re PR tree-optimization/49094 (ARM aligned(1) attribute is sometimes dropped)
[gcc.git] / gcc / tree-sra.c
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>
6
7 This file is part of GCC.
8
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.
13
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.
18
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/>. */
22
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.
27
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.
32
33 Both passes operate in four stages:
34
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.
38
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.
46
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.
50
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.
55
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).
60
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.
64
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.
67
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. */
73
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"
95
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 */
100
101 /* Global variable describing which aggregate reduction we are performing at
102 the moment. */
103 static enum sra_mode sra_mode;
104
105 struct assign_link;
106
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.
112
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.
118
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. */
122
123 struct access
124 {
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;
131
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;
138
139 /* The statement this access belongs to. */
140 gimple stmt;
141
142 /* Next group representative for this aggregate. */
143 struct access *next_grp;
144
145 /* Pointer to the group representative. Pointer to itself if the struct is
146 the representative. */
147 struct access *group_representative;
148
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;
152
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;
157
158 /* Pointers to the first and last element in the linked list of assign
159 links. */
160 struct assign_link *first_link, *last_link;
161
162 /* Pointer to the next access in the work queue. */
163 struct access *next_queued;
164
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;
169
170 /* Is this particular access write access? */
171 unsigned write : 1;
172
173 /* Is this access an artificial one created to scalarize some record
174 entirely? */
175 unsigned total_scalarization : 1;
176
177 /* Is this access an access to a non-addressable field? */
178 unsigned non_addressable : 1;
179
180 /* Is this access currently in the work queue? */
181 unsigned grp_queued : 1;
182
183 /* Does this group contain a write access? This flag is propagated down the
184 access tree. */
185 unsigned grp_write : 1;
186
187 /* Does this group contain a read access? This flag is propagated down the
188 access tree. */
189 unsigned grp_read : 1;
190
191 /* Does this group contain a read access that comes from an assignment
192 statement? This flag is propagated down the access tree. */
193 unsigned grp_assignment_read : 1;
194
195 /* Does this group contain a write access that comes from an assignment
196 statement? This flag is propagated down the access tree. */
197 unsigned grp_assignment_write : 1;
198
199 /* Does this group contain a read access through a scalar type? This flag is
200 not propagated in the access tree in any direction. */
201 unsigned grp_scalar_read : 1;
202
203 /* Does this group contain a write access through a scalar type? This flag
204 is not propagated in the access tree in any direction. */
205 unsigned grp_scalar_write : 1;
206
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;
211
212 /* Is the subtree rooted in this access fully covered by scalar
213 replacements? */
214 unsigned grp_covered : 1;
215
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;
219
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;
224
225 /* Does this access and/or group contain a write access through a
226 BIT_FIELD_REF? */
227 unsigned grp_partial_lhs : 1;
228
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;
233
234 /* Should TREE_NO_WARNING of a replacement be set? */
235 unsigned grp_no_warning : 1;
236
237 /* Is it possible that the group refers to data which might be (directly or
238 otherwise) modified? */
239 unsigned grp_maybe_modified : 1;
240
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;
245
246 /* Set when we discover that this pointer is not safe to dereference in the
247 caller. */
248 unsigned grp_not_necessarilly_dereferenced : 1;
249 };
250
251 typedef struct access *access_p;
252
253 DEF_VEC_P (access_p);
254 DEF_VEC_ALLOC_P (access_p, heap);
255
256 /* Alloc pool for allocating access structures. */
257 static alloc_pool access_pool;
258
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
263 {
264 struct access *lacc, *racc;
265 struct assign_link *next;
266 };
267
268 /* Alloc pool for allocating assign link structures. */
269 static alloc_pool link_pool;
270
271 /* Base (tree) -> Vector (VEC(access_p,heap) *) map. */
272 static struct pointer_map_t *base_access_vec;
273
274 /* Bitmap of candidates. */
275 static bitmap candidate_bitmap;
276
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;
280
281 /* Obstack for creation of fancy names. */
282 static struct obstack name_obstack;
283
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;
287
288 /* Number of parameters of the analyzed function when doing early ipa SRA. */
289 static int func_param_count;
290
291 /* scan_function sets the following to true if it encounters a call to
292 __builtin_apply_args. */
293 static bool encountered_apply_args;
294
295 /* Set by scan_function when it finds a recursive call. */
296 static bool encountered_recursive_call;
297
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;
301
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;
310
311 /* Representative of no accesses at all. */
312 static struct access no_accesses_representant;
313
314 /* Predicate to test the special value. */
315
316 static inline bool
317 no_accesses_p (struct access *access)
318 {
319 return access == &no_accesses_representant;
320 }
321
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. */
325
326 static struct
327 {
328 /* Number of processed aggregates is readily available in
329 analyze_all_variable_accesses and so is not stored here. */
330
331 /* Number of created scalar replacements. */
332 int replacements;
333
334 /* Number of times sra_modify_expr or sra_modify_assign themselves changed an
335 expression. */
336 int exprs;
337
338 /* Number of statements created by generate_subtree_copies. */
339 int subtree_copies;
340
341 /* Number of statements created by load_assign_lhs_subreplacements. */
342 int subreplacements;
343
344 /* Number of times sra_modify_assign has deleted a statement. */
345 int deleted;
346
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;
351
352 /* Number of parameters that were removed because they were unused. */
353 int deleted_unused_parameters;
354
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;
358
359 /* Number of aggregate parameters that were replaced by one or more of their
360 components. */
361 int aggregate_params_reduced;
362
363 /* Numbber of components created when splitting aggregate parameters. */
364 int param_reductions_created;
365 } sra_stats;
366
367 static void
368 dump_access (FILE *f, struct access *access, bool grp)
369 {
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, ", total_scalarization = %d, grp_read = %d, grp_write = %d, "
381 "grp_assignment_read = %d, grp_assignment_write = %d, "
382 "grp_scalar_read = %d, grp_scalar_write = %d, "
383 "grp_hint = %d, grp_covered = %d, "
384 "grp_unscalarizable_region = %d, grp_unscalarized_data = %d, "
385 "grp_partial_lhs = %d, grp_to_be_replaced = %d, "
386 "grp_maybe_modified = %d, "
387 "grp_not_necessarilly_dereferenced = %d\n",
388 access->total_scalarization, access->grp_read, access->grp_write,
389 access->grp_assignment_read, access->grp_assignment_write,
390 access->grp_scalar_read, access->grp_scalar_write,
391 access->grp_hint, access->grp_covered,
392 access->grp_unscalarizable_region, access->grp_unscalarized_data,
393 access->grp_partial_lhs, access->grp_to_be_replaced,
394 access->grp_maybe_modified,
395 access->grp_not_necessarilly_dereferenced);
396 else
397 fprintf (f, ", write = %d, total_scalarization = %d, "
398 "grp_partial_lhs = %d\n",
399 access->write, access->total_scalarization,
400 access->grp_partial_lhs);
401 }
402
403 /* Dump a subtree rooted in ACCESS to file F, indent by LEVEL. */
404
405 static void
406 dump_access_tree_1 (FILE *f, struct access *access, int level)
407 {
408 do
409 {
410 int i;
411
412 for (i = 0; i < level; i++)
413 fputs ("* ", dump_file);
414
415 dump_access (f, access, true);
416
417 if (access->first_child)
418 dump_access_tree_1 (f, access->first_child, level + 1);
419
420 access = access->next_sibling;
421 }
422 while (access);
423 }
424
425 /* Dump all access trees for a variable, given the pointer to the first root in
426 ACCESS. */
427
428 static void
429 dump_access_tree (FILE *f, struct access *access)
430 {
431 for (; access; access = access->next_grp)
432 dump_access_tree_1 (f, access, 0);
433 }
434
435 /* Return true iff ACC is non-NULL and has subaccesses. */
436
437 static inline bool
438 access_has_children_p (struct access *acc)
439 {
440 return acc && acc->first_child;
441 }
442
443 /* Return a vector of pointers to accesses for the variable given in BASE or
444 NULL if there is none. */
445
446 static VEC (access_p, heap) *
447 get_base_access_vector (tree base)
448 {
449 void **slot;
450
451 slot = pointer_map_contains (base_access_vec, base);
452 if (!slot)
453 return NULL;
454 else
455 return *(VEC (access_p, heap) **) slot;
456 }
457
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. */
460
461 static struct access *
462 find_access_in_subtree (struct access *access, HOST_WIDE_INT offset,
463 HOST_WIDE_INT size)
464 {
465 while (access && (access->offset != offset || access->size != size))
466 {
467 struct access *child = access->first_child;
468
469 while (child && (child->offset + child->size <= offset))
470 child = child->next_sibling;
471 access = child;
472 }
473
474 return access;
475 }
476
477 /* Return the first group representative for DECL or NULL if none exists. */
478
479 static struct access *
480 get_first_repr_for_decl (tree base)
481 {
482 VEC (access_p, heap) *access_vec;
483
484 access_vec = get_base_access_vector (base);
485 if (!access_vec)
486 return NULL;
487
488 return VEC_index (access_p, access_vec, 0);
489 }
490
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. */
494
495 static struct access *
496 get_var_base_offset_size_access (tree base, HOST_WIDE_INT offset,
497 HOST_WIDE_INT size)
498 {
499 struct access *access;
500
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;
506
507 return find_access_in_subtree (access, offset, size);
508 }
509
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)
513 {
514 gcc_assert (link->racc == racc);
515
516 if (!racc->first_link)
517 {
518 gcc_assert (!racc->last_link);
519 racc->first_link = link;
520 }
521 else
522 racc->last_link->next = link;
523
524 racc->last_link = link;
525 link->next = NULL;
526 }
527
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)
532 {
533 if (!old_racc->first_link)
534 {
535 gcc_assert (!old_racc->last_link);
536 return;
537 }
538
539 if (new_racc->first_link)
540 {
541 gcc_assert (!new_racc->last_link->next);
542 gcc_assert (!old_racc->last_link || !old_racc->last_link->next);
543
544 new_racc->last_link->next = old_racc->first_link;
545 new_racc->last_link = old_racc->last_link;
546 }
547 else
548 {
549 gcc_assert (!new_racc->last_link);
550
551 new_racc->first_link = old_racc->first_link;
552 new_racc->last_link = old_racc->last_link;
553 }
554 old_racc->first_link = old_racc->last_link = NULL;
555 }
556
557 /* Add ACCESS to the work queue (which is actually a stack). */
558
559 static void
560 add_access_to_work_queue (struct access *access)
561 {
562 if (!access->grp_queued)
563 {
564 gcc_assert (!access->next_queued);
565 access->next_queued = work_queue_head;
566 access->grp_queued = 1;
567 work_queue_head = access;
568 }
569 }
570
571 /* Pop an access from the work queue, and return it, assuming there is one. */
572
573 static struct access *
574 pop_access_from_work_queue (void)
575 {
576 struct access *access = work_queue_head;
577
578 work_queue_head = access->next_queued;
579 access->next_queued = NULL;
580 access->grp_queued = 0;
581 return access;
582 }
583
584
585 /* Allocate necessary structures. */
586
587 static void
588 sra_initialize (void)
589 {
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;
601 }
602
603 /* Hook fed to pointer_map_traverse, deallocate stored vectors. */
604
605 static bool
606 delete_base_accesses (const void *key ATTRIBUTE_UNUSED, void **value,
607 void *data ATTRIBUTE_UNUSED)
608 {
609 VEC (access_p, heap) *access_vec;
610 access_vec = (VEC (access_p, heap) *) *value;
611 VEC_free (access_p, heap, access_vec);
612
613 return true;
614 }
615
616 /* Deallocate all general structures. */
617
618 static void
619 sra_deinitialize (void)
620 {
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);
627
628 pointer_map_traverse (base_access_vec, delete_base_accesses, NULL);
629 pointer_map_destroy (base_access_vec);
630 }
631
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)
636 {
637 bitmap_clear_bit (candidate_bitmap, DECL_UID (decl));
638
639 if (dump_file && (dump_flags & TDF_DETAILS))
640 {
641 fprintf (dump_file, "! Disqualifying ");
642 print_generic_expr (dump_file, decl, 0);
643 fprintf (dump_file, " - %s\n", reason);
644 }
645 }
646
647 /* Return true iff the type contains a field or an element which does not allow
648 scalarization. */
649
650 static bool
651 type_internals_preclude_sra_p (tree type, const char **msg)
652 {
653 tree fld;
654 tree et;
655
656 switch (TREE_CODE (type))
657 {
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)
663 {
664 tree ft = TREE_TYPE (fld);
665
666 if (TREE_THIS_VOLATILE (fld))
667 {
668 *msg = "volatile structure field";
669 return true;
670 }
671 if (!DECL_FIELD_OFFSET (fld))
672 {
673 *msg = "no structure field offset";
674 return true;
675 }
676 if (!DECL_SIZE (fld))
677 {
678 *msg = "zero structure field size";
679 return true;
680 }
681 if (!host_integerp (DECL_FIELD_OFFSET (fld), 1))
682 {
683 *msg = "structure field offset not fixed";
684 return true;
685 }
686 if (!host_integerp (DECL_SIZE (fld), 1))
687 {
688 *msg = "structure field size not fixed";
689 return true;
690 }
691 if (AGGREGATE_TYPE_P (ft)
692 && int_bit_position (fld) % BITS_PER_UNIT != 0)
693 {
694 *msg = "structure field is bit field";
695 return true;
696 }
697
698 if (AGGREGATE_TYPE_P (ft) && type_internals_preclude_sra_p (ft, msg))
699 return true;
700 }
701
702 return false;
703
704 case ARRAY_TYPE:
705 et = TREE_TYPE (type);
706
707 if (TYPE_VOLATILE (et))
708 {
709 *msg = "element type is volatile";
710 return true;
711 }
712
713 if (AGGREGATE_TYPE_P (et) && type_internals_preclude_sra_p (et, msg))
714 return true;
715
716 return false;
717
718 default:
719 return false;
720 }
721 }
722
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. */
725
726 static tree
727 get_ssa_base_param (tree t)
728 {
729 if (TREE_CODE (t) == SSA_NAME)
730 {
731 if (SSA_NAME_IS_DEFAULT_DEF (t))
732 return SSA_NAME_VAR (t);
733 else
734 return NULL_TREE;
735 }
736 return t;
737 }
738
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. */
742
743 static void
744 mark_parm_dereference (tree base, HOST_WIDE_INT dist, gimple stmt)
745 {
746 basic_block bb = gimple_bb (stmt);
747 int idx, parm_index = 0;
748 tree parm;
749
750 if (bitmap_bit_p (final_bbs, bb->index))
751 return;
752
753 for (parm = DECL_ARGUMENTS (current_function_decl);
754 parm && parm != base;
755 parm = DECL_CHAIN (parm))
756 parm_index++;
757
758 gcc_assert (parm_index < func_param_count);
759
760 idx = bb->index * func_param_count + parm_index;
761 if (bb_dereferences[idx] < dist)
762 bb_dereferences[idx] = dist;
763 }
764
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. */
768
769 static struct access *
770 create_access_1 (tree base, HOST_WIDE_INT offset, HOST_WIDE_INT size)
771 {
772 VEC (access_p, heap) *vec;
773 struct access *access;
774 void **slot;
775
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;
781
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);
787
788 VEC_safe_push (access_p, heap, vec, access);
789
790 *((struct VEC (access_p,heap) **)
791 pointer_map_insert (base_access_vec, base)) = vec;
792
793 return access;
794 }
795
796 /* Create and insert access for EXPR. Return created access, or NULL if it is
797 not possible. */
798
799 static struct access *
800 create_access (tree expr, gimple stmt, bool write)
801 {
802 struct access *access;
803 HOST_WIDE_INT offset, size, max_size;
804 tree base = expr;
805 bool ptr, unscalarizable_region = false;
806
807 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
808
809 if (sra_mode == SRA_MODE_EARLY_IPA
810 && TREE_CODE (base) == MEM_REF)
811 {
812 base = get_ssa_base_param (TREE_OPERAND (base, 0));
813 if (!base)
814 return NULL;
815 ptr = true;
816 }
817 else
818 ptr = false;
819
820 if (!DECL_P (base) || !bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
821 return NULL;
822
823 if (sra_mode == SRA_MODE_EARLY_IPA)
824 {
825 if (size < 0 || size != max_size)
826 {
827 disqualify_candidate (base, "Encountered a variable sized access.");
828 return NULL;
829 }
830 if (TREE_CODE (expr) == COMPONENT_REF
831 && DECL_BIT_FIELD (TREE_OPERAND (expr, 1)))
832 {
833 disqualify_candidate (base, "Encountered a bit-field access.");
834 return NULL;
835 }
836 gcc_checking_assert ((offset % BITS_PER_UNIT) == 0);
837
838 if (ptr)
839 mark_parm_dereference (base, offset + size, stmt);
840 }
841 else
842 {
843 if (size != max_size)
844 {
845 size = max_size;
846 unscalarizable_region = true;
847 }
848 if (size < 0)
849 {
850 disqualify_candidate (base, "Encountered an unconstrained access.");
851 return NULL;
852 }
853 }
854
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;
861
862 if (TREE_CODE (expr) == COMPONENT_REF
863 && DECL_NONADDRESSABLE_P (TREE_OPERAND (expr, 1)))
864 access->non_addressable = 1;
865
866 return access;
867 }
868
869
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. */
873
874 static bool
875 type_consists_of_records_p (tree type)
876 {
877 tree fld;
878
879 if (TREE_CODE (type) != RECORD_TYPE)
880 return false;
881
882 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
883 if (TREE_CODE (fld) == FIELD_DECL)
884 {
885 tree ft = TREE_TYPE (fld);
886
887 if (DECL_BIT_FIELD (fld))
888 return false;
889
890 if (!is_gimple_reg_type (ft)
891 && !type_consists_of_records_p (ft))
892 return false;
893 }
894
895 return true;
896 }
897
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. */
903
904 static void
905 completely_scalarize_record (tree base, tree decl, HOST_WIDE_INT offset,
906 tree ref)
907 {
908 tree fld, decl_type = TREE_TYPE (decl);
909
910 for (fld = TYPE_FIELDS (decl_type); fld; fld = DECL_CHAIN (fld))
911 if (TREE_CODE (fld) == FIELD_DECL)
912 {
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);
917
918 if (is_gimple_reg_type (ft))
919 {
920 struct access *access;
921 HOST_WIDE_INT size;
922
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->total_scalarization = 1;
928 /* Accesses for intraprocedural SRA can have their stmt NULL. */
929 }
930 else
931 completely_scalarize_record (base, fld, pos, nref);
932 }
933 }
934
935
936 /* Search the given tree for a declaration by skipping handled components and
937 exclude it from the candidates. */
938
939 static void
940 disqualify_base_of_expr (tree t, const char *reason)
941 {
942 t = get_base_address (t);
943 if (sra_mode == SRA_MODE_EARLY_IPA
944 && TREE_CODE (t) == MEM_REF)
945 t = get_ssa_base_param (TREE_OPERAND (t, 0));
946
947 if (t && DECL_P (t))
948 disqualify_candidate (t, reason);
949 }
950
951 /* Scan expression EXPR and create access structures for all accesses to
952 candidates for scalarization. Return the created access or NULL if none is
953 created. */
954
955 static struct access *
956 build_access_from_expr_1 (tree expr, gimple stmt, bool write)
957 {
958 struct access *ret = NULL;
959 bool partial_ref;
960
961 if (TREE_CODE (expr) == BIT_FIELD_REF
962 || TREE_CODE (expr) == IMAGPART_EXPR
963 || TREE_CODE (expr) == REALPART_EXPR)
964 {
965 expr = TREE_OPERAND (expr, 0);
966 partial_ref = true;
967 }
968 else
969 partial_ref = false;
970
971 /* We need to dive through V_C_Es in order to get the size of its parameter
972 and not the result type. Ada produces such statements. We are also
973 capable of handling the topmost V_C_E but not any of those buried in other
974 handled components. */
975 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
976 expr = TREE_OPERAND (expr, 0);
977
978 if (contains_view_convert_expr_p (expr))
979 {
980 disqualify_base_of_expr (expr, "V_C_E under a different handled "
981 "component.");
982 return NULL;
983 }
984
985 switch (TREE_CODE (expr))
986 {
987 case MEM_REF:
988 if (TREE_CODE (TREE_OPERAND (expr, 0)) != ADDR_EXPR
989 && sra_mode != SRA_MODE_EARLY_IPA)
990 return NULL;
991 /* fall through */
992 case VAR_DECL:
993 case PARM_DECL:
994 case RESULT_DECL:
995 case COMPONENT_REF:
996 case ARRAY_REF:
997 case ARRAY_RANGE_REF:
998 ret = create_access (expr, stmt, write);
999 break;
1000
1001 default:
1002 break;
1003 }
1004
1005 if (write && partial_ref && ret)
1006 ret->grp_partial_lhs = 1;
1007
1008 return ret;
1009 }
1010
1011 /* Scan expression EXPR and create access structures for all accesses to
1012 candidates for scalarization. Return true if any access has been inserted.
1013 STMT must be the statement from which the expression is taken, WRITE must be
1014 true if the expression is a store and false otherwise. */
1015
1016 static bool
1017 build_access_from_expr (tree expr, gimple stmt, bool write)
1018 {
1019 struct access *access;
1020
1021 access = build_access_from_expr_1 (expr, stmt, write);
1022 if (access)
1023 {
1024 /* This means the aggregate is accesses as a whole in a way other than an
1025 assign statement and thus cannot be removed even if we had a scalar
1026 replacement for everything. */
1027 if (cannot_scalarize_away_bitmap)
1028 bitmap_set_bit (cannot_scalarize_away_bitmap, DECL_UID (access->base));
1029 return true;
1030 }
1031 return false;
1032 }
1033
1034 /* Disqualify LHS and RHS for scalarization if STMT must end its basic block in
1035 modes in which it matters, return true iff they have been disqualified. RHS
1036 may be NULL, in that case ignore it. If we scalarize an aggregate in
1037 intra-SRA we may need to add statements after each statement. This is not
1038 possible if a statement unconditionally has to end the basic block. */
1039 static bool
1040 disqualify_ops_if_throwing_stmt (gimple stmt, tree lhs, tree rhs)
1041 {
1042 if ((sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1043 && (stmt_can_throw_internal (stmt) || stmt_ends_bb_p (stmt)))
1044 {
1045 disqualify_base_of_expr (lhs, "LHS of a throwing stmt.");
1046 if (rhs)
1047 disqualify_base_of_expr (rhs, "RHS of a throwing stmt.");
1048 return true;
1049 }
1050 return false;
1051 }
1052
1053 /* Return true iff type of EXP is not sufficiently aligned. */
1054
1055 static bool
1056 tree_non_mode_aligned_mem_p (tree exp)
1057 {
1058 enum machine_mode mode = TYPE_MODE (TREE_TYPE (exp));
1059 unsigned int align;
1060
1061 if (TREE_CODE (exp) == SSA_NAME
1062 || mode == BLKmode
1063 || !STRICT_ALIGNMENT)
1064 return false;
1065
1066 align = get_object_alignment (exp, BIGGEST_ALIGNMENT);
1067 if (GET_MODE_ALIGNMENT (mode) > align)
1068 return true;
1069
1070 return false;
1071 }
1072
1073 /* Scan expressions occuring in STMT, create access structures for all accesses
1074 to candidates for scalarization and remove those candidates which occur in
1075 statements or expressions that prevent them from being split apart. Return
1076 true if any access has been inserted. */
1077
1078 static bool
1079 build_accesses_from_assign (gimple stmt)
1080 {
1081 tree lhs, rhs;
1082 struct access *lacc, *racc;
1083
1084 if (!gimple_assign_single_p (stmt))
1085 return false;
1086
1087 lhs = gimple_assign_lhs (stmt);
1088 rhs = gimple_assign_rhs1 (stmt);
1089
1090 if (disqualify_ops_if_throwing_stmt (stmt, lhs, rhs))
1091 return false;
1092
1093 racc = build_access_from_expr_1 (rhs, stmt, false);
1094 lacc = build_access_from_expr_1 (lhs, stmt, true);
1095
1096 if (lacc)
1097 {
1098 lacc->grp_assignment_write = 1;
1099 lacc->grp_unscalarizable_region |= tree_non_mode_aligned_mem_p (rhs);
1100 }
1101
1102 if (racc)
1103 {
1104 racc->grp_assignment_read = 1;
1105 if (should_scalarize_away_bitmap && !gimple_has_volatile_ops (stmt)
1106 && !is_gimple_reg_type (racc->type))
1107 bitmap_set_bit (should_scalarize_away_bitmap, DECL_UID (racc->base));
1108 racc->grp_unscalarizable_region |= tree_non_mode_aligned_mem_p (lhs);
1109 }
1110
1111 if (lacc && racc
1112 && (sra_mode == SRA_MODE_EARLY_INTRA || sra_mode == SRA_MODE_INTRA)
1113 && !lacc->grp_unscalarizable_region
1114 && !racc->grp_unscalarizable_region
1115 && AGGREGATE_TYPE_P (TREE_TYPE (lhs))
1116 /* FIXME: Turn the following line into an assert after PR 40058 is
1117 fixed. */
1118 && lacc->size == racc->size
1119 && useless_type_conversion_p (lacc->type, racc->type))
1120 {
1121 struct assign_link *link;
1122
1123 link = (struct assign_link *) pool_alloc (link_pool);
1124 memset (link, 0, sizeof (struct assign_link));
1125
1126 link->lacc = lacc;
1127 link->racc = racc;
1128
1129 add_link_to_rhs (racc, link);
1130 }
1131
1132 return lacc || racc;
1133 }
1134
1135 /* Callback of walk_stmt_load_store_addr_ops visit_addr used to determine
1136 GIMPLE_ASM operands with memory constrains which cannot be scalarized. */
1137
1138 static bool
1139 asm_visit_addr (gimple stmt ATTRIBUTE_UNUSED, tree op,
1140 void *data ATTRIBUTE_UNUSED)
1141 {
1142 op = get_base_address (op);
1143 if (op
1144 && DECL_P (op))
1145 disqualify_candidate (op, "Non-scalarizable GIMPLE_ASM operand.");
1146
1147 return false;
1148 }
1149
1150 /* Return true iff callsite CALL has at least as many actual arguments as there
1151 are formal parameters of the function currently processed by IPA-SRA. */
1152
1153 static inline bool
1154 callsite_has_enough_arguments_p (gimple call)
1155 {
1156 return gimple_call_num_args (call) >= (unsigned) func_param_count;
1157 }
1158
1159 /* Scan function and look for interesting expressions and create access
1160 structures for them. Return true iff any access is created. */
1161
1162 static bool
1163 scan_function (void)
1164 {
1165 basic_block bb;
1166 bool ret = false;
1167
1168 FOR_EACH_BB (bb)
1169 {
1170 gimple_stmt_iterator gsi;
1171 for (gsi = gsi_start_bb (bb); !gsi_end_p (gsi); gsi_next (&gsi))
1172 {
1173 gimple stmt = gsi_stmt (gsi);
1174 tree t;
1175 unsigned i;
1176
1177 if (final_bbs && stmt_can_throw_external (stmt))
1178 bitmap_set_bit (final_bbs, bb->index);
1179 switch (gimple_code (stmt))
1180 {
1181 case GIMPLE_RETURN:
1182 t = gimple_return_retval (stmt);
1183 if (t != NULL_TREE)
1184 ret |= build_access_from_expr (t, stmt, false);
1185 if (final_bbs)
1186 bitmap_set_bit (final_bbs, bb->index);
1187 break;
1188
1189 case GIMPLE_ASSIGN:
1190 ret |= build_accesses_from_assign (stmt);
1191 break;
1192
1193 case GIMPLE_CALL:
1194 for (i = 0; i < gimple_call_num_args (stmt); i++)
1195 ret |= build_access_from_expr (gimple_call_arg (stmt, i),
1196 stmt, false);
1197
1198 if (sra_mode == SRA_MODE_EARLY_IPA)
1199 {
1200 tree dest = gimple_call_fndecl (stmt);
1201 int flags = gimple_call_flags (stmt);
1202
1203 if (dest)
1204 {
1205 if (DECL_BUILT_IN_CLASS (dest) == BUILT_IN_NORMAL
1206 && DECL_FUNCTION_CODE (dest) == BUILT_IN_APPLY_ARGS)
1207 encountered_apply_args = true;
1208 if (cgraph_get_node (dest)
1209 == cgraph_get_node (current_function_decl))
1210 {
1211 encountered_recursive_call = true;
1212 if (!callsite_has_enough_arguments_p (stmt))
1213 encountered_unchangable_recursive_call = true;
1214 }
1215 }
1216
1217 if (final_bbs
1218 && (flags & (ECF_CONST | ECF_PURE)) == 0)
1219 bitmap_set_bit (final_bbs, bb->index);
1220 }
1221
1222 t = gimple_call_lhs (stmt);
1223 if (t && !disqualify_ops_if_throwing_stmt (stmt, t, NULL))
1224 ret |= build_access_from_expr (t, stmt, true);
1225 break;
1226
1227 case GIMPLE_ASM:
1228 walk_stmt_load_store_addr_ops (stmt, NULL, NULL, NULL,
1229 asm_visit_addr);
1230 if (final_bbs)
1231 bitmap_set_bit (final_bbs, bb->index);
1232
1233 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
1234 {
1235 t = TREE_VALUE (gimple_asm_input_op (stmt, i));
1236 ret |= build_access_from_expr (t, stmt, false);
1237 }
1238 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
1239 {
1240 t = TREE_VALUE (gimple_asm_output_op (stmt, i));
1241 ret |= build_access_from_expr (t, stmt, true);
1242 }
1243 break;
1244
1245 default:
1246 break;
1247 }
1248 }
1249 }
1250
1251 return ret;
1252 }
1253
1254 /* Helper of QSORT function. There are pointers to accesses in the array. An
1255 access is considered smaller than another if it has smaller offset or if the
1256 offsets are the same but is size is bigger. */
1257
1258 static int
1259 compare_access_positions (const void *a, const void *b)
1260 {
1261 const access_p *fp1 = (const access_p *) a;
1262 const access_p *fp2 = (const access_p *) b;
1263 const access_p f1 = *fp1;
1264 const access_p f2 = *fp2;
1265
1266 if (f1->offset != f2->offset)
1267 return f1->offset < f2->offset ? -1 : 1;
1268
1269 if (f1->size == f2->size)
1270 {
1271 if (f1->type == f2->type)
1272 return 0;
1273 /* Put any non-aggregate type before any aggregate type. */
1274 else if (!is_gimple_reg_type (f1->type)
1275 && is_gimple_reg_type (f2->type))
1276 return 1;
1277 else if (is_gimple_reg_type (f1->type)
1278 && !is_gimple_reg_type (f2->type))
1279 return -1;
1280 /* Put any complex or vector type before any other scalar type. */
1281 else if (TREE_CODE (f1->type) != COMPLEX_TYPE
1282 && TREE_CODE (f1->type) != VECTOR_TYPE
1283 && (TREE_CODE (f2->type) == COMPLEX_TYPE
1284 || TREE_CODE (f2->type) == VECTOR_TYPE))
1285 return 1;
1286 else if ((TREE_CODE (f1->type) == COMPLEX_TYPE
1287 || TREE_CODE (f1->type) == VECTOR_TYPE)
1288 && TREE_CODE (f2->type) != COMPLEX_TYPE
1289 && TREE_CODE (f2->type) != VECTOR_TYPE)
1290 return -1;
1291 /* Put the integral type with the bigger precision first. */
1292 else if (INTEGRAL_TYPE_P (f1->type)
1293 && INTEGRAL_TYPE_P (f2->type))
1294 return TYPE_PRECISION (f2->type) - TYPE_PRECISION (f1->type);
1295 /* Put any integral type with non-full precision last. */
1296 else if (INTEGRAL_TYPE_P (f1->type)
1297 && (TREE_INT_CST_LOW (TYPE_SIZE (f1->type))
1298 != TYPE_PRECISION (f1->type)))
1299 return 1;
1300 else if (INTEGRAL_TYPE_P (f2->type)
1301 && (TREE_INT_CST_LOW (TYPE_SIZE (f2->type))
1302 != TYPE_PRECISION (f2->type)))
1303 return -1;
1304 /* Stabilize the sort. */
1305 return TYPE_UID (f1->type) - TYPE_UID (f2->type);
1306 }
1307
1308 /* We want the bigger accesses first, thus the opposite operator in the next
1309 line: */
1310 return f1->size > f2->size ? -1 : 1;
1311 }
1312
1313
1314 /* Append a name of the declaration to the name obstack. A helper function for
1315 make_fancy_name. */
1316
1317 static void
1318 make_fancy_decl_name (tree decl)
1319 {
1320 char buffer[32];
1321
1322 tree name = DECL_NAME (decl);
1323 if (name)
1324 obstack_grow (&name_obstack, IDENTIFIER_POINTER (name),
1325 IDENTIFIER_LENGTH (name));
1326 else
1327 {
1328 sprintf (buffer, "D%u", DECL_UID (decl));
1329 obstack_grow (&name_obstack, buffer, strlen (buffer));
1330 }
1331 }
1332
1333 /* Helper for make_fancy_name. */
1334
1335 static void
1336 make_fancy_name_1 (tree expr)
1337 {
1338 char buffer[32];
1339 tree index;
1340
1341 if (DECL_P (expr))
1342 {
1343 make_fancy_decl_name (expr);
1344 return;
1345 }
1346
1347 switch (TREE_CODE (expr))
1348 {
1349 case COMPONENT_REF:
1350 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1351 obstack_1grow (&name_obstack, '$');
1352 make_fancy_decl_name (TREE_OPERAND (expr, 1));
1353 break;
1354
1355 case ARRAY_REF:
1356 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1357 obstack_1grow (&name_obstack, '$');
1358 /* Arrays with only one element may not have a constant as their
1359 index. */
1360 index = TREE_OPERAND (expr, 1);
1361 if (TREE_CODE (index) != INTEGER_CST)
1362 break;
1363 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (index));
1364 obstack_grow (&name_obstack, buffer, strlen (buffer));
1365 break;
1366
1367 case ADDR_EXPR:
1368 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1369 break;
1370
1371 case MEM_REF:
1372 make_fancy_name_1 (TREE_OPERAND (expr, 0));
1373 if (!integer_zerop (TREE_OPERAND (expr, 1)))
1374 {
1375 obstack_1grow (&name_obstack, '$');
1376 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC,
1377 TREE_INT_CST_LOW (TREE_OPERAND (expr, 1)));
1378 obstack_grow (&name_obstack, buffer, strlen (buffer));
1379 }
1380 break;
1381
1382 case BIT_FIELD_REF:
1383 case REALPART_EXPR:
1384 case IMAGPART_EXPR:
1385 gcc_unreachable (); /* we treat these as scalars. */
1386 break;
1387 default:
1388 break;
1389 }
1390 }
1391
1392 /* Create a human readable name for replacement variable of ACCESS. */
1393
1394 static char *
1395 make_fancy_name (tree expr)
1396 {
1397 make_fancy_name_1 (expr);
1398 obstack_1grow (&name_obstack, '\0');
1399 return XOBFINISH (&name_obstack, char *);
1400 }
1401
1402 /* Construct a MEM_REF that would reference a part of aggregate BASE of type
1403 EXP_TYPE at the given OFFSET. If BASE is something for which
1404 get_addr_base_and_unit_offset returns NULL, gsi must be non-NULL and is used
1405 to insert new statements either before or below the current one as specified
1406 by INSERT_AFTER. This function is not capable of handling bitfields. */
1407
1408 tree
1409 build_ref_for_offset (location_t loc, tree base, HOST_WIDE_INT offset,
1410 tree exp_type, gimple_stmt_iterator *gsi,
1411 bool insert_after)
1412 {
1413 tree prev_base = base;
1414 tree off;
1415 HOST_WIDE_INT base_offset;
1416
1417 gcc_checking_assert (offset % BITS_PER_UNIT == 0);
1418
1419 base = get_addr_base_and_unit_offset (base, &base_offset);
1420
1421 /* get_addr_base_and_unit_offset returns NULL for references with a variable
1422 offset such as array[var_index]. */
1423 if (!base)
1424 {
1425 gimple stmt;
1426 tree tmp, addr;
1427
1428 gcc_checking_assert (gsi);
1429 tmp = create_tmp_reg (build_pointer_type (TREE_TYPE (prev_base)), NULL);
1430 add_referenced_var (tmp);
1431 tmp = make_ssa_name (tmp, NULL);
1432 addr = build_fold_addr_expr (unshare_expr (prev_base));
1433 STRIP_USELESS_TYPE_CONVERSION (addr);
1434 stmt = gimple_build_assign (tmp, addr);
1435 gimple_set_location (stmt, loc);
1436 SSA_NAME_DEF_STMT (tmp) = stmt;
1437 if (insert_after)
1438 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
1439 else
1440 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
1441 update_stmt (stmt);
1442
1443 off = build_int_cst (reference_alias_ptr_type (prev_base),
1444 offset / BITS_PER_UNIT);
1445 base = tmp;
1446 }
1447 else if (TREE_CODE (base) == MEM_REF)
1448 {
1449 off = build_int_cst (TREE_TYPE (TREE_OPERAND (base, 1)),
1450 base_offset + offset / BITS_PER_UNIT);
1451 off = int_const_binop (PLUS_EXPR, TREE_OPERAND (base, 1), off);
1452 base = unshare_expr (TREE_OPERAND (base, 0));
1453 }
1454 else
1455 {
1456 off = build_int_cst (reference_alias_ptr_type (base),
1457 base_offset + offset / BITS_PER_UNIT);
1458 base = build_fold_addr_expr (unshare_expr (base));
1459 }
1460
1461 return fold_build2_loc (loc, MEM_REF, exp_type, base, off);
1462 }
1463
1464 /* Construct a memory reference to a part of an aggregate BASE at the given
1465 OFFSET and of the same type as MODEL. In case this is a reference to a
1466 component, the function will replicate the last COMPONENT_REF of model's
1467 expr to access it. GSI and INSERT_AFTER have the same meaning as in
1468 build_ref_for_offset. */
1469
1470 static tree
1471 build_ref_for_model (location_t loc, tree base, HOST_WIDE_INT offset,
1472 struct access *model, gimple_stmt_iterator *gsi,
1473 bool insert_after)
1474 {
1475 if (TREE_CODE (model->expr) == COMPONENT_REF)
1476 {
1477 tree t, exp_type, fld = TREE_OPERAND (model->expr, 1);
1478 tree cr_offset = component_ref_field_offset (model->expr);
1479
1480 gcc_assert (cr_offset && host_integerp (cr_offset, 1));
1481 offset -= TREE_INT_CST_LOW (cr_offset) * BITS_PER_UNIT;
1482 offset -= TREE_INT_CST_LOW (DECL_FIELD_BIT_OFFSET (fld));
1483 exp_type = TREE_TYPE (TREE_OPERAND (model->expr, 0));
1484 t = build_ref_for_offset (loc, base, offset, exp_type, gsi, insert_after);
1485 return fold_build3_loc (loc, COMPONENT_REF, model->type, t, fld,
1486 TREE_OPERAND (model->expr, 2));
1487 }
1488 else
1489 return build_ref_for_offset (loc, base, offset, model->type,
1490 gsi, insert_after);
1491 }
1492
1493 /* Construct a memory reference consisting of component_refs and array_refs to
1494 a part of an aggregate *RES (which is of type TYPE). The requested part
1495 should have type EXP_TYPE at be the given OFFSET. This function might not
1496 succeed, it returns true when it does and only then *RES points to something
1497 meaningful. This function should be used only to build expressions that we
1498 might need to present to user (e.g. in warnings). In all other situations,
1499 build_ref_for_model or build_ref_for_offset should be used instead. */
1500
1501 static bool
1502 build_user_friendly_ref_for_offset (tree *res, tree type, HOST_WIDE_INT offset,
1503 tree exp_type)
1504 {
1505 while (1)
1506 {
1507 tree fld;
1508 tree tr_size, index, minidx;
1509 HOST_WIDE_INT el_size;
1510
1511 if (offset == 0 && exp_type
1512 && types_compatible_p (exp_type, type))
1513 return true;
1514
1515 switch (TREE_CODE (type))
1516 {
1517 case UNION_TYPE:
1518 case QUAL_UNION_TYPE:
1519 case RECORD_TYPE:
1520 for (fld = TYPE_FIELDS (type); fld; fld = DECL_CHAIN (fld))
1521 {
1522 HOST_WIDE_INT pos, size;
1523 tree expr, *expr_ptr;
1524
1525 if (TREE_CODE (fld) != FIELD_DECL)
1526 continue;
1527
1528 pos = int_bit_position (fld);
1529 gcc_assert (TREE_CODE (type) == RECORD_TYPE || pos == 0);
1530 tr_size = DECL_SIZE (fld);
1531 if (!tr_size || !host_integerp (tr_size, 1))
1532 continue;
1533 size = tree_low_cst (tr_size, 1);
1534 if (size == 0)
1535 {
1536 if (pos != offset)
1537 continue;
1538 }
1539 else if (pos > offset || (pos + size) <= offset)
1540 continue;
1541
1542 expr = build3 (COMPONENT_REF, TREE_TYPE (fld), *res, fld,
1543 NULL_TREE);
1544 expr_ptr = &expr;
1545 if (build_user_friendly_ref_for_offset (expr_ptr, TREE_TYPE (fld),
1546 offset - pos, exp_type))
1547 {
1548 *res = expr;
1549 return true;
1550 }
1551 }
1552 return false;
1553
1554 case ARRAY_TYPE:
1555 tr_size = TYPE_SIZE (TREE_TYPE (type));
1556 if (!tr_size || !host_integerp (tr_size, 1))
1557 return false;
1558 el_size = tree_low_cst (tr_size, 1);
1559
1560 minidx = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1561 if (TREE_CODE (minidx) != INTEGER_CST || el_size == 0)
1562 return false;
1563 index = build_int_cst (TYPE_DOMAIN (type), offset / el_size);
1564 if (!integer_zerop (minidx))
1565 index = int_const_binop (PLUS_EXPR, index, minidx);
1566 *res = build4 (ARRAY_REF, TREE_TYPE (type), *res, index,
1567 NULL_TREE, NULL_TREE);
1568 offset = offset % el_size;
1569 type = TREE_TYPE (type);
1570 break;
1571
1572 default:
1573 if (offset != 0)
1574 return false;
1575
1576 if (exp_type)
1577 return false;
1578 else
1579 return true;
1580 }
1581 }
1582 }
1583
1584 /* Return true iff TYPE is stdarg va_list type. */
1585
1586 static inline bool
1587 is_va_list_type (tree type)
1588 {
1589 return TYPE_MAIN_VARIANT (type) == TYPE_MAIN_VARIANT (va_list_type_node);
1590 }
1591
1592 /* Print message to dump file why a variable was rejected. */
1593
1594 static void
1595 reject (tree var, const char *msg)
1596 {
1597 if (dump_file && (dump_flags & TDF_DETAILS))
1598 {
1599 fprintf (dump_file, "Rejected (%d): %s: ", DECL_UID (var), msg);
1600 print_generic_expr (dump_file, var, 0);
1601 fprintf (dump_file, "\n");
1602 }
1603 }
1604
1605 /* The very first phase of intraprocedural SRA. It marks in candidate_bitmap
1606 those with type which is suitable for scalarization. */
1607
1608 static bool
1609 find_var_candidates (void)
1610 {
1611 tree var, type;
1612 referenced_var_iterator rvi;
1613 bool ret = false;
1614 const char *msg;
1615
1616 FOR_EACH_REFERENCED_VAR (cfun, var, rvi)
1617 {
1618 if (TREE_CODE (var) != VAR_DECL && TREE_CODE (var) != PARM_DECL)
1619 continue;
1620 type = TREE_TYPE (var);
1621
1622 if (!AGGREGATE_TYPE_P (type))
1623 {
1624 reject (var, "not aggregate");
1625 continue;
1626 }
1627 if (needs_to_live_in_memory (var))
1628 {
1629 reject (var, "needs to live in memory");
1630 continue;
1631 }
1632 if (TREE_THIS_VOLATILE (var))
1633 {
1634 reject (var, "is volatile");
1635 continue;
1636 }
1637 if (!COMPLETE_TYPE_P (type))
1638 {
1639 reject (var, "has incomplete type");
1640 continue;
1641 }
1642 if (!host_integerp (TYPE_SIZE (type), 1))
1643 {
1644 reject (var, "type size not fixed");
1645 continue;
1646 }
1647 if (tree_low_cst (TYPE_SIZE (type), 1) == 0)
1648 {
1649 reject (var, "type size is zero");
1650 continue;
1651 }
1652 if (type_internals_preclude_sra_p (type, &msg))
1653 {
1654 reject (var, msg);
1655 continue;
1656 }
1657 if (/* Fix for PR 41089. tree-stdarg.c needs to have va_lists intact but
1658 we also want to schedule it rather late. Thus we ignore it in
1659 the early pass. */
1660 (sra_mode == SRA_MODE_EARLY_INTRA
1661 && is_va_list_type (type)))
1662 {
1663 reject (var, "is va_list");
1664 continue;
1665 }
1666
1667 bitmap_set_bit (candidate_bitmap, DECL_UID (var));
1668
1669 if (dump_file && (dump_flags & TDF_DETAILS))
1670 {
1671 fprintf (dump_file, "Candidate (%d): ", DECL_UID (var));
1672 print_generic_expr (dump_file, var, 0);
1673 fprintf (dump_file, "\n");
1674 }
1675 ret = true;
1676 }
1677
1678 return ret;
1679 }
1680
1681 /* Sort all accesses for the given variable, check for partial overlaps and
1682 return NULL if there are any. If there are none, pick a representative for
1683 each combination of offset and size and create a linked list out of them.
1684 Return the pointer to the first representative and make sure it is the first
1685 one in the vector of accesses. */
1686
1687 static struct access *
1688 sort_and_splice_var_accesses (tree var)
1689 {
1690 int i, j, access_count;
1691 struct access *res, **prev_acc_ptr = &res;
1692 VEC (access_p, heap) *access_vec;
1693 bool first = true;
1694 HOST_WIDE_INT low = -1, high = 0;
1695
1696 access_vec = get_base_access_vector (var);
1697 if (!access_vec)
1698 return NULL;
1699 access_count = VEC_length (access_p, access_vec);
1700
1701 /* Sort by <OFFSET, SIZE>. */
1702 VEC_qsort (access_p, access_vec, compare_access_positions);
1703
1704 i = 0;
1705 while (i < access_count)
1706 {
1707 struct access *access = VEC_index (access_p, access_vec, i);
1708 bool grp_write = access->write;
1709 bool grp_read = !access->write;
1710 bool grp_scalar_write = access->write
1711 && is_gimple_reg_type (access->type);
1712 bool grp_scalar_read = !access->write
1713 && is_gimple_reg_type (access->type);
1714 bool grp_assignment_read = access->grp_assignment_read;
1715 bool grp_assignment_write = access->grp_assignment_write;
1716 bool multiple_scalar_reads = false;
1717 bool total_scalarization = access->total_scalarization;
1718 bool grp_partial_lhs = access->grp_partial_lhs;
1719 bool first_scalar = is_gimple_reg_type (access->type);
1720 bool unscalarizable_region = access->grp_unscalarizable_region;
1721
1722 if (first || access->offset >= high)
1723 {
1724 first = false;
1725 low = access->offset;
1726 high = access->offset + access->size;
1727 }
1728 else if (access->offset > low && access->offset + access->size > high)
1729 return NULL;
1730 else
1731 gcc_assert (access->offset >= low
1732 && access->offset + access->size <= high);
1733
1734 j = i + 1;
1735 while (j < access_count)
1736 {
1737 struct access *ac2 = VEC_index (access_p, access_vec, j);
1738 if (ac2->offset != access->offset || ac2->size != access->size)
1739 break;
1740 if (ac2->write)
1741 {
1742 grp_write = true;
1743 grp_scalar_write = (grp_scalar_write
1744 || is_gimple_reg_type (ac2->type));
1745 }
1746 else
1747 {
1748 grp_read = true;
1749 if (is_gimple_reg_type (ac2->type))
1750 {
1751 if (grp_scalar_read)
1752 multiple_scalar_reads = true;
1753 else
1754 grp_scalar_read = true;
1755 }
1756 }
1757 grp_assignment_read |= ac2->grp_assignment_read;
1758 grp_assignment_write |= ac2->grp_assignment_write;
1759 grp_partial_lhs |= ac2->grp_partial_lhs;
1760 unscalarizable_region |= ac2->grp_unscalarizable_region;
1761 total_scalarization |= ac2->total_scalarization;
1762 relink_to_new_repr (access, ac2);
1763
1764 /* If there are both aggregate-type and scalar-type accesses with
1765 this combination of size and offset, the comparison function
1766 should have put the scalars first. */
1767 gcc_assert (first_scalar || !is_gimple_reg_type (ac2->type));
1768 ac2->group_representative = access;
1769 j++;
1770 }
1771
1772 i = j;
1773
1774 access->group_representative = access;
1775 access->grp_write = grp_write;
1776 access->grp_read = grp_read;
1777 access->grp_scalar_read = grp_scalar_read;
1778 access->grp_scalar_write = grp_scalar_write;
1779 access->grp_assignment_read = grp_assignment_read;
1780 access->grp_assignment_write = grp_assignment_write;
1781 access->grp_hint = multiple_scalar_reads || total_scalarization;
1782 access->grp_partial_lhs = grp_partial_lhs;
1783 access->grp_unscalarizable_region = unscalarizable_region;
1784 if (access->first_link)
1785 add_access_to_work_queue (access);
1786
1787 *prev_acc_ptr = access;
1788 prev_acc_ptr = &access->next_grp;
1789 }
1790
1791 gcc_assert (res == VEC_index (access_p, access_vec, 0));
1792 return res;
1793 }
1794
1795 /* Create a variable for the given ACCESS which determines the type, name and a
1796 few other properties. Return the variable declaration and store it also to
1797 ACCESS->replacement. */
1798
1799 static tree
1800 create_access_replacement (struct access *access, bool rename)
1801 {
1802 tree repl;
1803
1804 repl = create_tmp_var (access->type, "SR");
1805 get_var_ann (repl);
1806 add_referenced_var (repl);
1807 if (rename)
1808 mark_sym_for_renaming (repl);
1809
1810 if (!access->grp_partial_lhs
1811 && (TREE_CODE (access->type) == COMPLEX_TYPE
1812 || TREE_CODE (access->type) == VECTOR_TYPE))
1813 DECL_GIMPLE_REG_P (repl) = 1;
1814
1815 DECL_SOURCE_LOCATION (repl) = DECL_SOURCE_LOCATION (access->base);
1816 DECL_ARTIFICIAL (repl) = 1;
1817 DECL_IGNORED_P (repl) = DECL_IGNORED_P (access->base);
1818
1819 if (DECL_NAME (access->base)
1820 && !DECL_IGNORED_P (access->base)
1821 && !DECL_ARTIFICIAL (access->base))
1822 {
1823 char *pretty_name = make_fancy_name (access->expr);
1824 tree debug_expr = unshare_expr (access->expr), d;
1825
1826 DECL_NAME (repl) = get_identifier (pretty_name);
1827 obstack_free (&name_obstack, pretty_name);
1828
1829 /* Get rid of any SSA_NAMEs embedded in debug_expr,
1830 as DECL_DEBUG_EXPR isn't considered when looking for still
1831 used SSA_NAMEs and thus they could be freed. All debug info
1832 generation cares is whether something is constant or variable
1833 and that get_ref_base_and_extent works properly on the
1834 expression. */
1835 for (d = debug_expr; handled_component_p (d); d = TREE_OPERAND (d, 0))
1836 switch (TREE_CODE (d))
1837 {
1838 case ARRAY_REF:
1839 case ARRAY_RANGE_REF:
1840 if (TREE_OPERAND (d, 1)
1841 && TREE_CODE (TREE_OPERAND (d, 1)) == SSA_NAME)
1842 TREE_OPERAND (d, 1) = SSA_NAME_VAR (TREE_OPERAND (d, 1));
1843 if (TREE_OPERAND (d, 3)
1844 && TREE_CODE (TREE_OPERAND (d, 3)) == SSA_NAME)
1845 TREE_OPERAND (d, 3) = SSA_NAME_VAR (TREE_OPERAND (d, 3));
1846 /* FALLTHRU */
1847 case COMPONENT_REF:
1848 if (TREE_OPERAND (d, 2)
1849 && TREE_CODE (TREE_OPERAND (d, 2)) == SSA_NAME)
1850 TREE_OPERAND (d, 2) = SSA_NAME_VAR (TREE_OPERAND (d, 2));
1851 break;
1852 default:
1853 break;
1854 }
1855 SET_DECL_DEBUG_EXPR (repl, debug_expr);
1856 DECL_DEBUG_EXPR_IS_FROM (repl) = 1;
1857 if (access->grp_no_warning)
1858 TREE_NO_WARNING (repl) = 1;
1859 else
1860 TREE_NO_WARNING (repl) = TREE_NO_WARNING (access->base);
1861 }
1862 else
1863 TREE_NO_WARNING (repl) = 1;
1864
1865 if (dump_file)
1866 {
1867 fprintf (dump_file, "Created a replacement for ");
1868 print_generic_expr (dump_file, access->base, 0);
1869 fprintf (dump_file, " offset: %u, size: %u: ",
1870 (unsigned) access->offset, (unsigned) access->size);
1871 print_generic_expr (dump_file, repl, 0);
1872 fprintf (dump_file, "\n");
1873 }
1874 sra_stats.replacements++;
1875
1876 return repl;
1877 }
1878
1879 /* Return ACCESS scalar replacement, create it if it does not exist yet. */
1880
1881 static inline tree
1882 get_access_replacement (struct access *access)
1883 {
1884 gcc_assert (access->grp_to_be_replaced);
1885
1886 if (!access->replacement_decl)
1887 access->replacement_decl = create_access_replacement (access, true);
1888 return access->replacement_decl;
1889 }
1890
1891 /* Return ACCESS scalar replacement, create it if it does not exist yet but do
1892 not mark it for renaming. */
1893
1894 static inline tree
1895 get_unrenamed_access_replacement (struct access *access)
1896 {
1897 gcc_assert (!access->grp_to_be_replaced);
1898
1899 if (!access->replacement_decl)
1900 access->replacement_decl = create_access_replacement (access, false);
1901 return access->replacement_decl;
1902 }
1903
1904
1905 /* Build a subtree of accesses rooted in *ACCESS, and move the pointer in the
1906 linked list along the way. Stop when *ACCESS is NULL or the access pointed
1907 to it is not "within" the root. Return false iff some accesses partially
1908 overlap. */
1909
1910 static bool
1911 build_access_subtree (struct access **access)
1912 {
1913 struct access *root = *access, *last_child = NULL;
1914 HOST_WIDE_INT limit = root->offset + root->size;
1915
1916 *access = (*access)->next_grp;
1917 while (*access && (*access)->offset + (*access)->size <= limit)
1918 {
1919 if (!last_child)
1920 root->first_child = *access;
1921 else
1922 last_child->next_sibling = *access;
1923 last_child = *access;
1924
1925 if (!build_access_subtree (access))
1926 return false;
1927 }
1928
1929 if (*access && (*access)->offset < limit)
1930 return false;
1931
1932 return true;
1933 }
1934
1935 /* Build a tree of access representatives, ACCESS is the pointer to the first
1936 one, others are linked in a list by the next_grp field. Return false iff
1937 some accesses partially overlap. */
1938
1939 static bool
1940 build_access_trees (struct access *access)
1941 {
1942 while (access)
1943 {
1944 struct access *root = access;
1945
1946 if (!build_access_subtree (&access))
1947 return false;
1948 root->next_grp = access;
1949 }
1950 return true;
1951 }
1952
1953 /* Return true if expr contains some ARRAY_REFs into a variable bounded
1954 array. */
1955
1956 static bool
1957 expr_with_var_bounded_array_refs_p (tree expr)
1958 {
1959 while (handled_component_p (expr))
1960 {
1961 if (TREE_CODE (expr) == ARRAY_REF
1962 && !host_integerp (array_ref_low_bound (expr), 0))
1963 return true;
1964 expr = TREE_OPERAND (expr, 0);
1965 }
1966 return false;
1967 }
1968
1969 /* Analyze the subtree of accesses rooted in ROOT, scheduling replacements when
1970 both seeming beneficial and when ALLOW_REPLACEMENTS allows it. Also set all
1971 sorts of access flags appropriately along the way, notably always set
1972 grp_read and grp_assign_read according to MARK_READ and grp_write when
1973 MARK_WRITE is true.
1974
1975 Creating a replacement for a scalar access is considered beneficial if its
1976 grp_hint is set (this means we are either attempting total scalarization or
1977 there is more than one direct read access) or according to the following
1978 table:
1979
1980 Access written to through a scalar type (once or more times)
1981 |
1982 | Written to in an assignment statement
1983 | |
1984 | | Access read as scalar _once_
1985 | | |
1986 | | | Read in an assignment statement
1987 | | | |
1988 | | | | Scalarize Comment
1989 -----------------------------------------------------------------------------
1990 0 0 0 0 No access for the scalar
1991 0 0 0 1 No access for the scalar
1992 0 0 1 0 No Single read - won't help
1993 0 0 1 1 No The same case
1994 0 1 0 0 No access for the scalar
1995 0 1 0 1 No access for the scalar
1996 0 1 1 0 Yes s = *g; return s.i;
1997 0 1 1 1 Yes The same case as above
1998 1 0 0 0 No Won't help
1999 1 0 0 1 Yes s.i = 1; *g = s;
2000 1 0 1 0 Yes s.i = 5; g = s.i;
2001 1 0 1 1 Yes The same case as above
2002 1 1 0 0 No Won't help.
2003 1 1 0 1 Yes s.i = 1; *g = s;
2004 1 1 1 0 Yes s = *g; return s.i;
2005 1 1 1 1 Yes Any of the above yeses */
2006
2007 static bool
2008 analyze_access_subtree (struct access *root, struct access *parent,
2009 bool allow_replacements)
2010 {
2011 struct access *child;
2012 HOST_WIDE_INT limit = root->offset + root->size;
2013 HOST_WIDE_INT covered_to = root->offset;
2014 bool scalar = is_gimple_reg_type (root->type);
2015 bool hole = false, sth_created = false;
2016
2017 if (parent)
2018 {
2019 if (parent->grp_read)
2020 root->grp_read = 1;
2021 if (parent->grp_assignment_read)
2022 root->grp_assignment_read = 1;
2023 if (parent->grp_write)
2024 root->grp_write = 1;
2025 if (parent->grp_assignment_write)
2026 root->grp_assignment_write = 1;
2027 }
2028
2029 if (root->grp_unscalarizable_region)
2030 allow_replacements = false;
2031
2032 if (allow_replacements && expr_with_var_bounded_array_refs_p (root->expr))
2033 allow_replacements = false;
2034
2035 for (child = root->first_child; child; child = child->next_sibling)
2036 {
2037 if (!hole && child->offset < covered_to)
2038 hole = true;
2039 else
2040 covered_to += child->size;
2041
2042 sth_created |= analyze_access_subtree (child, root,
2043 allow_replacements && !scalar);
2044
2045 root->grp_unscalarized_data |= child->grp_unscalarized_data;
2046 hole |= !child->grp_covered;
2047 }
2048
2049 if (allow_replacements && scalar && !root->first_child
2050 && (root->grp_hint
2051 || ((root->grp_scalar_read || root->grp_assignment_read)
2052 && (root->grp_scalar_write || root->grp_assignment_write))))
2053 {
2054 if (dump_file && (dump_flags & TDF_DETAILS))
2055 {
2056 fprintf (dump_file, "Marking ");
2057 print_generic_expr (dump_file, root->base, 0);
2058 fprintf (dump_file, " offset: %u, size: %u: ",
2059 (unsigned) root->offset, (unsigned) root->size);
2060 fprintf (dump_file, " to be replaced.\n");
2061 }
2062
2063 root->grp_to_be_replaced = 1;
2064 sth_created = true;
2065 hole = false;
2066 }
2067 else if (covered_to < limit)
2068 hole = true;
2069
2070 if (sth_created && !hole)
2071 {
2072 root->grp_covered = 1;
2073 return true;
2074 }
2075 if (root->grp_write || TREE_CODE (root->base) == PARM_DECL)
2076 root->grp_unscalarized_data = 1; /* not covered and written to */
2077 if (sth_created)
2078 return true;
2079 return false;
2080 }
2081
2082 /* Analyze all access trees linked by next_grp by the means of
2083 analyze_access_subtree. */
2084 static bool
2085 analyze_access_trees (struct access *access)
2086 {
2087 bool ret = false;
2088
2089 while (access)
2090 {
2091 if (analyze_access_subtree (access, NULL, true))
2092 ret = true;
2093 access = access->next_grp;
2094 }
2095
2096 return ret;
2097 }
2098
2099 /* Return true iff a potential new child of LACC at offset OFFSET and with size
2100 SIZE would conflict with an already existing one. If exactly such a child
2101 already exists in LACC, store a pointer to it in EXACT_MATCH. */
2102
2103 static bool
2104 child_would_conflict_in_lacc (struct access *lacc, HOST_WIDE_INT norm_offset,
2105 HOST_WIDE_INT size, struct access **exact_match)
2106 {
2107 struct access *child;
2108
2109 for (child = lacc->first_child; child; child = child->next_sibling)
2110 {
2111 if (child->offset == norm_offset && child->size == size)
2112 {
2113 *exact_match = child;
2114 return true;
2115 }
2116
2117 if (child->offset < norm_offset + size
2118 && child->offset + child->size > norm_offset)
2119 return true;
2120 }
2121
2122 return false;
2123 }
2124
2125 /* Create a new child access of PARENT, with all properties just like MODEL
2126 except for its offset and with its grp_write false and grp_read true.
2127 Return the new access or NULL if it cannot be created. Note that this access
2128 is created long after all splicing and sorting, it's not located in any
2129 access vector and is automatically a representative of its group. */
2130
2131 static struct access *
2132 create_artificial_child_access (struct access *parent, struct access *model,
2133 HOST_WIDE_INT new_offset)
2134 {
2135 struct access *access;
2136 struct access **child;
2137 tree expr = parent->base;
2138
2139 gcc_assert (!model->grp_unscalarizable_region);
2140
2141 access = (struct access *) pool_alloc (access_pool);
2142 memset (access, 0, sizeof (struct access));
2143 if (!build_user_friendly_ref_for_offset (&expr, TREE_TYPE (expr), new_offset,
2144 model->type))
2145 {
2146 access->grp_no_warning = true;
2147 expr = build_ref_for_model (EXPR_LOCATION (parent->base), parent->base,
2148 new_offset, model, NULL, false);
2149 }
2150
2151 access->base = parent->base;
2152 access->expr = expr;
2153 access->offset = new_offset;
2154 access->size = model->size;
2155 access->type = model->type;
2156 access->grp_write = true;
2157 access->grp_read = false;
2158
2159 child = &parent->first_child;
2160 while (*child && (*child)->offset < new_offset)
2161 child = &(*child)->next_sibling;
2162
2163 access->next_sibling = *child;
2164 *child = access;
2165
2166 return access;
2167 }
2168
2169
2170 /* Propagate all subaccesses of RACC across an assignment link to LACC. Return
2171 true if any new subaccess was created. Additionally, if RACC is a scalar
2172 access but LACC is not, change the type of the latter, if possible. */
2173
2174 static bool
2175 propagate_subaccesses_across_link (struct access *lacc, struct access *racc)
2176 {
2177 struct access *rchild;
2178 HOST_WIDE_INT norm_delta = lacc->offset - racc->offset;
2179 bool ret = false;
2180
2181 if (is_gimple_reg_type (lacc->type)
2182 || lacc->grp_unscalarizable_region
2183 || racc->grp_unscalarizable_region)
2184 return false;
2185
2186 if (!lacc->first_child && !racc->first_child
2187 && is_gimple_reg_type (racc->type))
2188 {
2189 tree t = lacc->base;
2190
2191 lacc->type = racc->type;
2192 if (build_user_friendly_ref_for_offset (&t, TREE_TYPE (t), lacc->offset,
2193 racc->type))
2194 lacc->expr = t;
2195 else
2196 {
2197 lacc->expr = build_ref_for_model (EXPR_LOCATION (lacc->base),
2198 lacc->base, lacc->offset,
2199 racc, NULL, false);
2200 lacc->grp_no_warning = true;
2201 }
2202 return false;
2203 }
2204
2205 for (rchild = racc->first_child; rchild; rchild = rchild->next_sibling)
2206 {
2207 struct access *new_acc = NULL;
2208 HOST_WIDE_INT norm_offset = rchild->offset + norm_delta;
2209
2210 if (rchild->grp_unscalarizable_region)
2211 continue;
2212
2213 if (child_would_conflict_in_lacc (lacc, norm_offset, rchild->size,
2214 &new_acc))
2215 {
2216 if (new_acc)
2217 {
2218 rchild->grp_hint = 1;
2219 new_acc->grp_hint |= new_acc->grp_read;
2220 if (rchild->first_child)
2221 ret |= propagate_subaccesses_across_link (new_acc, rchild);
2222 }
2223 continue;
2224 }
2225
2226 rchild->grp_hint = 1;
2227 new_acc = create_artificial_child_access (lacc, rchild, norm_offset);
2228 if (new_acc)
2229 {
2230 ret = true;
2231 if (racc->first_child)
2232 propagate_subaccesses_across_link (new_acc, rchild);
2233 }
2234 }
2235
2236 return ret;
2237 }
2238
2239 /* Propagate all subaccesses across assignment links. */
2240
2241 static void
2242 propagate_all_subaccesses (void)
2243 {
2244 while (work_queue_head)
2245 {
2246 struct access *racc = pop_access_from_work_queue ();
2247 struct assign_link *link;
2248
2249 gcc_assert (racc->first_link);
2250
2251 for (link = racc->first_link; link; link = link->next)
2252 {
2253 struct access *lacc = link->lacc;
2254
2255 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (lacc->base)))
2256 continue;
2257 lacc = lacc->group_representative;
2258 if (propagate_subaccesses_across_link (lacc, racc)
2259 && lacc->first_link)
2260 add_access_to_work_queue (lacc);
2261 }
2262 }
2263 }
2264
2265 /* Go through all accesses collected throughout the (intraprocedural) analysis
2266 stage, exclude overlapping ones, identify representatives and build trees
2267 out of them, making decisions about scalarization on the way. Return true
2268 iff there are any to-be-scalarized variables after this stage. */
2269
2270 static bool
2271 analyze_all_variable_accesses (void)
2272 {
2273 int res = 0;
2274 bitmap tmp = BITMAP_ALLOC (NULL);
2275 bitmap_iterator bi;
2276 unsigned i, max_total_scalarization_size;
2277
2278 max_total_scalarization_size = UNITS_PER_WORD * BITS_PER_UNIT
2279 * MOVE_RATIO (optimize_function_for_speed_p (cfun));
2280
2281 EXECUTE_IF_SET_IN_BITMAP (candidate_bitmap, 0, i, bi)
2282 if (bitmap_bit_p (should_scalarize_away_bitmap, i)
2283 && !bitmap_bit_p (cannot_scalarize_away_bitmap, i))
2284 {
2285 tree var = referenced_var (i);
2286
2287 if (TREE_CODE (var) == VAR_DECL
2288 && ((unsigned) tree_low_cst (TYPE_SIZE (TREE_TYPE (var)), 1)
2289 <= max_total_scalarization_size)
2290 && type_consists_of_records_p (TREE_TYPE (var)))
2291 {
2292 completely_scalarize_record (var, var, 0, var);
2293 if (dump_file && (dump_flags & TDF_DETAILS))
2294 {
2295 fprintf (dump_file, "Will attempt to totally scalarize ");
2296 print_generic_expr (dump_file, var, 0);
2297 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2298 }
2299 }
2300 }
2301
2302 bitmap_copy (tmp, candidate_bitmap);
2303 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2304 {
2305 tree var = referenced_var (i);
2306 struct access *access;
2307
2308 access = sort_and_splice_var_accesses (var);
2309 if (!access || !build_access_trees (access))
2310 disqualify_candidate (var,
2311 "No or inhibitingly overlapping accesses.");
2312 }
2313
2314 propagate_all_subaccesses ();
2315
2316 bitmap_copy (tmp, candidate_bitmap);
2317 EXECUTE_IF_SET_IN_BITMAP (tmp, 0, i, bi)
2318 {
2319 tree var = referenced_var (i);
2320 struct access *access = get_first_repr_for_decl (var);
2321
2322 if (analyze_access_trees (access))
2323 {
2324 res++;
2325 if (dump_file && (dump_flags & TDF_DETAILS))
2326 {
2327 fprintf (dump_file, "\nAccess trees for ");
2328 print_generic_expr (dump_file, var, 0);
2329 fprintf (dump_file, " (UID: %u): \n", DECL_UID (var));
2330 dump_access_tree (dump_file, access);
2331 fprintf (dump_file, "\n");
2332 }
2333 }
2334 else
2335 disqualify_candidate (var, "No scalar replacements to be created.");
2336 }
2337
2338 BITMAP_FREE (tmp);
2339
2340 if (res)
2341 {
2342 statistics_counter_event (cfun, "Scalarized aggregates", res);
2343 return true;
2344 }
2345 else
2346 return false;
2347 }
2348
2349 /* Generate statements copying scalar replacements of accesses within a subtree
2350 into or out of AGG. ACCESS, all its children, siblings and their children
2351 are to be processed. AGG is an aggregate type expression (can be a
2352 declaration but does not have to be, it can for example also be a mem_ref or
2353 a series of handled components). TOP_OFFSET is the offset of the processed
2354 subtree which has to be subtracted from offsets of individual accesses to
2355 get corresponding offsets for AGG. If CHUNK_SIZE is non-null, copy only
2356 replacements in the interval <start_offset, start_offset + chunk_size>,
2357 otherwise copy all. GSI is a statement iterator used to place the new
2358 statements. WRITE should be true when the statements should write from AGG
2359 to the replacement and false if vice versa. if INSERT_AFTER is true, new
2360 statements will be added after the current statement in GSI, they will be
2361 added before the statement otherwise. */
2362
2363 static void
2364 generate_subtree_copies (struct access *access, tree agg,
2365 HOST_WIDE_INT top_offset,
2366 HOST_WIDE_INT start_offset, HOST_WIDE_INT chunk_size,
2367 gimple_stmt_iterator *gsi, bool write,
2368 bool insert_after, location_t loc)
2369 {
2370 do
2371 {
2372 if (chunk_size && access->offset >= start_offset + chunk_size)
2373 return;
2374
2375 if (access->grp_to_be_replaced
2376 && (chunk_size == 0
2377 || access->offset + access->size > start_offset))
2378 {
2379 tree expr, repl = get_access_replacement (access);
2380 gimple stmt;
2381
2382 expr = build_ref_for_model (loc, agg, access->offset - top_offset,
2383 access, gsi, insert_after);
2384
2385 if (write)
2386 {
2387 if (access->grp_partial_lhs)
2388 expr = force_gimple_operand_gsi (gsi, expr, true, NULL_TREE,
2389 !insert_after,
2390 insert_after ? GSI_NEW_STMT
2391 : GSI_SAME_STMT);
2392 stmt = gimple_build_assign (repl, expr);
2393 }
2394 else
2395 {
2396 TREE_NO_WARNING (repl) = 1;
2397 if (access->grp_partial_lhs)
2398 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2399 !insert_after,
2400 insert_after ? GSI_NEW_STMT
2401 : GSI_SAME_STMT);
2402 stmt = gimple_build_assign (expr, repl);
2403 }
2404 gimple_set_location (stmt, loc);
2405
2406 if (insert_after)
2407 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2408 else
2409 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2410 update_stmt (stmt);
2411 sra_stats.subtree_copies++;
2412 }
2413
2414 if (access->first_child)
2415 generate_subtree_copies (access->first_child, agg, top_offset,
2416 start_offset, chunk_size, gsi,
2417 write, insert_after, loc);
2418
2419 access = access->next_sibling;
2420 }
2421 while (access);
2422 }
2423
2424 /* Assign zero to all scalar replacements in an access subtree. ACCESS is the
2425 the root of the subtree to be processed. GSI is the statement iterator used
2426 for inserting statements which are added after the current statement if
2427 INSERT_AFTER is true or before it otherwise. */
2428
2429 static void
2430 init_subtree_with_zero (struct access *access, gimple_stmt_iterator *gsi,
2431 bool insert_after, location_t loc)
2432
2433 {
2434 struct access *child;
2435
2436 if (access->grp_to_be_replaced)
2437 {
2438 gimple stmt;
2439
2440 stmt = gimple_build_assign (get_access_replacement (access),
2441 build_zero_cst (access->type));
2442 if (insert_after)
2443 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2444 else
2445 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2446 update_stmt (stmt);
2447 gimple_set_location (stmt, loc);
2448 }
2449
2450 for (child = access->first_child; child; child = child->next_sibling)
2451 init_subtree_with_zero (child, gsi, insert_after, loc);
2452 }
2453
2454 /* Search for an access representative for the given expression EXPR and
2455 return it or NULL if it cannot be found. */
2456
2457 static struct access *
2458 get_access_for_expr (tree expr)
2459 {
2460 HOST_WIDE_INT offset, size, max_size;
2461 tree base;
2462
2463 /* FIXME: This should not be necessary but Ada produces V_C_Es with a type of
2464 a different size than the size of its argument and we need the latter
2465 one. */
2466 if (TREE_CODE (expr) == VIEW_CONVERT_EXPR)
2467 expr = TREE_OPERAND (expr, 0);
2468
2469 base = get_ref_base_and_extent (expr, &offset, &size, &max_size);
2470 if (max_size == -1 || !DECL_P (base))
2471 return NULL;
2472
2473 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (base)))
2474 return NULL;
2475
2476 return get_var_base_offset_size_access (base, offset, max_size);
2477 }
2478
2479 /* Replace the expression EXPR with a scalar replacement if there is one and
2480 generate other statements to do type conversion or subtree copying if
2481 necessary. GSI is used to place newly created statements, WRITE is true if
2482 the expression is being written to (it is on a LHS of a statement or output
2483 in an assembly statement). */
2484
2485 static bool
2486 sra_modify_expr (tree *expr, gimple_stmt_iterator *gsi, bool write)
2487 {
2488 location_t loc;
2489 struct access *access;
2490 tree type, bfr;
2491
2492 if (TREE_CODE (*expr) == BIT_FIELD_REF)
2493 {
2494 bfr = *expr;
2495 expr = &TREE_OPERAND (*expr, 0);
2496 }
2497 else
2498 bfr = NULL_TREE;
2499
2500 if (TREE_CODE (*expr) == REALPART_EXPR || TREE_CODE (*expr) == IMAGPART_EXPR)
2501 expr = &TREE_OPERAND (*expr, 0);
2502 access = get_access_for_expr (*expr);
2503 if (!access)
2504 return false;
2505 type = TREE_TYPE (*expr);
2506
2507 loc = gimple_location (gsi_stmt (*gsi));
2508 if (access->grp_to_be_replaced)
2509 {
2510 tree repl = get_access_replacement (access);
2511 /* If we replace a non-register typed access simply use the original
2512 access expression to extract the scalar component afterwards.
2513 This happens if scalarizing a function return value or parameter
2514 like in gcc.c-torture/execute/20041124-1.c, 20050316-1.c and
2515 gcc.c-torture/compile/20011217-1.c.
2516
2517 We also want to use this when accessing a complex or vector which can
2518 be accessed as a different type too, potentially creating a need for
2519 type conversion (see PR42196) and when scalarized unions are involved
2520 in assembler statements (see PR42398). */
2521 if (!useless_type_conversion_p (type, access->type))
2522 {
2523 tree ref;
2524
2525 ref = build_ref_for_model (loc, access->base, access->offset, access,
2526 NULL, false);
2527
2528 if (write)
2529 {
2530 gimple stmt;
2531
2532 if (access->grp_partial_lhs)
2533 ref = force_gimple_operand_gsi (gsi, ref, true, NULL_TREE,
2534 false, GSI_NEW_STMT);
2535 stmt = gimple_build_assign (repl, ref);
2536 gimple_set_location (stmt, loc);
2537 gsi_insert_after (gsi, stmt, GSI_NEW_STMT);
2538 }
2539 else
2540 {
2541 gimple stmt;
2542
2543 if (access->grp_partial_lhs)
2544 repl = force_gimple_operand_gsi (gsi, repl, true, NULL_TREE,
2545 true, GSI_SAME_STMT);
2546 stmt = gimple_build_assign (ref, repl);
2547 gimple_set_location (stmt, loc);
2548 gsi_insert_before (gsi, stmt, GSI_SAME_STMT);
2549 }
2550 }
2551 else
2552 *expr = repl;
2553 sra_stats.exprs++;
2554 }
2555
2556 if (access->first_child)
2557 {
2558 HOST_WIDE_INT start_offset, chunk_size;
2559 if (bfr
2560 && host_integerp (TREE_OPERAND (bfr, 1), 1)
2561 && host_integerp (TREE_OPERAND (bfr, 2), 1))
2562 {
2563 chunk_size = tree_low_cst (TREE_OPERAND (bfr, 1), 1);
2564 start_offset = access->offset
2565 + tree_low_cst (TREE_OPERAND (bfr, 2), 1);
2566 }
2567 else
2568 start_offset = chunk_size = 0;
2569
2570 generate_subtree_copies (access->first_child, access->base, 0,
2571 start_offset, chunk_size, gsi, write, write,
2572 loc);
2573 }
2574 return true;
2575 }
2576
2577 /* Where scalar replacements of the RHS have been written to when a replacement
2578 of a LHS of an assigments cannot be direclty loaded from a replacement of
2579 the RHS. */
2580 enum unscalarized_data_handling { SRA_UDH_NONE, /* Nothing done so far. */
2581 SRA_UDH_RIGHT, /* Data flushed to the RHS. */
2582 SRA_UDH_LEFT }; /* Data flushed to the LHS. */
2583
2584 /* Store all replacements in the access tree rooted in TOP_RACC either to their
2585 base aggregate if there are unscalarized data or directly to LHS of the
2586 statement that is pointed to by GSI otherwise. */
2587
2588 static enum unscalarized_data_handling
2589 handle_unscalarized_data_in_subtree (struct access *top_racc,
2590 gimple_stmt_iterator *gsi)
2591 {
2592 if (top_racc->grp_unscalarized_data)
2593 {
2594 generate_subtree_copies (top_racc->first_child, top_racc->base, 0, 0, 0,
2595 gsi, false, false,
2596 gimple_location (gsi_stmt (*gsi)));
2597 return SRA_UDH_RIGHT;
2598 }
2599 else
2600 {
2601 tree lhs = gimple_assign_lhs (gsi_stmt (*gsi));
2602 generate_subtree_copies (top_racc->first_child, lhs, top_racc->offset,
2603 0, 0, gsi, false, false,
2604 gimple_location (gsi_stmt (*gsi)));
2605 return SRA_UDH_LEFT;
2606 }
2607 }
2608
2609
2610 /* Try to generate statements to load all sub-replacements in an access subtree
2611 formed by children of LACC from scalar replacements in the TOP_RACC subtree.
2612 If that is not possible, refresh the TOP_RACC base aggregate and load the
2613 accesses from it. LEFT_OFFSET is the offset of the left whole subtree being
2614 copied. NEW_GSI is stmt iterator used for statement insertions after the
2615 original assignment, OLD_GSI is used to insert statements before the
2616 assignment. *REFRESHED keeps the information whether we have needed to
2617 refresh replacements of the LHS and from which side of the assignments this
2618 takes place. */
2619
2620 static void
2621 load_assign_lhs_subreplacements (struct access *lacc, struct access *top_racc,
2622 HOST_WIDE_INT left_offset,
2623 gimple_stmt_iterator *old_gsi,
2624 gimple_stmt_iterator *new_gsi,
2625 enum unscalarized_data_handling *refreshed)
2626 {
2627 location_t loc = gimple_location (gsi_stmt (*old_gsi));
2628 for (lacc = lacc->first_child; lacc; lacc = lacc->next_sibling)
2629 {
2630 if (lacc->grp_to_be_replaced)
2631 {
2632 struct access *racc;
2633 HOST_WIDE_INT offset = lacc->offset - left_offset + top_racc->offset;
2634 gimple stmt;
2635 tree rhs;
2636
2637 racc = find_access_in_subtree (top_racc, offset, lacc->size);
2638 if (racc && racc->grp_to_be_replaced)
2639 {
2640 rhs = get_access_replacement (racc);
2641 if (!useless_type_conversion_p (lacc->type, racc->type))
2642 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, lacc->type, rhs);
2643 }
2644 else
2645 {
2646 /* No suitable access on the right hand side, need to load from
2647 the aggregate. See if we have to update it first... */
2648 if (*refreshed == SRA_UDH_NONE)
2649 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2650 old_gsi);
2651
2652 if (*refreshed == SRA_UDH_LEFT)
2653 rhs = build_ref_for_model (loc, lacc->base, lacc->offset, lacc,
2654 new_gsi, true);
2655 else
2656 rhs = build_ref_for_model (loc, top_racc->base, offset, lacc,
2657 new_gsi, true);
2658 }
2659
2660 stmt = gimple_build_assign (get_access_replacement (lacc), rhs);
2661 gsi_insert_after (new_gsi, stmt, GSI_NEW_STMT);
2662 gimple_set_location (stmt, loc);
2663 update_stmt (stmt);
2664 sra_stats.subreplacements++;
2665 }
2666 else if (*refreshed == SRA_UDH_NONE
2667 && lacc->grp_read && !lacc->grp_covered)
2668 *refreshed = handle_unscalarized_data_in_subtree (top_racc,
2669 old_gsi);
2670
2671 if (lacc->first_child)
2672 load_assign_lhs_subreplacements (lacc, top_racc, left_offset,
2673 old_gsi, new_gsi, refreshed);
2674 }
2675 }
2676
2677 /* Result code for SRA assignment modification. */
2678 enum assignment_mod_result { SRA_AM_NONE, /* nothing done for the stmt */
2679 SRA_AM_MODIFIED, /* stmt changed but not
2680 removed */
2681 SRA_AM_REMOVED }; /* stmt eliminated */
2682
2683 /* Modify assignments with a CONSTRUCTOR on their RHS. STMT contains a pointer
2684 to the assignment and GSI is the statement iterator pointing at it. Returns
2685 the same values as sra_modify_assign. */
2686
2687 static enum assignment_mod_result
2688 sra_modify_constructor_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2689 {
2690 tree lhs = gimple_assign_lhs (*stmt);
2691 struct access *acc;
2692 location_t loc;
2693
2694 acc = get_access_for_expr (lhs);
2695 if (!acc)
2696 return SRA_AM_NONE;
2697
2698 loc = gimple_location (*stmt);
2699 if (VEC_length (constructor_elt,
2700 CONSTRUCTOR_ELTS (gimple_assign_rhs1 (*stmt))) > 0)
2701 {
2702 /* I have never seen this code path trigger but if it can happen the
2703 following should handle it gracefully. */
2704 if (access_has_children_p (acc))
2705 generate_subtree_copies (acc->first_child, acc->base, 0, 0, 0, gsi,
2706 true, true, loc);
2707 return SRA_AM_MODIFIED;
2708 }
2709
2710 if (acc->grp_covered)
2711 {
2712 init_subtree_with_zero (acc, gsi, false, loc);
2713 unlink_stmt_vdef (*stmt);
2714 gsi_remove (gsi, true);
2715 return SRA_AM_REMOVED;
2716 }
2717 else
2718 {
2719 init_subtree_with_zero (acc, gsi, true, loc);
2720 return SRA_AM_MODIFIED;
2721 }
2722 }
2723
2724 /* Create and return a new suitable default definition SSA_NAME for RACC which
2725 is an access describing an uninitialized part of an aggregate that is being
2726 loaded. */
2727
2728 static tree
2729 get_repl_default_def_ssa_name (struct access *racc)
2730 {
2731 tree repl, decl;
2732
2733 decl = get_unrenamed_access_replacement (racc);
2734
2735 repl = gimple_default_def (cfun, decl);
2736 if (!repl)
2737 {
2738 repl = make_ssa_name (decl, gimple_build_nop ());
2739 set_default_def (decl, repl);
2740 }
2741
2742 return repl;
2743 }
2744
2745 /* Return true if REF has a COMPONENT_REF with a bit-field field declaration
2746 somewhere in it. */
2747
2748 static inline bool
2749 contains_bitfld_comp_ref_p (const_tree ref)
2750 {
2751 while (handled_component_p (ref))
2752 {
2753 if (TREE_CODE (ref) == COMPONENT_REF
2754 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1)))
2755 return true;
2756 ref = TREE_OPERAND (ref, 0);
2757 }
2758
2759 return false;
2760 }
2761
2762 /* Return true if REF has an VIEW_CONVERT_EXPR or a COMPONENT_REF with a
2763 bit-field field declaration somewhere in it. */
2764
2765 static inline bool
2766 contains_vce_or_bfcref_p (const_tree ref)
2767 {
2768 while (handled_component_p (ref))
2769 {
2770 if (TREE_CODE (ref) == VIEW_CONVERT_EXPR
2771 || (TREE_CODE (ref) == COMPONENT_REF
2772 && DECL_BIT_FIELD (TREE_OPERAND (ref, 1))))
2773 return true;
2774 ref = TREE_OPERAND (ref, 0);
2775 }
2776
2777 return false;
2778 }
2779
2780 /* Examine both sides of the assignment statement pointed to by STMT, replace
2781 them with a scalare replacement if there is one and generate copying of
2782 replacements if scalarized aggregates have been used in the assignment. GSI
2783 is used to hold generated statements for type conversions and subtree
2784 copying. */
2785
2786 static enum assignment_mod_result
2787 sra_modify_assign (gimple *stmt, gimple_stmt_iterator *gsi)
2788 {
2789 struct access *lacc, *racc;
2790 tree lhs, rhs;
2791 bool modify_this_stmt = false;
2792 bool force_gimple_rhs = false;
2793 location_t loc;
2794 gimple_stmt_iterator orig_gsi = *gsi;
2795
2796 if (!gimple_assign_single_p (*stmt))
2797 return SRA_AM_NONE;
2798 lhs = gimple_assign_lhs (*stmt);
2799 rhs = gimple_assign_rhs1 (*stmt);
2800
2801 if (TREE_CODE (rhs) == CONSTRUCTOR)
2802 return sra_modify_constructor_assign (stmt, gsi);
2803
2804 if (TREE_CODE (rhs) == REALPART_EXPR || TREE_CODE (lhs) == REALPART_EXPR
2805 || TREE_CODE (rhs) == IMAGPART_EXPR || TREE_CODE (lhs) == IMAGPART_EXPR
2806 || TREE_CODE (rhs) == BIT_FIELD_REF || TREE_CODE (lhs) == BIT_FIELD_REF)
2807 {
2808 modify_this_stmt = sra_modify_expr (gimple_assign_rhs1_ptr (*stmt),
2809 gsi, false);
2810 modify_this_stmt |= sra_modify_expr (gimple_assign_lhs_ptr (*stmt),
2811 gsi, true);
2812 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2813 }
2814
2815 lacc = get_access_for_expr (lhs);
2816 racc = get_access_for_expr (rhs);
2817 if (!lacc && !racc)
2818 return SRA_AM_NONE;
2819
2820 loc = gimple_location (*stmt);
2821 if (lacc && lacc->grp_to_be_replaced)
2822 {
2823 lhs = get_access_replacement (lacc);
2824 gimple_assign_set_lhs (*stmt, lhs);
2825 modify_this_stmt = true;
2826 if (lacc->grp_partial_lhs)
2827 force_gimple_rhs = true;
2828 sra_stats.exprs++;
2829 }
2830
2831 if (racc && racc->grp_to_be_replaced)
2832 {
2833 rhs = get_access_replacement (racc);
2834 modify_this_stmt = true;
2835 if (racc->grp_partial_lhs)
2836 force_gimple_rhs = true;
2837 sra_stats.exprs++;
2838 }
2839
2840 if (modify_this_stmt)
2841 {
2842 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2843 {
2844 /* If we can avoid creating a VIEW_CONVERT_EXPR do so.
2845 ??? This should move to fold_stmt which we simply should
2846 call after building a VIEW_CONVERT_EXPR here. */
2847 if (AGGREGATE_TYPE_P (TREE_TYPE (lhs))
2848 && !contains_bitfld_comp_ref_p (lhs)
2849 && !access_has_children_p (lacc))
2850 {
2851 lhs = build_ref_for_model (loc, lhs, 0, racc, gsi, false);
2852 gimple_assign_set_lhs (*stmt, lhs);
2853 }
2854 else if (AGGREGATE_TYPE_P (TREE_TYPE (rhs))
2855 && !contains_vce_or_bfcref_p (rhs)
2856 && !access_has_children_p (racc))
2857 rhs = build_ref_for_model (loc, rhs, 0, lacc, gsi, false);
2858
2859 if (!useless_type_conversion_p (TREE_TYPE (lhs), TREE_TYPE (rhs)))
2860 {
2861 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR, TREE_TYPE (lhs),
2862 rhs);
2863 if (is_gimple_reg_type (TREE_TYPE (lhs))
2864 && TREE_CODE (lhs) != SSA_NAME)
2865 force_gimple_rhs = true;
2866 }
2867 }
2868 }
2869
2870 /* From this point on, the function deals with assignments in between
2871 aggregates when at least one has scalar reductions of some of its
2872 components. There are three possible scenarios: Both the LHS and RHS have
2873 to-be-scalarized components, 2) only the RHS has or 3) only the LHS has.
2874
2875 In the first case, we would like to load the LHS components from RHS
2876 components whenever possible. If that is not possible, we would like to
2877 read it directly from the RHS (after updating it by storing in it its own
2878 components). If there are some necessary unscalarized data in the LHS,
2879 those will be loaded by the original assignment too. If neither of these
2880 cases happen, the original statement can be removed. Most of this is done
2881 by load_assign_lhs_subreplacements.
2882
2883 In the second case, we would like to store all RHS scalarized components
2884 directly into LHS and if they cover the aggregate completely, remove the
2885 statement too. In the third case, we want the LHS components to be loaded
2886 directly from the RHS (DSE will remove the original statement if it
2887 becomes redundant).
2888
2889 This is a bit complex but manageable when types match and when unions do
2890 not cause confusion in a way that we cannot really load a component of LHS
2891 from the RHS or vice versa (the access representing this level can have
2892 subaccesses that are accessible only through a different union field at a
2893 higher level - different from the one used in the examined expression).
2894 Unions are fun.
2895
2896 Therefore, I specially handle a fourth case, happening when there is a
2897 specific type cast or it is impossible to locate a scalarized subaccess on
2898 the other side of the expression. If that happens, I simply "refresh" the
2899 RHS by storing in it is scalarized components leave the original statement
2900 there to do the copying and then load the scalar replacements of the LHS.
2901 This is what the first branch does. */
2902
2903 if (modify_this_stmt
2904 || gimple_has_volatile_ops (*stmt)
2905 || contains_vce_or_bfcref_p (rhs)
2906 || contains_vce_or_bfcref_p (lhs))
2907 {
2908 if (access_has_children_p (racc))
2909 generate_subtree_copies (racc->first_child, racc->base, 0, 0, 0,
2910 gsi, false, false, loc);
2911 if (access_has_children_p (lacc))
2912 generate_subtree_copies (lacc->first_child, lacc->base, 0, 0, 0,
2913 gsi, true, true, loc);
2914 sra_stats.separate_lhs_rhs_handling++;
2915 }
2916 else
2917 {
2918 if (access_has_children_p (lacc) && access_has_children_p (racc))
2919 {
2920 gimple_stmt_iterator orig_gsi = *gsi;
2921 enum unscalarized_data_handling refreshed;
2922
2923 if (lacc->grp_read && !lacc->grp_covered)
2924 refreshed = handle_unscalarized_data_in_subtree (racc, gsi);
2925 else
2926 refreshed = SRA_UDH_NONE;
2927
2928 load_assign_lhs_subreplacements (lacc, racc, lacc->offset,
2929 &orig_gsi, gsi, &refreshed);
2930 if (refreshed != SRA_UDH_RIGHT)
2931 {
2932 gsi_next (gsi);
2933 unlink_stmt_vdef (*stmt);
2934 gsi_remove (&orig_gsi, true);
2935 sra_stats.deleted++;
2936 return SRA_AM_REMOVED;
2937 }
2938 }
2939 else
2940 {
2941 if (racc)
2942 {
2943 if (!racc->grp_to_be_replaced && !racc->grp_unscalarized_data)
2944 {
2945 if (dump_file)
2946 {
2947 fprintf (dump_file, "Removing load: ");
2948 print_gimple_stmt (dump_file, *stmt, 0, 0);
2949 }
2950
2951 if (TREE_CODE (lhs) == SSA_NAME)
2952 {
2953 rhs = get_repl_default_def_ssa_name (racc);
2954 if (!useless_type_conversion_p (TREE_TYPE (lhs),
2955 TREE_TYPE (rhs)))
2956 rhs = fold_build1_loc (loc, VIEW_CONVERT_EXPR,
2957 TREE_TYPE (lhs), rhs);
2958 }
2959 else
2960 {
2961 if (racc->first_child)
2962 generate_subtree_copies (racc->first_child, lhs,
2963 racc->offset, 0, 0, gsi,
2964 false, false, loc);
2965
2966 gcc_assert (*stmt == gsi_stmt (*gsi));
2967 unlink_stmt_vdef (*stmt);
2968 gsi_remove (gsi, true);
2969 sra_stats.deleted++;
2970 return SRA_AM_REMOVED;
2971 }
2972 }
2973 else if (racc->first_child)
2974 generate_subtree_copies (racc->first_child, lhs, racc->offset,
2975 0, 0, gsi, false, true, loc);
2976 }
2977 if (access_has_children_p (lacc))
2978 generate_subtree_copies (lacc->first_child, rhs, lacc->offset,
2979 0, 0, gsi, true, true, loc);
2980 }
2981 }
2982
2983 /* This gimplification must be done after generate_subtree_copies, lest we
2984 insert the subtree copies in the middle of the gimplified sequence. */
2985 if (force_gimple_rhs)
2986 rhs = force_gimple_operand_gsi (&orig_gsi, rhs, true, NULL_TREE,
2987 true, GSI_SAME_STMT);
2988 if (gimple_assign_rhs1 (*stmt) != rhs)
2989 {
2990 modify_this_stmt = true;
2991 gimple_assign_set_rhs_from_tree (&orig_gsi, rhs);
2992 gcc_assert (*stmt == gsi_stmt (orig_gsi));
2993 }
2994
2995 return modify_this_stmt ? SRA_AM_MODIFIED : SRA_AM_NONE;
2996 }
2997
2998 /* Traverse the function body and all modifications as decided in
2999 analyze_all_variable_accesses. Return true iff the CFG has been
3000 changed. */
3001
3002 static bool
3003 sra_modify_function_body (void)
3004 {
3005 bool cfg_changed = false;
3006 basic_block bb;
3007
3008 FOR_EACH_BB (bb)
3009 {
3010 gimple_stmt_iterator gsi = gsi_start_bb (bb);
3011 while (!gsi_end_p (gsi))
3012 {
3013 gimple stmt = gsi_stmt (gsi);
3014 enum assignment_mod_result assign_result;
3015 bool modified = false, deleted = false;
3016 tree *t;
3017 unsigned i;
3018
3019 switch (gimple_code (stmt))
3020 {
3021 case GIMPLE_RETURN:
3022 t = gimple_return_retval_ptr (stmt);
3023 if (*t != NULL_TREE)
3024 modified |= sra_modify_expr (t, &gsi, false);
3025 break;
3026
3027 case GIMPLE_ASSIGN:
3028 assign_result = sra_modify_assign (&stmt, &gsi);
3029 modified |= assign_result == SRA_AM_MODIFIED;
3030 deleted = assign_result == SRA_AM_REMOVED;
3031 break;
3032
3033 case GIMPLE_CALL:
3034 /* Operands must be processed before the lhs. */
3035 for (i = 0; i < gimple_call_num_args (stmt); i++)
3036 {
3037 t = gimple_call_arg_ptr (stmt, i);
3038 modified |= sra_modify_expr (t, &gsi, false);
3039 }
3040
3041 if (gimple_call_lhs (stmt))
3042 {
3043 t = gimple_call_lhs_ptr (stmt);
3044 modified |= sra_modify_expr (t, &gsi, true);
3045 }
3046 break;
3047
3048 case GIMPLE_ASM:
3049 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
3050 {
3051 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
3052 modified |= sra_modify_expr (t, &gsi, false);
3053 }
3054 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
3055 {
3056 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
3057 modified |= sra_modify_expr (t, &gsi, true);
3058 }
3059 break;
3060
3061 default:
3062 break;
3063 }
3064
3065 if (modified)
3066 {
3067 update_stmt (stmt);
3068 if (maybe_clean_eh_stmt (stmt)
3069 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
3070 cfg_changed = true;
3071 }
3072 if (!deleted)
3073 gsi_next (&gsi);
3074 }
3075 }
3076
3077 return cfg_changed;
3078 }
3079
3080 /* Generate statements initializing scalar replacements of parts of function
3081 parameters. */
3082
3083 static void
3084 initialize_parameter_reductions (void)
3085 {
3086 gimple_stmt_iterator gsi;
3087 gimple_seq seq = NULL;
3088 tree parm;
3089
3090 for (parm = DECL_ARGUMENTS (current_function_decl);
3091 parm;
3092 parm = DECL_CHAIN (parm))
3093 {
3094 VEC (access_p, heap) *access_vec;
3095 struct access *access;
3096
3097 if (!bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3098 continue;
3099 access_vec = get_base_access_vector (parm);
3100 if (!access_vec)
3101 continue;
3102
3103 if (!seq)
3104 {
3105 seq = gimple_seq_alloc ();
3106 gsi = gsi_start (seq);
3107 }
3108
3109 for (access = VEC_index (access_p, access_vec, 0);
3110 access;
3111 access = access->next_grp)
3112 generate_subtree_copies (access, parm, 0, 0, 0, &gsi, true, true,
3113 EXPR_LOCATION (parm));
3114 }
3115
3116 if (seq)
3117 gsi_insert_seq_on_edge_immediate (single_succ_edge (ENTRY_BLOCK_PTR), seq);
3118 }
3119
3120 /* The "main" function of intraprocedural SRA passes. Runs the analysis and if
3121 it reveals there are components of some aggregates to be scalarized, it runs
3122 the required transformations. */
3123 static unsigned int
3124 perform_intra_sra (void)
3125 {
3126 int ret = 0;
3127 sra_initialize ();
3128
3129 if (!find_var_candidates ())
3130 goto out;
3131
3132 if (!scan_function ())
3133 goto out;
3134
3135 if (!analyze_all_variable_accesses ())
3136 goto out;
3137
3138 if (sra_modify_function_body ())
3139 ret = TODO_update_ssa | TODO_cleanup_cfg;
3140 else
3141 ret = TODO_update_ssa;
3142 initialize_parameter_reductions ();
3143
3144 statistics_counter_event (cfun, "Scalar replacements created",
3145 sra_stats.replacements);
3146 statistics_counter_event (cfun, "Modified expressions", sra_stats.exprs);
3147 statistics_counter_event (cfun, "Subtree copy stmts",
3148 sra_stats.subtree_copies);
3149 statistics_counter_event (cfun, "Subreplacement stmts",
3150 sra_stats.subreplacements);
3151 statistics_counter_event (cfun, "Deleted stmts", sra_stats.deleted);
3152 statistics_counter_event (cfun, "Separate LHS and RHS handling",
3153 sra_stats.separate_lhs_rhs_handling);
3154
3155 out:
3156 sra_deinitialize ();
3157 return ret;
3158 }
3159
3160 /* Perform early intraprocedural SRA. */
3161 static unsigned int
3162 early_intra_sra (void)
3163 {
3164 sra_mode = SRA_MODE_EARLY_INTRA;
3165 return perform_intra_sra ();
3166 }
3167
3168 /* Perform "late" intraprocedural SRA. */
3169 static unsigned int
3170 late_intra_sra (void)
3171 {
3172 sra_mode = SRA_MODE_INTRA;
3173 return perform_intra_sra ();
3174 }
3175
3176
3177 static bool
3178 gate_intra_sra (void)
3179 {
3180 return flag_tree_sra != 0 && dbg_cnt (tree_sra);
3181 }
3182
3183
3184 struct gimple_opt_pass pass_sra_early =
3185 {
3186 {
3187 GIMPLE_PASS,
3188 "esra", /* name */
3189 gate_intra_sra, /* gate */
3190 early_intra_sra, /* execute */
3191 NULL, /* sub */
3192 NULL, /* next */
3193 0, /* static_pass_number */
3194 TV_TREE_SRA, /* tv_id */
3195 PROP_cfg | PROP_ssa, /* properties_required */
3196 0, /* properties_provided */
3197 0, /* properties_destroyed */
3198 0, /* todo_flags_start */
3199 TODO_update_ssa
3200 | TODO_ggc_collect
3201 | TODO_verify_ssa /* todo_flags_finish */
3202 }
3203 };
3204
3205 struct gimple_opt_pass pass_sra =
3206 {
3207 {
3208 GIMPLE_PASS,
3209 "sra", /* name */
3210 gate_intra_sra, /* gate */
3211 late_intra_sra, /* execute */
3212 NULL, /* sub */
3213 NULL, /* next */
3214 0, /* static_pass_number */
3215 TV_TREE_SRA, /* tv_id */
3216 PROP_cfg | PROP_ssa, /* properties_required */
3217 0, /* properties_provided */
3218 0, /* properties_destroyed */
3219 TODO_update_address_taken, /* todo_flags_start */
3220 TODO_update_ssa
3221 | TODO_ggc_collect
3222 | TODO_verify_ssa /* todo_flags_finish */
3223 }
3224 };
3225
3226
3227 /* Return true iff PARM (which must be a parm_decl) is an unused scalar
3228 parameter. */
3229
3230 static bool
3231 is_unused_scalar_param (tree parm)
3232 {
3233 tree name;
3234 return (is_gimple_reg (parm)
3235 && (!(name = gimple_default_def (cfun, parm))
3236 || has_zero_uses (name)));
3237 }
3238
3239 /* Scan immediate uses of a default definition SSA name of a parameter PARM and
3240 examine whether there are any direct or otherwise infeasible ones. If so,
3241 return true, otherwise return false. PARM must be a gimple register with a
3242 non-NULL default definition. */
3243
3244 static bool
3245 ptr_parm_has_direct_uses (tree parm)
3246 {
3247 imm_use_iterator ui;
3248 gimple stmt;
3249 tree name = gimple_default_def (cfun, parm);
3250 bool ret = false;
3251
3252 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
3253 {
3254 int uses_ok = 0;
3255 use_operand_p use_p;
3256
3257 if (is_gimple_debug (stmt))
3258 continue;
3259
3260 /* Valid uses include dereferences on the lhs and the rhs. */
3261 if (gimple_has_lhs (stmt))
3262 {
3263 tree lhs = gimple_get_lhs (stmt);
3264 while (handled_component_p (lhs))
3265 lhs = TREE_OPERAND (lhs, 0);
3266 if (TREE_CODE (lhs) == MEM_REF
3267 && TREE_OPERAND (lhs, 0) == name
3268 && integer_zerop (TREE_OPERAND (lhs, 1))
3269 && types_compatible_p (TREE_TYPE (lhs),
3270 TREE_TYPE (TREE_TYPE (name))))
3271 uses_ok++;
3272 }
3273 if (gimple_assign_single_p (stmt))
3274 {
3275 tree rhs = gimple_assign_rhs1 (stmt);
3276 while (handled_component_p (rhs))
3277 rhs = TREE_OPERAND (rhs, 0);
3278 if (TREE_CODE (rhs) == MEM_REF
3279 && TREE_OPERAND (rhs, 0) == name
3280 && integer_zerop (TREE_OPERAND (rhs, 1))
3281 && types_compatible_p (TREE_TYPE (rhs),
3282 TREE_TYPE (TREE_TYPE (name))))
3283 uses_ok++;
3284 }
3285 else if (is_gimple_call (stmt))
3286 {
3287 unsigned i;
3288 for (i = 0; i < gimple_call_num_args (stmt); ++i)
3289 {
3290 tree arg = gimple_call_arg (stmt, i);
3291 while (handled_component_p (arg))
3292 arg = TREE_OPERAND (arg, 0);
3293 if (TREE_CODE (arg) == MEM_REF
3294 && TREE_OPERAND (arg, 0) == name
3295 && integer_zerop (TREE_OPERAND (arg, 1))
3296 && types_compatible_p (TREE_TYPE (arg),
3297 TREE_TYPE (TREE_TYPE (name))))
3298 uses_ok++;
3299 }
3300 }
3301
3302 /* If the number of valid uses does not match the number of
3303 uses in this stmt there is an unhandled use. */
3304 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
3305 --uses_ok;
3306
3307 if (uses_ok != 0)
3308 ret = true;
3309
3310 if (ret)
3311 BREAK_FROM_IMM_USE_STMT (ui);
3312 }
3313
3314 return ret;
3315 }
3316
3317 /* Identify candidates for reduction for IPA-SRA based on their type and mark
3318 them in candidate_bitmap. Note that these do not necessarily include
3319 parameter which are unused and thus can be removed. Return true iff any
3320 such candidate has been found. */
3321
3322 static bool
3323 find_param_candidates (void)
3324 {
3325 tree parm;
3326 int count = 0;
3327 bool ret = false;
3328 const char *msg;
3329
3330 for (parm = DECL_ARGUMENTS (current_function_decl);
3331 parm;
3332 parm = DECL_CHAIN (parm))
3333 {
3334 tree type = TREE_TYPE (parm);
3335
3336 count++;
3337
3338 if (TREE_THIS_VOLATILE (parm)
3339 || TREE_ADDRESSABLE (parm)
3340 || (!is_gimple_reg_type (type) && is_va_list_type (type)))
3341 continue;
3342
3343 if (is_unused_scalar_param (parm))
3344 {
3345 ret = true;
3346 continue;
3347 }
3348
3349 if (POINTER_TYPE_P (type))
3350 {
3351 type = TREE_TYPE (type);
3352
3353 if (TREE_CODE (type) == FUNCTION_TYPE
3354 || TYPE_VOLATILE (type)
3355 || (TREE_CODE (type) == ARRAY_TYPE
3356 && TYPE_NONALIASED_COMPONENT (type))
3357 || !is_gimple_reg (parm)
3358 || is_va_list_type (type)
3359 || ptr_parm_has_direct_uses (parm))
3360 continue;
3361 }
3362 else if (!AGGREGATE_TYPE_P (type))
3363 continue;
3364
3365 if (!COMPLETE_TYPE_P (type)
3366 || !host_integerp (TYPE_SIZE (type), 1)
3367 || tree_low_cst (TYPE_SIZE (type), 1) == 0
3368 || (AGGREGATE_TYPE_P (type)
3369 && type_internals_preclude_sra_p (type, &msg)))
3370 continue;
3371
3372 bitmap_set_bit (candidate_bitmap, DECL_UID (parm));
3373 ret = true;
3374 if (dump_file && (dump_flags & TDF_DETAILS))
3375 {
3376 fprintf (dump_file, "Candidate (%d): ", DECL_UID (parm));
3377 print_generic_expr (dump_file, parm, 0);
3378 fprintf (dump_file, "\n");
3379 }
3380 }
3381
3382 func_param_count = count;
3383 return ret;
3384 }
3385
3386 /* Callback of walk_aliased_vdefs, marks the access passed as DATA as
3387 maybe_modified. */
3388
3389 static bool
3390 mark_maybe_modified (ao_ref *ao ATTRIBUTE_UNUSED, tree vdef ATTRIBUTE_UNUSED,
3391 void *data)
3392 {
3393 struct access *repr = (struct access *) data;
3394
3395 repr->grp_maybe_modified = 1;
3396 return true;
3397 }
3398
3399 /* Analyze what representatives (in linked lists accessible from
3400 REPRESENTATIVES) can be modified by side effects of statements in the
3401 current function. */
3402
3403 static void
3404 analyze_modified_params (VEC (access_p, heap) *representatives)
3405 {
3406 int i;
3407
3408 for (i = 0; i < func_param_count; i++)
3409 {
3410 struct access *repr;
3411
3412 for (repr = VEC_index (access_p, representatives, i);
3413 repr;
3414 repr = repr->next_grp)
3415 {
3416 struct access *access;
3417 bitmap visited;
3418 ao_ref ar;
3419
3420 if (no_accesses_p (repr))
3421 continue;
3422 if (!POINTER_TYPE_P (TREE_TYPE (repr->base))
3423 || repr->grp_maybe_modified)
3424 continue;
3425
3426 ao_ref_init (&ar, repr->expr);
3427 visited = BITMAP_ALLOC (NULL);
3428 for (access = repr; access; access = access->next_sibling)
3429 {
3430 /* All accesses are read ones, otherwise grp_maybe_modified would
3431 be trivially set. */
3432 walk_aliased_vdefs (&ar, gimple_vuse (access->stmt),
3433 mark_maybe_modified, repr, &visited);
3434 if (repr->grp_maybe_modified)
3435 break;
3436 }
3437 BITMAP_FREE (visited);
3438 }
3439 }
3440 }
3441
3442 /* Propagate distances in bb_dereferences in the opposite direction than the
3443 control flow edges, in each step storing the maximum of the current value
3444 and the minimum of all successors. These steps are repeated until the table
3445 stabilizes. Note that BBs which might terminate the functions (according to
3446 final_bbs bitmap) never updated in this way. */
3447
3448 static void
3449 propagate_dereference_distances (void)
3450 {
3451 VEC (basic_block, heap) *queue;
3452 basic_block bb;
3453
3454 queue = VEC_alloc (basic_block, heap, last_basic_block_for_function (cfun));
3455 VEC_quick_push (basic_block, queue, ENTRY_BLOCK_PTR);
3456 FOR_EACH_BB (bb)
3457 {
3458 VEC_quick_push (basic_block, queue, bb);
3459 bb->aux = bb;
3460 }
3461
3462 while (!VEC_empty (basic_block, queue))
3463 {
3464 edge_iterator ei;
3465 edge e;
3466 bool change = false;
3467 int i;
3468
3469 bb = VEC_pop (basic_block, queue);
3470 bb->aux = NULL;
3471
3472 if (bitmap_bit_p (final_bbs, bb->index))
3473 continue;
3474
3475 for (i = 0; i < func_param_count; i++)
3476 {
3477 int idx = bb->index * func_param_count + i;
3478 bool first = true;
3479 HOST_WIDE_INT inh = 0;
3480
3481 FOR_EACH_EDGE (e, ei, bb->succs)
3482 {
3483 int succ_idx = e->dest->index * func_param_count + i;
3484
3485 if (e->src == EXIT_BLOCK_PTR)
3486 continue;
3487
3488 if (first)
3489 {
3490 first = false;
3491 inh = bb_dereferences [succ_idx];
3492 }
3493 else if (bb_dereferences [succ_idx] < inh)
3494 inh = bb_dereferences [succ_idx];
3495 }
3496
3497 if (!first && bb_dereferences[idx] < inh)
3498 {
3499 bb_dereferences[idx] = inh;
3500 change = true;
3501 }
3502 }
3503
3504 if (change && !bitmap_bit_p (final_bbs, bb->index))
3505 FOR_EACH_EDGE (e, ei, bb->preds)
3506 {
3507 if (e->src->aux)
3508 continue;
3509
3510 e->src->aux = e->src;
3511 VEC_quick_push (basic_block, queue, e->src);
3512 }
3513 }
3514
3515 VEC_free (basic_block, heap, queue);
3516 }
3517
3518 /* Dump a dereferences TABLE with heading STR to file F. */
3519
3520 static void
3521 dump_dereferences_table (FILE *f, const char *str, HOST_WIDE_INT *table)
3522 {
3523 basic_block bb;
3524
3525 fprintf (dump_file, str);
3526 FOR_BB_BETWEEN (bb, ENTRY_BLOCK_PTR, EXIT_BLOCK_PTR, next_bb)
3527 {
3528 fprintf (f, "%4i %i ", bb->index, bitmap_bit_p (final_bbs, bb->index));
3529 if (bb != EXIT_BLOCK_PTR)
3530 {
3531 int i;
3532 for (i = 0; i < func_param_count; i++)
3533 {
3534 int idx = bb->index * func_param_count + i;
3535 fprintf (f, " %4" HOST_WIDE_INT_PRINT "d", table[idx]);
3536 }
3537 }
3538 fprintf (f, "\n");
3539 }
3540 fprintf (dump_file, "\n");
3541 }
3542
3543 /* Determine what (parts of) parameters passed by reference that are not
3544 assigned to are not certainly dereferenced in this function and thus the
3545 dereferencing cannot be safely moved to the caller without potentially
3546 introducing a segfault. Mark such REPRESENTATIVES as
3547 grp_not_necessarilly_dereferenced.
3548
3549 The dereferenced maximum "distance," i.e. the offset + size of the accessed
3550 part is calculated rather than simple booleans are calculated for each
3551 pointer parameter to handle cases when only a fraction of the whole
3552 aggregate is allocated (see testsuite/gcc.c-torture/execute/ipa-sra-2.c for
3553 an example).
3554
3555 The maximum dereference distances for each pointer parameter and BB are
3556 already stored in bb_dereference. This routine simply propagates these
3557 values upwards by propagate_dereference_distances and then compares the
3558 distances of individual parameters in the ENTRY BB to the equivalent
3559 distances of each representative of a (fraction of a) parameter. */
3560
3561 static void
3562 analyze_caller_dereference_legality (VEC (access_p, heap) *representatives)
3563 {
3564 int i;
3565
3566 if (dump_file && (dump_flags & TDF_DETAILS))
3567 dump_dereferences_table (dump_file,
3568 "Dereference table before propagation:\n",
3569 bb_dereferences);
3570
3571 propagate_dereference_distances ();
3572
3573 if (dump_file && (dump_flags & TDF_DETAILS))
3574 dump_dereferences_table (dump_file,
3575 "Dereference table after propagation:\n",
3576 bb_dereferences);
3577
3578 for (i = 0; i < func_param_count; i++)
3579 {
3580 struct access *repr = VEC_index (access_p, representatives, i);
3581 int idx = ENTRY_BLOCK_PTR->index * func_param_count + i;
3582
3583 if (!repr || no_accesses_p (repr))
3584 continue;
3585
3586 do
3587 {
3588 if ((repr->offset + repr->size) > bb_dereferences[idx])
3589 repr->grp_not_necessarilly_dereferenced = 1;
3590 repr = repr->next_grp;
3591 }
3592 while (repr);
3593 }
3594 }
3595
3596 /* Return the representative access for the parameter declaration PARM if it is
3597 a scalar passed by reference which is not written to and the pointer value
3598 is not used directly. Thus, if it is legal to dereference it in the caller
3599 and we can rule out modifications through aliases, such parameter should be
3600 turned into one passed by value. Return NULL otherwise. */
3601
3602 static struct access *
3603 unmodified_by_ref_scalar_representative (tree parm)
3604 {
3605 int i, access_count;
3606 struct access *repr;
3607 VEC (access_p, heap) *access_vec;
3608
3609 access_vec = get_base_access_vector (parm);
3610 gcc_assert (access_vec);
3611 repr = VEC_index (access_p, access_vec, 0);
3612 if (repr->write)
3613 return NULL;
3614 repr->group_representative = repr;
3615
3616 access_count = VEC_length (access_p, access_vec);
3617 for (i = 1; i < access_count; i++)
3618 {
3619 struct access *access = VEC_index (access_p, access_vec, i);
3620 if (access->write)
3621 return NULL;
3622 access->group_representative = repr;
3623 access->next_sibling = repr->next_sibling;
3624 repr->next_sibling = access;
3625 }
3626
3627 repr->grp_read = 1;
3628 repr->grp_scalar_ptr = 1;
3629 return repr;
3630 }
3631
3632 /* Return true iff this access precludes IPA-SRA of the parameter it is
3633 associated with. */
3634
3635 static bool
3636 access_precludes_ipa_sra_p (struct access *access)
3637 {
3638 /* Avoid issues such as the second simple testcase in PR 42025. The problem
3639 is incompatible assign in a call statement (and possibly even in asm
3640 statements). This can be relaxed by using a new temporary but only for
3641 non-TREE_ADDRESSABLE types and is probably not worth the complexity. (In
3642 intraprocedural SRA we deal with this by keeping the old aggregate around,
3643 something we cannot do in IPA-SRA.) */
3644 if (access->write
3645 && (is_gimple_call (access->stmt)
3646 || gimple_code (access->stmt) == GIMPLE_ASM))
3647 return true;
3648
3649 return false;
3650 }
3651
3652
3653 /* Sort collected accesses for parameter PARM, identify representatives for
3654 each accessed region and link them together. Return NULL if there are
3655 different but overlapping accesses, return the special ptr value meaning
3656 there are no accesses for this parameter if that is the case and return the
3657 first representative otherwise. Set *RO_GRP if there is a group of accesses
3658 with only read (i.e. no write) accesses. */
3659
3660 static struct access *
3661 splice_param_accesses (tree parm, bool *ro_grp)
3662 {
3663 int i, j, access_count, group_count;
3664 int agg_size, total_size = 0;
3665 struct access *access, *res, **prev_acc_ptr = &res;
3666 VEC (access_p, heap) *access_vec;
3667
3668 access_vec = get_base_access_vector (parm);
3669 if (!access_vec)
3670 return &no_accesses_representant;
3671 access_count = VEC_length (access_p, access_vec);
3672
3673 VEC_qsort (access_p, access_vec, compare_access_positions);
3674
3675 i = 0;
3676 total_size = 0;
3677 group_count = 0;
3678 while (i < access_count)
3679 {
3680 bool modification;
3681 tree a1_alias_type;
3682 access = VEC_index (access_p, access_vec, i);
3683 modification = access->write;
3684 if (access_precludes_ipa_sra_p (access))
3685 return NULL;
3686 a1_alias_type = reference_alias_ptr_type (access->expr);
3687
3688 /* Access is about to become group representative unless we find some
3689 nasty overlap which would preclude us from breaking this parameter
3690 apart. */
3691
3692 j = i + 1;
3693 while (j < access_count)
3694 {
3695 struct access *ac2 = VEC_index (access_p, access_vec, j);
3696 if (ac2->offset != access->offset)
3697 {
3698 /* All or nothing law for parameters. */
3699 if (access->offset + access->size > ac2->offset)
3700 return NULL;
3701 else
3702 break;
3703 }
3704 else if (ac2->size != access->size)
3705 return NULL;
3706
3707 if (access_precludes_ipa_sra_p (ac2)
3708 || (ac2->type != access->type
3709 && (TREE_ADDRESSABLE (ac2->type)
3710 || TREE_ADDRESSABLE (access->type)))
3711 || (reference_alias_ptr_type (ac2->expr) != a1_alias_type))
3712 return NULL;
3713
3714 modification |= ac2->write;
3715 ac2->group_representative = access;
3716 ac2->next_sibling = access->next_sibling;
3717 access->next_sibling = ac2;
3718 j++;
3719 }
3720
3721 group_count++;
3722 access->grp_maybe_modified = modification;
3723 if (!modification)
3724 *ro_grp = true;
3725 *prev_acc_ptr = access;
3726 prev_acc_ptr = &access->next_grp;
3727 total_size += access->size;
3728 i = j;
3729 }
3730
3731 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3732 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3733 else
3734 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3735 if (total_size >= agg_size)
3736 return NULL;
3737
3738 gcc_assert (group_count > 0);
3739 return res;
3740 }
3741
3742 /* Decide whether parameters with representative accesses given by REPR should
3743 be reduced into components. */
3744
3745 static int
3746 decide_one_param_reduction (struct access *repr)
3747 {
3748 int total_size, cur_parm_size, agg_size, new_param_count, parm_size_limit;
3749 bool by_ref;
3750 tree parm;
3751
3752 parm = repr->base;
3753 cur_parm_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (parm)), 1);
3754 gcc_assert (cur_parm_size > 0);
3755
3756 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3757 {
3758 by_ref = true;
3759 agg_size = tree_low_cst (TYPE_SIZE (TREE_TYPE (TREE_TYPE (parm))), 1);
3760 }
3761 else
3762 {
3763 by_ref = false;
3764 agg_size = cur_parm_size;
3765 }
3766
3767 if (dump_file)
3768 {
3769 struct access *acc;
3770 fprintf (dump_file, "Evaluating PARAM group sizes for ");
3771 print_generic_expr (dump_file, parm, 0);
3772 fprintf (dump_file, " (UID: %u): \n", DECL_UID (parm));
3773 for (acc = repr; acc; acc = acc->next_grp)
3774 dump_access (dump_file, acc, true);
3775 }
3776
3777 total_size = 0;
3778 new_param_count = 0;
3779
3780 for (; repr; repr = repr->next_grp)
3781 {
3782 gcc_assert (parm == repr->base);
3783
3784 /* Taking the address of a non-addressable field is verboten. */
3785 if (by_ref && repr->non_addressable)
3786 return 0;
3787
3788 if (!by_ref || (!repr->grp_maybe_modified
3789 && !repr->grp_not_necessarilly_dereferenced))
3790 total_size += repr->size;
3791 else
3792 total_size += cur_parm_size;
3793
3794 new_param_count++;
3795 }
3796
3797 gcc_assert (new_param_count > 0);
3798
3799 if (optimize_function_for_size_p (cfun))
3800 parm_size_limit = cur_parm_size;
3801 else
3802 parm_size_limit = (PARAM_VALUE (PARAM_IPA_SRA_PTR_GROWTH_FACTOR)
3803 * cur_parm_size);
3804
3805 if (total_size < agg_size
3806 && total_size <= parm_size_limit)
3807 {
3808 if (dump_file)
3809 fprintf (dump_file, " ....will be split into %i components\n",
3810 new_param_count);
3811 return new_param_count;
3812 }
3813 else
3814 return 0;
3815 }
3816
3817 /* The order of the following enums is important, we need to do extra work for
3818 UNUSED_PARAMS, BY_VAL_ACCESSES and UNMODIF_BY_REF_ACCESSES. */
3819 enum ipa_splicing_result { NO_GOOD_ACCESS, UNUSED_PARAMS, BY_VAL_ACCESSES,
3820 MODIF_BY_REF_ACCESSES, UNMODIF_BY_REF_ACCESSES };
3821
3822 /* Identify representatives of all accesses to all candidate parameters for
3823 IPA-SRA. Return result based on what representatives have been found. */
3824
3825 static enum ipa_splicing_result
3826 splice_all_param_accesses (VEC (access_p, heap) **representatives)
3827 {
3828 enum ipa_splicing_result result = NO_GOOD_ACCESS;
3829 tree parm;
3830 struct access *repr;
3831
3832 *representatives = VEC_alloc (access_p, heap, func_param_count);
3833
3834 for (parm = DECL_ARGUMENTS (current_function_decl);
3835 parm;
3836 parm = DECL_CHAIN (parm))
3837 {
3838 if (is_unused_scalar_param (parm))
3839 {
3840 VEC_quick_push (access_p, *representatives,
3841 &no_accesses_representant);
3842 if (result == NO_GOOD_ACCESS)
3843 result = UNUSED_PARAMS;
3844 }
3845 else if (POINTER_TYPE_P (TREE_TYPE (parm))
3846 && is_gimple_reg_type (TREE_TYPE (TREE_TYPE (parm)))
3847 && bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3848 {
3849 repr = unmodified_by_ref_scalar_representative (parm);
3850 VEC_quick_push (access_p, *representatives, repr);
3851 if (repr)
3852 result = UNMODIF_BY_REF_ACCESSES;
3853 }
3854 else if (bitmap_bit_p (candidate_bitmap, DECL_UID (parm)))
3855 {
3856 bool ro_grp = false;
3857 repr = splice_param_accesses (parm, &ro_grp);
3858 VEC_quick_push (access_p, *representatives, repr);
3859
3860 if (repr && !no_accesses_p (repr))
3861 {
3862 if (POINTER_TYPE_P (TREE_TYPE (parm)))
3863 {
3864 if (ro_grp)
3865 result = UNMODIF_BY_REF_ACCESSES;
3866 else if (result < MODIF_BY_REF_ACCESSES)
3867 result = MODIF_BY_REF_ACCESSES;
3868 }
3869 else if (result < BY_VAL_ACCESSES)
3870 result = BY_VAL_ACCESSES;
3871 }
3872 else if (no_accesses_p (repr) && (result == NO_GOOD_ACCESS))
3873 result = UNUSED_PARAMS;
3874 }
3875 else
3876 VEC_quick_push (access_p, *representatives, NULL);
3877 }
3878
3879 if (result == NO_GOOD_ACCESS)
3880 {
3881 VEC_free (access_p, heap, *representatives);
3882 *representatives = NULL;
3883 return NO_GOOD_ACCESS;
3884 }
3885
3886 return result;
3887 }
3888
3889 /* Return the index of BASE in PARMS. Abort if it is not found. */
3890
3891 static inline int
3892 get_param_index (tree base, VEC(tree, heap) *parms)
3893 {
3894 int i, len;
3895
3896 len = VEC_length (tree, parms);
3897 for (i = 0; i < len; i++)
3898 if (VEC_index (tree, parms, i) == base)
3899 return i;
3900 gcc_unreachable ();
3901 }
3902
3903 /* Convert the decisions made at the representative level into compact
3904 parameter adjustments. REPRESENTATIVES are pointers to first
3905 representatives of each param accesses, ADJUSTMENTS_COUNT is the expected
3906 final number of adjustments. */
3907
3908 static ipa_parm_adjustment_vec
3909 turn_representatives_into_adjustments (VEC (access_p, heap) *representatives,
3910 int adjustments_count)
3911 {
3912 VEC (tree, heap) *parms;
3913 ipa_parm_adjustment_vec adjustments;
3914 tree parm;
3915 int i;
3916
3917 gcc_assert (adjustments_count > 0);
3918 parms = ipa_get_vector_of_formal_parms (current_function_decl);
3919 adjustments = VEC_alloc (ipa_parm_adjustment_t, heap, adjustments_count);
3920 parm = DECL_ARGUMENTS (current_function_decl);
3921 for (i = 0; i < func_param_count; i++, parm = DECL_CHAIN (parm))
3922 {
3923 struct access *repr = VEC_index (access_p, representatives, i);
3924
3925 if (!repr || no_accesses_p (repr))
3926 {
3927 struct ipa_parm_adjustment *adj;
3928
3929 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3930 memset (adj, 0, sizeof (*adj));
3931 adj->base_index = get_param_index (parm, parms);
3932 adj->base = parm;
3933 if (!repr)
3934 adj->copy_param = 1;
3935 else
3936 adj->remove_param = 1;
3937 }
3938 else
3939 {
3940 struct ipa_parm_adjustment *adj;
3941 int index = get_param_index (parm, parms);
3942
3943 for (; repr; repr = repr->next_grp)
3944 {
3945 adj = VEC_quick_push (ipa_parm_adjustment_t, adjustments, NULL);
3946 memset (adj, 0, sizeof (*adj));
3947 gcc_assert (repr->base == parm);
3948 adj->base_index = index;
3949 adj->base = repr->base;
3950 adj->type = repr->type;
3951 adj->alias_ptr_type = reference_alias_ptr_type (repr->expr);
3952 adj->offset = repr->offset;
3953 adj->by_ref = (POINTER_TYPE_P (TREE_TYPE (repr->base))
3954 && (repr->grp_maybe_modified
3955 || repr->grp_not_necessarilly_dereferenced));
3956
3957 }
3958 }
3959 }
3960 VEC_free (tree, heap, parms);
3961 return adjustments;
3962 }
3963
3964 /* Analyze the collected accesses and produce a plan what to do with the
3965 parameters in the form of adjustments, NULL meaning nothing. */
3966
3967 static ipa_parm_adjustment_vec
3968 analyze_all_param_acesses (void)
3969 {
3970 enum ipa_splicing_result repr_state;
3971 bool proceed = false;
3972 int i, adjustments_count = 0;
3973 VEC (access_p, heap) *representatives;
3974 ipa_parm_adjustment_vec adjustments;
3975
3976 repr_state = splice_all_param_accesses (&representatives);
3977 if (repr_state == NO_GOOD_ACCESS)
3978 return NULL;
3979
3980 /* If there are any parameters passed by reference which are not modified
3981 directly, we need to check whether they can be modified indirectly. */
3982 if (repr_state == UNMODIF_BY_REF_ACCESSES)
3983 {
3984 analyze_caller_dereference_legality (representatives);
3985 analyze_modified_params (representatives);
3986 }
3987
3988 for (i = 0; i < func_param_count; i++)
3989 {
3990 struct access *repr = VEC_index (access_p, representatives, i);
3991
3992 if (repr && !no_accesses_p (repr))
3993 {
3994 if (repr->grp_scalar_ptr)
3995 {
3996 adjustments_count++;
3997 if (repr->grp_not_necessarilly_dereferenced
3998 || repr->grp_maybe_modified)
3999 VEC_replace (access_p, representatives, i, NULL);
4000 else
4001 {
4002 proceed = true;
4003 sra_stats.scalar_by_ref_to_by_val++;
4004 }
4005 }
4006 else
4007 {
4008 int new_components = decide_one_param_reduction (repr);
4009
4010 if (new_components == 0)
4011 {
4012 VEC_replace (access_p, representatives, i, NULL);
4013 adjustments_count++;
4014 }
4015 else
4016 {
4017 adjustments_count += new_components;
4018 sra_stats.aggregate_params_reduced++;
4019 sra_stats.param_reductions_created += new_components;
4020 proceed = true;
4021 }
4022 }
4023 }
4024 else
4025 {
4026 if (no_accesses_p (repr))
4027 {
4028 proceed = true;
4029 sra_stats.deleted_unused_parameters++;
4030 }
4031 adjustments_count++;
4032 }
4033 }
4034
4035 if (!proceed && dump_file)
4036 fprintf (dump_file, "NOT proceeding to change params.\n");
4037
4038 if (proceed)
4039 adjustments = turn_representatives_into_adjustments (representatives,
4040 adjustments_count);
4041 else
4042 adjustments = NULL;
4043
4044 VEC_free (access_p, heap, representatives);
4045 return adjustments;
4046 }
4047
4048 /* If a parameter replacement identified by ADJ does not yet exist in the form
4049 of declaration, create it and record it, otherwise return the previously
4050 created one. */
4051
4052 static tree
4053 get_replaced_param_substitute (struct ipa_parm_adjustment *adj)
4054 {
4055 tree repl;
4056 if (!adj->new_ssa_base)
4057 {
4058 char *pretty_name = make_fancy_name (adj->base);
4059
4060 repl = create_tmp_reg (TREE_TYPE (adj->base), "ISR");
4061 DECL_NAME (repl) = get_identifier (pretty_name);
4062 obstack_free (&name_obstack, pretty_name);
4063
4064 get_var_ann (repl);
4065 add_referenced_var (repl);
4066 adj->new_ssa_base = repl;
4067 }
4068 else
4069 repl = adj->new_ssa_base;
4070 return repl;
4071 }
4072
4073 /* Find the first adjustment for a particular parameter BASE in a vector of
4074 ADJUSTMENTS which is not a copy_param. Return NULL if there is no such
4075 adjustment. */
4076
4077 static struct ipa_parm_adjustment *
4078 get_adjustment_for_base (ipa_parm_adjustment_vec adjustments, tree base)
4079 {
4080 int i, len;
4081
4082 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4083 for (i = 0; i < len; i++)
4084 {
4085 struct ipa_parm_adjustment *adj;
4086
4087 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4088 if (!adj->copy_param && adj->base == base)
4089 return adj;
4090 }
4091
4092 return NULL;
4093 }
4094
4095 /* If the statement STMT defines an SSA_NAME of a parameter which is to be
4096 removed because its value is not used, replace the SSA_NAME with a one
4097 relating to a created VAR_DECL together all of its uses and return true.
4098 ADJUSTMENTS is a pointer to an adjustments vector. */
4099
4100 static bool
4101 replace_removed_params_ssa_names (gimple stmt,
4102 ipa_parm_adjustment_vec adjustments)
4103 {
4104 struct ipa_parm_adjustment *adj;
4105 tree lhs, decl, repl, name;
4106
4107 if (gimple_code (stmt) == GIMPLE_PHI)
4108 lhs = gimple_phi_result (stmt);
4109 else if (is_gimple_assign (stmt))
4110 lhs = gimple_assign_lhs (stmt);
4111 else if (is_gimple_call (stmt))
4112 lhs = gimple_call_lhs (stmt);
4113 else
4114 gcc_unreachable ();
4115
4116 if (TREE_CODE (lhs) != SSA_NAME)
4117 return false;
4118 decl = SSA_NAME_VAR (lhs);
4119 if (TREE_CODE (decl) != PARM_DECL)
4120 return false;
4121
4122 adj = get_adjustment_for_base (adjustments, decl);
4123 if (!adj)
4124 return false;
4125
4126 repl = get_replaced_param_substitute (adj);
4127 name = make_ssa_name (repl, stmt);
4128
4129 if (dump_file)
4130 {
4131 fprintf (dump_file, "replacing an SSA name of a removed param ");
4132 print_generic_expr (dump_file, lhs, 0);
4133 fprintf (dump_file, " with ");
4134 print_generic_expr (dump_file, name, 0);
4135 fprintf (dump_file, "\n");
4136 }
4137
4138 if (is_gimple_assign (stmt))
4139 gimple_assign_set_lhs (stmt, name);
4140 else if (is_gimple_call (stmt))
4141 gimple_call_set_lhs (stmt, name);
4142 else
4143 gimple_phi_set_result (stmt, name);
4144
4145 replace_uses_by (lhs, name);
4146 release_ssa_name (lhs);
4147 return true;
4148 }
4149
4150 /* If the expression *EXPR should be replaced by a reduction of a parameter, do
4151 so. ADJUSTMENTS is a pointer to a vector of adjustments. CONVERT
4152 specifies whether the function should care about type incompatibility the
4153 current and new expressions. If it is false, the function will leave
4154 incompatibility issues to the caller. Return true iff the expression
4155 was modified. */
4156
4157 static bool
4158 sra_ipa_modify_expr (tree *expr, bool convert,
4159 ipa_parm_adjustment_vec adjustments)
4160 {
4161 int i, len;
4162 struct ipa_parm_adjustment *adj, *cand = NULL;
4163 HOST_WIDE_INT offset, size, max_size;
4164 tree base, src;
4165
4166 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4167
4168 if (TREE_CODE (*expr) == BIT_FIELD_REF
4169 || TREE_CODE (*expr) == IMAGPART_EXPR
4170 || TREE_CODE (*expr) == REALPART_EXPR)
4171 {
4172 expr = &TREE_OPERAND (*expr, 0);
4173 convert = true;
4174 }
4175
4176 base = get_ref_base_and_extent (*expr, &offset, &size, &max_size);
4177 if (!base || size == -1 || max_size == -1)
4178 return false;
4179
4180 if (TREE_CODE (base) == MEM_REF)
4181 {
4182 offset += mem_ref_offset (base).low * BITS_PER_UNIT;
4183 base = TREE_OPERAND (base, 0);
4184 }
4185
4186 base = get_ssa_base_param (base);
4187 if (!base || TREE_CODE (base) != PARM_DECL)
4188 return false;
4189
4190 for (i = 0; i < len; i++)
4191 {
4192 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4193
4194 if (adj->base == base &&
4195 (adj->offset == offset || adj->remove_param))
4196 {
4197 cand = adj;
4198 break;
4199 }
4200 }
4201 if (!cand || cand->copy_param || cand->remove_param)
4202 return false;
4203
4204 if (cand->by_ref)
4205 src = build_simple_mem_ref (cand->reduction);
4206 else
4207 src = cand->reduction;
4208
4209 if (dump_file && (dump_flags & TDF_DETAILS))
4210 {
4211 fprintf (dump_file, "About to replace expr ");
4212 print_generic_expr (dump_file, *expr, 0);
4213 fprintf (dump_file, " with ");
4214 print_generic_expr (dump_file, src, 0);
4215 fprintf (dump_file, "\n");
4216 }
4217
4218 if (convert && !useless_type_conversion_p (TREE_TYPE (*expr), cand->type))
4219 {
4220 tree vce = build1 (VIEW_CONVERT_EXPR, TREE_TYPE (*expr), src);
4221 *expr = vce;
4222 }
4223 else
4224 *expr = src;
4225 return true;
4226 }
4227
4228 /* If the statement pointed to by STMT_PTR contains any expressions that need
4229 to replaced with a different one as noted by ADJUSTMENTS, do so. Handle any
4230 potential type incompatibilities (GSI is used to accommodate conversion
4231 statements and must point to the statement). Return true iff the statement
4232 was modified. */
4233
4234 static bool
4235 sra_ipa_modify_assign (gimple *stmt_ptr, gimple_stmt_iterator *gsi,
4236 ipa_parm_adjustment_vec adjustments)
4237 {
4238 gimple stmt = *stmt_ptr;
4239 tree *lhs_p, *rhs_p;
4240 bool any;
4241
4242 if (!gimple_assign_single_p (stmt))
4243 return false;
4244
4245 rhs_p = gimple_assign_rhs1_ptr (stmt);
4246 lhs_p = gimple_assign_lhs_ptr (stmt);
4247
4248 any = sra_ipa_modify_expr (rhs_p, false, adjustments);
4249 any |= sra_ipa_modify_expr (lhs_p, false, adjustments);
4250 if (any)
4251 {
4252 tree new_rhs = NULL_TREE;
4253
4254 if (!useless_type_conversion_p (TREE_TYPE (*lhs_p), TREE_TYPE (*rhs_p)))
4255 {
4256 if (TREE_CODE (*rhs_p) == CONSTRUCTOR)
4257 {
4258 /* V_C_Es of constructors can cause trouble (PR 42714). */
4259 if (is_gimple_reg_type (TREE_TYPE (*lhs_p)))
4260 *rhs_p = build_zero_cst (TREE_TYPE (*lhs_p));
4261 else
4262 *rhs_p = build_constructor (TREE_TYPE (*lhs_p), 0);
4263 }
4264 else
4265 new_rhs = fold_build1_loc (gimple_location (stmt),
4266 VIEW_CONVERT_EXPR, TREE_TYPE (*lhs_p),
4267 *rhs_p);
4268 }
4269 else if (REFERENCE_CLASS_P (*rhs_p)
4270 && is_gimple_reg_type (TREE_TYPE (*lhs_p))
4271 && !is_gimple_reg (*lhs_p))
4272 /* This can happen when an assignment in between two single field
4273 structures is turned into an assignment in between two pointers to
4274 scalars (PR 42237). */
4275 new_rhs = *rhs_p;
4276
4277 if (new_rhs)
4278 {
4279 tree tmp = force_gimple_operand_gsi (gsi, new_rhs, true, NULL_TREE,
4280 true, GSI_SAME_STMT);
4281
4282 gimple_assign_set_rhs_from_tree (gsi, tmp);
4283 }
4284
4285 return true;
4286 }
4287
4288 return false;
4289 }
4290
4291 /* Traverse the function body and all modifications as described in
4292 ADJUSTMENTS. Return true iff the CFG has been changed. */
4293
4294 static bool
4295 ipa_sra_modify_function_body (ipa_parm_adjustment_vec adjustments)
4296 {
4297 bool cfg_changed = false;
4298 basic_block bb;
4299
4300 FOR_EACH_BB (bb)
4301 {
4302 gimple_stmt_iterator gsi;
4303
4304 for (gsi = gsi_start_phis (bb); !gsi_end_p (gsi); gsi_next (&gsi))
4305 replace_removed_params_ssa_names (gsi_stmt (gsi), adjustments);
4306
4307 gsi = gsi_start_bb (bb);
4308 while (!gsi_end_p (gsi))
4309 {
4310 gimple stmt = gsi_stmt (gsi);
4311 bool modified = false;
4312 tree *t;
4313 unsigned i;
4314
4315 switch (gimple_code (stmt))
4316 {
4317 case GIMPLE_RETURN:
4318 t = gimple_return_retval_ptr (stmt);
4319 if (*t != NULL_TREE)
4320 modified |= sra_ipa_modify_expr (t, true, adjustments);
4321 break;
4322
4323 case GIMPLE_ASSIGN:
4324 modified |= sra_ipa_modify_assign (&stmt, &gsi, adjustments);
4325 modified |= replace_removed_params_ssa_names (stmt, adjustments);
4326 break;
4327
4328 case GIMPLE_CALL:
4329 /* Operands must be processed before the lhs. */
4330 for (i = 0; i < gimple_call_num_args (stmt); i++)
4331 {
4332 t = gimple_call_arg_ptr (stmt, i);
4333 modified |= sra_ipa_modify_expr (t, true, adjustments);
4334 }
4335
4336 if (gimple_call_lhs (stmt))
4337 {
4338 t = gimple_call_lhs_ptr (stmt);
4339 modified |= sra_ipa_modify_expr (t, false, adjustments);
4340 modified |= replace_removed_params_ssa_names (stmt,
4341 adjustments);
4342 }
4343 break;
4344
4345 case GIMPLE_ASM:
4346 for (i = 0; i < gimple_asm_ninputs (stmt); i++)
4347 {
4348 t = &TREE_VALUE (gimple_asm_input_op (stmt, i));
4349 modified |= sra_ipa_modify_expr (t, true, adjustments);
4350 }
4351 for (i = 0; i < gimple_asm_noutputs (stmt); i++)
4352 {
4353 t = &TREE_VALUE (gimple_asm_output_op (stmt, i));
4354 modified |= sra_ipa_modify_expr (t, false, adjustments);
4355 }
4356 break;
4357
4358 default:
4359 break;
4360 }
4361
4362 if (modified)
4363 {
4364 update_stmt (stmt);
4365 if (maybe_clean_eh_stmt (stmt)
4366 && gimple_purge_dead_eh_edges (gimple_bb (stmt)))
4367 cfg_changed = true;
4368 }
4369 gsi_next (&gsi);
4370 }
4371 }
4372
4373 return cfg_changed;
4374 }
4375
4376 /* Call gimple_debug_bind_reset_value on all debug statements describing
4377 gimple register parameters that are being removed or replaced. */
4378
4379 static void
4380 sra_ipa_reset_debug_stmts (ipa_parm_adjustment_vec adjustments)
4381 {
4382 int i, len;
4383 gimple_stmt_iterator *gsip = NULL, gsi;
4384
4385 if (MAY_HAVE_DEBUG_STMTS && single_succ_p (ENTRY_BLOCK_PTR))
4386 {
4387 gsi = gsi_after_labels (single_succ (ENTRY_BLOCK_PTR));
4388 gsip = &gsi;
4389 }
4390 len = VEC_length (ipa_parm_adjustment_t, adjustments);
4391 for (i = 0; i < len; i++)
4392 {
4393 struct ipa_parm_adjustment *adj;
4394 imm_use_iterator ui;
4395 gimple stmt, def_temp;
4396 tree name, vexpr, copy = NULL_TREE;
4397 use_operand_p use_p;
4398
4399 adj = VEC_index (ipa_parm_adjustment_t, adjustments, i);
4400 if (adj->copy_param || !is_gimple_reg (adj->base))
4401 continue;
4402 name = gimple_default_def (cfun, adj->base);
4403 vexpr = NULL;
4404 if (name)
4405 FOR_EACH_IMM_USE_STMT (stmt, ui, name)
4406 {
4407 /* All other users must have been removed by
4408 ipa_sra_modify_function_body. */
4409 gcc_assert (is_gimple_debug (stmt));
4410 if (vexpr == NULL && gsip != NULL)
4411 {
4412 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4413 vexpr = make_node (DEBUG_EXPR_DECL);
4414 def_temp = gimple_build_debug_source_bind (vexpr, adj->base,
4415 NULL);
4416 DECL_ARTIFICIAL (vexpr) = 1;
4417 TREE_TYPE (vexpr) = TREE_TYPE (name);
4418 DECL_MODE (vexpr) = DECL_MODE (adj->base);
4419 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4420 }
4421 if (vexpr)
4422 {
4423 FOR_EACH_IMM_USE_ON_STMT (use_p, ui)
4424 SET_USE (use_p, vexpr);
4425 }
4426 else
4427 gimple_debug_bind_reset_value (stmt);
4428 update_stmt (stmt);
4429 }
4430 /* Create a VAR_DECL for debug info purposes. */
4431 if (!DECL_IGNORED_P (adj->base))
4432 {
4433 copy = build_decl (DECL_SOURCE_LOCATION (current_function_decl),
4434 VAR_DECL, DECL_NAME (adj->base),
4435 TREE_TYPE (adj->base));
4436 if (DECL_PT_UID_SET_P (adj->base))
4437 SET_DECL_PT_UID (copy, DECL_PT_UID (adj->base));
4438 TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (adj->base);
4439 TREE_READONLY (copy) = TREE_READONLY (adj->base);
4440 TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (adj->base);
4441 DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (adj->base);
4442 DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (adj->base);
4443 DECL_IGNORED_P (copy) = DECL_IGNORED_P (adj->base);
4444 DECL_ABSTRACT_ORIGIN (copy) = DECL_ORIGIN (adj->base);
4445 DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
4446 SET_DECL_RTL (copy, 0);
4447 TREE_USED (copy) = 1;
4448 DECL_CONTEXT (copy) = current_function_decl;
4449 add_referenced_var (copy);
4450 add_local_decl (cfun, copy);
4451 DECL_CHAIN (copy) =
4452 BLOCK_VARS (DECL_INITIAL (current_function_decl));
4453 BLOCK_VARS (DECL_INITIAL (current_function_decl)) = copy;
4454 }
4455 if (gsip != NULL && copy && target_for_debug_bind (adj->base))
4456 {
4457 gcc_assert (TREE_CODE (adj->base) == PARM_DECL);
4458 if (vexpr)
4459 def_temp = gimple_build_debug_bind (copy, vexpr, NULL);
4460 else
4461 def_temp = gimple_build_debug_source_bind (copy, adj->base,
4462 NULL);
4463 gsi_insert_before (gsip, def_temp, GSI_SAME_STMT);
4464 }
4465 }
4466 }
4467
4468 /* Return false iff all callers have at least as many actual arguments as there
4469 are formal parameters in the current function. */
4470
4471 static bool
4472 not_all_callers_have_enough_arguments_p (struct cgraph_node *node,
4473 void *data ATTRIBUTE_UNUSED)
4474 {
4475 struct cgraph_edge *cs;
4476 for (cs = node->callers; cs; cs = cs->next_caller)
4477 if (!callsite_has_enough_arguments_p (cs->call_stmt))
4478 return true;
4479
4480 return false;
4481 }
4482
4483 /* Convert all callers of NODE. */
4484
4485 static bool
4486 convert_callers_for_node (struct cgraph_node *node,
4487 void *data)
4488 {
4489 ipa_parm_adjustment_vec adjustments = (ipa_parm_adjustment_vec)data;
4490 bitmap recomputed_callers = BITMAP_ALLOC (NULL);
4491 struct cgraph_edge *cs;
4492
4493 for (cs = node->callers; cs; cs = cs->next_caller)
4494 {
4495 current_function_decl = cs->caller->decl;
4496 push_cfun (DECL_STRUCT_FUNCTION (cs->caller->decl));
4497
4498 if (dump_file)
4499 fprintf (dump_file, "Adjusting call (%i -> %i) %s -> %s\n",
4500 cs->caller->uid, cs->callee->uid,
4501 cgraph_node_name (cs->caller),
4502 cgraph_node_name (cs->callee));
4503
4504 ipa_modify_call_arguments (cs, cs->call_stmt, adjustments);
4505
4506 pop_cfun ();
4507 }
4508
4509 for (cs = node->callers; cs; cs = cs->next_caller)
4510 if (bitmap_set_bit (recomputed_callers, cs->caller->uid)
4511 && gimple_in_ssa_p (DECL_STRUCT_FUNCTION (cs->caller->decl)))
4512 compute_inline_parameters (cs->caller, true);
4513 BITMAP_FREE (recomputed_callers);
4514
4515 return true;
4516 }
4517
4518 /* Convert all callers of NODE to pass parameters as given in ADJUSTMENTS. */
4519
4520 static void
4521 convert_callers (struct cgraph_node *node, tree old_decl,
4522 ipa_parm_adjustment_vec adjustments)
4523 {
4524 tree old_cur_fndecl = current_function_decl;
4525 basic_block this_block;
4526
4527 cgraph_for_node_and_aliases (node, convert_callers_for_node,
4528 adjustments, false);
4529
4530 current_function_decl = old_cur_fndecl;
4531
4532 if (!encountered_recursive_call)
4533 return;
4534
4535 FOR_EACH_BB (this_block)
4536 {
4537 gimple_stmt_iterator gsi;
4538
4539 for (gsi = gsi_start_bb (this_block); !gsi_end_p (gsi); gsi_next (&gsi))
4540 {
4541 gimple stmt = gsi_stmt (gsi);
4542 tree call_fndecl;
4543 if (gimple_code (stmt) != GIMPLE_CALL)
4544 continue;
4545 call_fndecl = gimple_call_fndecl (stmt);
4546 if (call_fndecl == old_decl)
4547 {
4548 if (dump_file)
4549 fprintf (dump_file, "Adjusting recursive call");
4550 gimple_call_set_fndecl (stmt, node->decl);
4551 ipa_modify_call_arguments (NULL, stmt, adjustments);
4552 }
4553 }
4554 }
4555
4556 return;
4557 }
4558
4559 /* Perform all the modification required in IPA-SRA for NODE to have parameters
4560 as given in ADJUSTMENTS. Return true iff the CFG has been changed. */
4561
4562 static bool
4563 modify_function (struct cgraph_node *node, ipa_parm_adjustment_vec adjustments)
4564 {
4565 struct cgraph_node *new_node;
4566 bool cfg_changed;
4567 VEC (cgraph_edge_p, heap) * redirect_callers = collect_callers_of_node (node);
4568
4569 rebuild_cgraph_edges ();
4570 pop_cfun ();
4571 current_function_decl = NULL_TREE;
4572
4573 new_node = cgraph_function_versioning (node, redirect_callers, NULL, NULL,
4574 NULL, NULL, "isra");
4575 current_function_decl = new_node->decl;
4576 push_cfun (DECL_STRUCT_FUNCTION (new_node->decl));
4577
4578 ipa_modify_formal_parameters (current_function_decl, adjustments, "ISRA");
4579 cfg_changed = ipa_sra_modify_function_body (adjustments);
4580 sra_ipa_reset_debug_stmts (adjustments);
4581 convert_callers (new_node, node->decl, adjustments);
4582 cgraph_make_node_local (new_node);
4583 return cfg_changed;
4584 }
4585
4586 /* Return false the function is apparently unsuitable for IPA-SRA based on it's
4587 attributes, return true otherwise. NODE is the cgraph node of the current
4588 function. */
4589
4590 static bool
4591 ipa_sra_preliminary_function_checks (struct cgraph_node *node)
4592 {
4593 if (!cgraph_node_can_be_local_p (node))
4594 {
4595 if (dump_file)
4596 fprintf (dump_file, "Function not local to this compilation unit.\n");
4597 return false;
4598 }
4599
4600 if (!node->local.can_change_signature)
4601 {
4602 if (dump_file)
4603 fprintf (dump_file, "Function can not change signature.\n");
4604 return false;
4605 }
4606
4607 if (!tree_versionable_function_p (node->decl))
4608 {
4609 if (dump_file)
4610 fprintf (dump_file, "Function is not versionable.\n");
4611 return false;
4612 }
4613
4614 if (DECL_VIRTUAL_P (current_function_decl))
4615 {
4616 if (dump_file)
4617 fprintf (dump_file, "Function is a virtual method.\n");
4618 return false;
4619 }
4620
4621 if ((DECL_COMDAT (node->decl) || DECL_EXTERNAL (node->decl))
4622 && inline_summary(node)->size >= MAX_INLINE_INSNS_AUTO)
4623 {
4624 if (dump_file)
4625 fprintf (dump_file, "Function too big to be made truly local.\n");
4626 return false;
4627 }
4628
4629 if (!node->callers)
4630 {
4631 if (dump_file)
4632 fprintf (dump_file,
4633 "Function has no callers in this compilation unit.\n");
4634 return false;
4635 }
4636
4637 if (cfun->stdarg)
4638 {
4639 if (dump_file)
4640 fprintf (dump_file, "Function uses stdarg. \n");
4641 return false;
4642 }
4643
4644 if (TYPE_ATTRIBUTES (TREE_TYPE (node->decl)))
4645 return false;
4646
4647 return true;
4648 }
4649
4650 /* Perform early interprocedural SRA. */
4651
4652 static unsigned int
4653 ipa_early_sra (void)
4654 {
4655 struct cgraph_node *node = cgraph_get_node (current_function_decl);
4656 ipa_parm_adjustment_vec adjustments;
4657 int ret = 0;
4658
4659 if (!ipa_sra_preliminary_function_checks (node))
4660 return 0;
4661
4662 sra_initialize ();
4663 sra_mode = SRA_MODE_EARLY_IPA;
4664
4665 if (!find_param_candidates ())
4666 {
4667 if (dump_file)
4668 fprintf (dump_file, "Function has no IPA-SRA candidates.\n");
4669 goto simple_out;
4670 }
4671
4672 if (cgraph_for_node_and_aliases (node, not_all_callers_have_enough_arguments_p,
4673 NULL, true))
4674 {
4675 if (dump_file)
4676 fprintf (dump_file, "There are callers with insufficient number of "
4677 "arguments.\n");
4678 goto simple_out;
4679 }
4680
4681 bb_dereferences = XCNEWVEC (HOST_WIDE_INT,
4682 func_param_count
4683 * last_basic_block_for_function (cfun));
4684 final_bbs = BITMAP_ALLOC (NULL);
4685
4686 scan_function ();
4687 if (encountered_apply_args)
4688 {
4689 if (dump_file)
4690 fprintf (dump_file, "Function calls __builtin_apply_args().\n");
4691 goto out;
4692 }
4693
4694 if (encountered_unchangable_recursive_call)
4695 {
4696 if (dump_file)
4697 fprintf (dump_file, "Function calls itself with insufficient "
4698 "number of arguments.\n");
4699 goto out;
4700 }
4701
4702 adjustments = analyze_all_param_acesses ();
4703 if (!adjustments)
4704 goto out;
4705 if (dump_file)
4706 ipa_dump_param_adjustments (dump_file, adjustments, current_function_decl);
4707
4708 if (modify_function (node, adjustments))
4709 ret = TODO_update_ssa | TODO_cleanup_cfg;
4710 else
4711 ret = TODO_update_ssa;
4712 VEC_free (ipa_parm_adjustment_t, heap, adjustments);
4713
4714 statistics_counter_event (cfun, "Unused parameters deleted",
4715 sra_stats.deleted_unused_parameters);
4716 statistics_counter_event (cfun, "Scalar parameters converted to by-value",
4717 sra_stats.scalar_by_ref_to_by_val);
4718 statistics_counter_event (cfun, "Aggregate parameters broken up",
4719 sra_stats.aggregate_params_reduced);
4720 statistics_counter_event (cfun, "Aggregate parameter components created",
4721 sra_stats.param_reductions_created);
4722
4723 out:
4724 BITMAP_FREE (final_bbs);
4725 free (bb_dereferences);
4726 simple_out:
4727 sra_deinitialize ();
4728 return ret;
4729 }
4730
4731 /* Return if early ipa sra shall be performed. */
4732 static bool
4733 ipa_early_sra_gate (void)
4734 {
4735 return flag_ipa_sra && dbg_cnt (eipa_sra);
4736 }
4737
4738 struct gimple_opt_pass pass_early_ipa_sra =
4739 {
4740 {
4741 GIMPLE_PASS,
4742 "eipa_sra", /* name */
4743 ipa_early_sra_gate, /* gate */
4744 ipa_early_sra, /* execute */
4745 NULL, /* sub */
4746 NULL, /* next */
4747 0, /* static_pass_number */
4748 TV_IPA_SRA, /* tv_id */
4749 0, /* properties_required */
4750 0, /* properties_provided */
4751 0, /* properties_destroyed */
4752 0, /* todo_flags_start */
4753 TODO_dump_cgraph /* todo_flags_finish */
4754 }
4755 };