
4050 if (repr && !no_accesses_p (repr))
4052 if (repr->grp_scalar_ptr)
4054 adjustments_count++;
4055 if (repr->grp_not_necessarilly_dereferenced
4056 || repr->grp_maybe_modified)
4057 VEC_replace (access_p, representatives, i, NULL);
4058 else
4060 proceed = true;
4061 sra_stats.scalar_by_ref_to_by_val++;
4064 else
4066 int new_components = decide_one_param_reduction (repr);
4068 if (new_components == 0)
4070 VEC_replace (access_p, representatives, i, NULL);
4071 adjustments_count++;
4073 else
4075 adjustments_count += new_components;
4076 sra_stats.aggregate_params_reduced++;
4077 sra_stats.param_reductions_created += new_components;
4078 proceed = true;
4082 else
4084 if (no_accesses_p (repr))
4086 proceed = true;
4087 sra_stats.deleted_unused_parameters++;
4089 adjustments_count++;
4093 if (!proceed && dump_file)
4094 fprintf (dump_file, "NOT proceeding to change params.\n");
4096 if (proceed)
4097 adjustments = turn_representatives_into_adjustments (representatives,
4098 adjustments_count);
4099 else
4100 adjustments = NULL;
4102 VEC_free (access_p, heap, representatives);
4103 return adjustments;
4106 /* If a parameter replacement identified by ADJ does not yet exist in the form
4107 of declaration, create it and record it, otherwise return the previously
4108 created one. */
4110 static tree
4111 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4113 tree repl;
4114 if (!adj->new_ssa_base)
4116 char *pretty_name = make_fancy_name (adj->base);
4118 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4119 DECL_NAME (repl) = get_identifier (pretty_name);
4120 obstack_free (&name_obstack, pretty_name);
4122 add_referenced_var (repl);
4123 adj->new_ssa_base = repl;
4125 else
4126 repl = adj->new_ssa_base;
4127 return repl;
4130 /* Find the first adjustment for a particular parameter BASE in a vector of
4131 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4132 adjustment. */
4134 static struct ipa_parm_adjustment *
4135 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4137 int i, len;
4139 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4140 for (i = 0; i < len; i++)
4142 struct ipa_parm_adjustment *adj;
4144 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4145 if (!adj->copy_param && adj->base == base)
4146 return adj;
4149 return NULL;
4152 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4153 removed because its value is not used, replace the SSA_NAME with a one
4154 relating to a created VAR_DECL together all of its uses and return true.
4155 ADJUSTMENTS is a pointer to an adjustments vector. */
4157 static bool
4158 replace_removed_params_ssa_names (gimple stmt,
4159 ipa_parm_adjustment_vec adjustments)
4161 struct ipa_parm_adjustment *adj;
4162 tree lhs, decl, repl, name;
4164 if (gimple_code (stmt) == GIMPLE_PHI)
4165 lhs = gimple_phi_result (stmt);
4166 else if (is_gimple_assign (stmt))
4167 lhs = gimple_assign_lhs (stmt);
4168 else if (is_gimple_call (stmt))
4169 lhs = gimple_call_lhs (stmt);
4170 else
4171 gcc_unreachable ();
4173 if (TREE_CODE (lhs) != SSA_NAME)
4174 return false;
4175 decl = SSA_NAME_VAR (lhs);
4176 if (TREE_CODE (decl) != PARM_DECL)
4177 return false;
4179 adj = get_adjustment_for_base (adjustments, decl);
4180 if (!adj)
4181 return false;
4183 repl = get_replaced_param_substitute (adj);
4184 name = make_ssa_name (repl, stmt);
4186 if (dump_file)
4188 fprintf (dump_file, "replacing an SSA name of a removed param ");
4189 print_generic_expr (dump_file, lhs, 0);
4190 fprintf (dump_file, " with ");
4191 print_generic_expr (dump_file, name, 0);
4192 fprintf (dump_file, "\n");
4195 if (is_gimple_assign (stmt))
4196 gimple_assign_set_lhs (stmt, name);
4197 else if (is_gimple_call (stmt))
4198 gimple_call_set_lhs (stmt, name);
4199 else
4200 gimple_phi_set_result (stmt, name);
4202 replace_uses_by (lhs, name);
4203 release_ssa_name (lhs);
4204 return true;
4207 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4208 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4209 specifies whether the function should care about type incompatibility the
4210 current and new expressions. If it is false, the function will leave
4211 incompatibility issues to the caller. Return true iff the expression
4212 was modified. */
4214 static bool
4215 sra_ipa_modify_expr (tree *expr, bool convert,
4216 ipa_parm_adjustment_vec adjustments)
4218 int i, len;
4219 struct ipa_parm_adjustment *adj, *cand = NULL;
4220 HOST_WIDE_INT offset, size, max_size;
4221 tree base, src;
4223 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4225 if (TREE_CODE (*expr) == BIT_FIELD_REF
4226 || TREE_CODE (*expr) == IMAGPART_EXPR
4227 || TREE_CODE (*expr) == REALPART_EXPR)
4229 expr = &TREE_OPERAND (*expr, 0);
4230 convert = true;
4233 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4234 if (!base || size == -1 || max_size == -1)
4235 return false;
4237 if (TREE_CODE (base) == MEM_REF)
4239 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4240 base = TREE_OPERAND (base, 0);
4243 base = get_ssa_base_param (base);
4244 if (!base || TREE_CODE (base) != PARM_DECL)
4245 return false;
4247 for (i = 0; i < len; i++)
4249 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4251 if (adj->base == base &&
4252 (adj->offset == offset || adj->remove_param))
4254 cand = adj;
4255 break;
4258 if (!cand || cand->copy_param || cand->remove_param)
4259 return false;
4261 if (cand->by_ref)
4262 src = build_simple_mem_ref (cand->reduction);
4263 else
4264 src = cand->reduction;
4266 if (dump_file && (dump_flags & TDF_DETAILS))
4268 fprintf (dump_file, "About to replace expr ");
4269 print_generic_expr (dump_file, *expr, 0);
4270 fprintf (dump_file, " with ");
4271 print_generic_expr (dump_file, src, 0);
4272 fprintf (dump_file, "\n");
4275 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4277 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4278 *expr = vce;
4280 else
4281 *expr = src;
4282 return true;
4285 /* If the statement pointed to by STMT_PTR contains any expressions that need
4286 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4287 potential type incompatibilities (GSI is used to accommodate conversion
4288 statements and must point to the statement). Return true iff the statement
4289 was modified. */
4291 static bool
4292 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4293 ipa_parm_adjustment_vec adjustments)
4295 gimple stmt = *stmt_ptr;
4296 tree *lhs_p, *rhs_p;
4297 bool any;
4299 if (!gimple_assign_single_p (stmt))
4300 return false;
4302 rhs_p = gimple_assign_rhs1_ptr (stmt);
4303 lhs_p = gimple_assign_lhs_ptr (stmt);
4305 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4306 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4307 if (any)
4309 tree new_rhs = NULL_TREE;
4311 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4313 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4315 /* V_C_Es of constructors can cause trouble (PR 42714). */
4316 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4317 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4318 else
4319 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4321 else
4322 new_rhs = fold_build1_loc (gimple_location (stmt),
4323 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4324 *rhs_p);
4326 else if (REFERENCE_CLASS_P (*rhs_p)
4327 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4328 && !is_gimple_reg (*lhs_p))
4329 /* This can happen when an assignment in between two single field
4330 structures is turned into an assignment in between two pointers to
4331 scalars (PR 42237). */
4332 new_rhs = *rhs_p;
4334 if (new_rhs)
4336 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4337 true, GSI_SAME_STMT);
4339 gimple_assign_set_rhs_from_tree (gsi, tmp);
4342 return true;
4345 return false;
4348 /* Traverse the function body and all modifications as described in
4349 ADJUSTMENTS. Return true iff the CFG has been changed. */
4351 static bool
4352 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4354 bool cfg_changed = false;
4355 basic_block bb;
4357 FOR_EACH_BB (bb)
4359 gimple_stmt_iterator gsi;
4361 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4362 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4364 gsi = gsi_start_bb (bb);
4365 while (!gsi_end_p (gsi))
4367 gimple stmt = gsi_stmt (gsi);
4368 bool modified = false;
4369 tree *t;
4370 unsigned i;
4372 switch (gimple_code (stmt))
4374 case GIMPLE_RETURN:
4375 t = gimple_return_retval_ptr (stmt);
4376 if (*t != NULL_TREE)
4377 modified |= sra_ipa_modify_expr (t, true, adjustments);
4378 break;
4380 case GIMPLE_ASSIGN:
4381 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4382 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4383 break;
4385 case GIMPLE_CALL:
4386 /* Operands must be processed before the lhs. */
4387 for (i = 0; i < gimple_call_num_args (stmt); i++)
4389 t = gimple_call_arg_ptr (stmt, i);
4390 modified |= sra_ipa_modify_expr (t, true, adjustments);
4393 if (gimple_call_lhs (stmt))
4395 t = gimple_call_lhs_ptr (stmt);
4396 modified |= sra_ipa_modify_expr (t, false, adjustments);
4397 modified |= replace_removed_params_ssa_names (stmt,
4398 adjustments);
4400 break;
4402 case GIMPLE_ASM:
4403 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4405 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4406 modified |= sra_ipa_modify_expr (t, true, adjustments);
4408 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4410 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4411 modified |= sra_ipa_modify_expr (t, false, adjustments);
4413 break;
4415 default:
4416 break;
4419 if (modified)
4421 update_stmt (stmt);
4422 if (maybe_clean_eh_stmt (stmt)
4423 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4424 cfg_changed = true;
4426 gsi_next (&gsi);
4430 return cfg_changed;
4433 /* Call gimple_debug_bind_reset_value on all debug statements describing
4434 gimple register parameters that are being removed or replaced. */
4436 static void
4437 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4439 int i, len;
4440 gimple_stmt_iterator *gsip = NULL, gsi;
4442 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR))
4444 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
4445 gsip = &gsi;
4447 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4448 for (i = 0; i < len; i++)
4450 struct ipa_parm_adjustment *adj;
4451 imm_use_iterator ui;
4452 gimple stmt, def_temp;
4453 tree name, vexpr, copy = NULL_TREE;
4454 use_operand_p use_p;
4456 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4457 if (adj->copy_param || !is_gimple_reg (adj->base))
4458 continue;
4459 name = gimple_default_def (cfun, adj->base);
4460 vexpr = NULL;
4461 if (name)
4462 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4464 /* All other users must have been removed by
4465 ipa_sra_modify_function_body. */
4466 gcc_assert (is_gimple_debug (stmt));
4467 if (vexpr == NULL && gsip != NULL)
4469 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4470 vexpr = make_node (DEBUG_EXPR_DECL);
4471 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4472 NULL);
4473 DECL_ARTIFICIAL (vexpr) = 1;
4474 TREE_TYPE (vexpr) = TREE_TYPE (name);
4475 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4476 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4478 if (vexpr)
4480 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4481 SET_USE (use_p, vexpr);
4483 else
4484 gimple_debug_bind_reset_value (stmt);
4485 update_stmt (stmt);
4487 /* Create a VAR_DECL for debug info purposes. */
4488 if (!DECL_IGNORED_P (adj->base))
4490 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4491 VAR_DECL, DECL_NAME (adj->base),
4492 TREE_TYPE (adj->base));
4493 if (DECL_PT_UID_SET_P (adj->base))
4494 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4495 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4496 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4497 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4498 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4499 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4500 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4501 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4502 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4503 SET_DECL_RTL (copy, 0);
4504 TREE_USED (copy) = 1;
4505 DECL_CONTEXT (copy) = current_function_decl;
4506 add_referenced_var (copy);
4507 add_local_decl (cfun, copy);
4508 DECL_CHAIN (copy) =
4509 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4510 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4512 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4514 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4515 if (vexpr)
4516 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4517 else
4518 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4519 NULL);
4520 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4525 /* Return false iff all callers have at least as many actual arguments as there
4526 are formal parameters in the current function. */
4528 static bool
4529 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4530 void *data ATTRIBUTE_UNUSED)
4532 struct cgraph_edge *cs;
4533 for (cs = node->callers; cs; cs = cs->next_caller)
4534 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4535 return true;
4537 return false;
4540 /* Convert all callers of NODE. */
4542 static bool
4543 convert_callers_for_node (struct cgraph_node *node,
4544 void *data)
4546 ipa_parm_adjustment_vec adjustments = (ipa_parm_adjustment_vec)data;
4547 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4548 struct cgraph_edge *cs;
4550 for (cs = node->callers; cs; cs = cs->next_caller)
4552 current_function_decl = cs->caller->decl;
4553 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4555 if (dump_file)
4556 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4557 cs->caller->uid, cs->callee->uid,
4558 cgraph_node_name (cs->caller),
4559 cgraph_node_name (cs->callee));
4561 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4563 pop_cfun ();
4566 for (cs = node->callers; cs; cs = cs->next_caller)
4567 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4568 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4569 compute_inline_parameters (cs->caller, true);
4570 BITMAP_FREE (recomputed_callers);
4572 return true;
4575 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4577 static void
4578 convert_callers (struct cgraph_node *node, tree old_decl,
4579 ipa_parm_adjustment_vec adjustments)
4581 tree old_cur_fndecl = current_function_decl;
4582 basic_block this_block;
4584 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4585 adjustments, false);
4587 current_function_decl = old_cur_fndecl;
4589 if (!encountered_recursive_call)
4590 return;
4592 FOR_EACH_BB (this_block)
4594 gimple_stmt_iterator gsi;
4596 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4598 gimple stmt = gsi_stmt (gsi);
4599 tree call_fndecl;
4600 if (gimple_code (stmt) != GIMPLE_CALL)
4601 continue;
4602 call_fndecl = gimple_call_fndecl (stmt);
4603 if (call_fndecl == old_decl)
4605 if (dump_file)
4606 fprintf (dump_file, "Adjusting recursive call");
4607 gimple_call_set_fndecl (stmt, node->decl);
4608 ipa_modify_call_arguments (NULL, stmt, adjustments);
4613 return;
4616 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4617 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4619 static bool
4620 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4622 struct cgraph_node *new_node;
4623 bool cfg_changed;
4624 VEC (cgraph_edge_p, heap) * redirect_callers = collect_callers_of_node (node);
4626 rebuild_cgraph_edges ();
4627 free_dominance_info (CDI_DOMINATORS);
4628 pop_cfun ();
4629 current_function_decl = NULL_TREE;
4631 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4632 NULL, NULL, "isra");
4633 current_function_decl = new_node->decl;
4634 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4636 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4637 cfg_changed = ipa_sra_modify_function_body (adjustments);
4638 sra_ipa_reset_debug_stmts (adjustments);
4639 convert_callers (new_node, node->decl, adjustments);
4640 cgraph_make_node_local (new_node);
4641 return cfg_changed;
4644 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4645 attributes, return true otherwise. NODE is the cgraph node of the current
4646 function. */
4648 static bool
4649 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4651 if (!cgraph_node_can_be_local_p (node))
4653 if (dump_file)
4654 fprintf (dump_file, "Function not local to this compilation unit.\n");
4655 return false;
4658 if (!node->local.can_change_signature)
4660 if (dump_file)
4661 fprintf (dump_file, "Function can not change signature.\n");
4662 return false;
4665 if (!tree_versionable_function_p (node->decl))
4667 if (dump_file)
4668 fprintf (dump_file, "Function is not versionable.\n");
4669 return false;
4672 if (DECL_VIRTUAL_P (current_function_decl))
4674 if (dump_file)
4675 fprintf (dump_file, "Function is a virtual method.\n");
4676 return false;
4679 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4680 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4682 if (dump_file)
4683 fprintf (dump_file, "Function too big to be made truly local.\n");
4684 return false;
4687 if (!node->callers)
4689 if (dump_file)
4690 fprintf (dump_file,
4691 "Function has no callers in this compilation unit.\n");
4692 return false;
4695 if (cfun->stdarg)
4697 if (dump_file)
4698 fprintf (dump_file, "Function uses stdarg. \n");
4699 return false;
4702 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4703 return false;
4705 return true;
4708 /* Perform early interprocedural SRA. */
4710 static unsigned int
4711 ipa_early_sra (void)
4713 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4714 ipa_parm_adjustment_vec adjustments;
4715 int ret = 0;
4717 if (!ipa_sra_preliminary_function_checks (node))
4718 return 0;
4720 sra_initialize ();
4721 sra_mode = SRA_MODE_EARLY_IPA;
4723 if (!find_param_candidates ())
4725 if (dump_file)
4726 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4727 goto simple_out;
4730 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4731 NULL, true))
4733 if (dump_file)
4734 fprintf (dump_file, "There are callers with insufficient number of "
4735 "arguments.\n");
4736 goto simple_out;
4739 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4740 func_param_count
4741 * last_basic_block_for_function (cfun));
4742 final_bbs = BITMAP_ALLOC (NULL);
4744 scan_function ();
4745 if (encountered_apply_args)
4747 if (dump_file)
4748 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4749 goto out;
4752 if (encountered_unchangable_recursive_call)
4754 if (dump_file)
4755 fprintf (dump_file, "Function calls itself with insufficient "
4756 "number of arguments.\n");
4757 goto out;
4760 adjustments = analyze_all_param_acesses ();
4761 if (!adjustments)
4762 goto out;
4763 if (dump_file)
4764 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4766 if (modify_function (node, adjustments))
4767 ret = TODO_update_ssa | TODO_cleanup_cfg;
4768 else
4769 ret = TODO_update_ssa;
4770 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4772 statistics_counter_event (cfun, "Unused parameters deleted",
4773 sra_stats.deleted_unused_parameters);
4774 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4775 sra_stats.scalar_by_ref_to_by_val);
4776 statistics_counter_event (cfun, "Aggregate parameters broken up",
4777 sra_stats.aggregate_params_reduced);
4778 statistics_counter_event (cfun, "Aggregate parameter components created",
4779 sra_stats.param_reductions_created);
4781 out:
4782 BITMAP_FREE (final_bbs);
4783 free (bb_dereferences);
4784 simple_out:
4785 sra_deinitialize ();
4786 return ret;
4789 /* Return if early ipa sra shall be performed. */
4790 static bool
4791 ipa_early_sra_gate (void)
4793 return flag_ipa_sra && dbg_cnt (eipa_sra);
4796 struct gimple_opt_pass pass_early_ipa_sra =
4799 GIMPLE_PASS,
4800 "eipa_sra", /* name */
4801 ipa_early_sra_gate, /* gate */
4802 ipa_early_sra, /* execute */
4803 NULL, /* sub */
4804 NULL, /* next */
4805 0, /* static_pass_number */
4806 TV_IPA_SRA, /* tv_id */
4807 0, /* properties_required */
4808 0, /* properties_provided */
4809 0, /* properties_destroyed */
4810 0, /* todo_flags_start */
4811 TODO_dump_cgraph /* todo_flags_finish */