tree-sra.c (sra_build_assignment): Disable assertion checking for now.
[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) 2003, 2004, 2005, 2006, 2007
5 Free Software Foundation, Inc.
6 Contributed by Diego Novillo <dnovillo@redhat.com>
7
8 This file is part of GCC.
9
10 GCC is free software; you can redistribute it and/or modify it
11 under the terms of the GNU General Public License as published by the
12 Free Software Foundation; either version 2, or (at your option) any
13 later version.
14
15 GCC is distributed in the hope that it will be useful, but WITHOUT
16 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
17 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
18 for more details.
19
20 You should have received a copy of the GNU General Public License
21 along with GCC; see the file COPYING. If not, write to the Free
22 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
23 02110-1301, USA. */
24
25 #include "config.h"
26 #include "system.h"
27 #include "coretypes.h"
28 #include "tm.h"
29 #include "ggc.h"
30 #include "tree.h"
31
32 /* These RTL headers are needed for basic-block.h. */
33 #include "rtl.h"
34 #include "tm_p.h"
35 #include "hard-reg-set.h"
36 #include "basic-block.h"
37 #include "diagnostic.h"
38 #include "langhooks.h"
39 #include "tree-inline.h"
40 #include "tree-flow.h"
41 #include "tree-gimple.h"
42 #include "tree-dump.h"
43 #include "tree-pass.h"
44 #include "timevar.h"
45 #include "flags.h"
46 #include "bitmap.h"
47 #include "obstack.h"
48 #include "target.h"
49 /* expr.h is needed for MOVE_RATIO. */
50 #include "expr.h"
51 #include "params.h"
52
53
54 /* This object of this pass is to replace a non-addressable aggregate with a
55 set of independent variables. Most of the time, all of these variables
56 will be scalars. But a secondary objective is to break up larger
57 aggregates into smaller aggregates. In the process we may find that some
58 bits of the larger aggregate can be deleted as unreferenced.
59
60 This substitution is done globally. More localized substitutions would
61 be the purvey of a load-store motion pass.
62
63 The optimization proceeds in phases:
64
65 (1) Identify variables that have types that are candidates for
66 decomposition.
67
68 (2) Scan the function looking for the ways these variables are used.
69 In particular we're interested in the number of times a variable
70 (or member) is needed as a complete unit, and the number of times
71 a variable (or member) is copied.
72
73 (3) Based on the usage profile, instantiate substitution variables.
74
75 (4) Scan the function making replacements.
76 */
77
78
79 /* True if this is the "early" pass, before inlining. */
80 static bool early_sra;
81
82 /* The set of todo flags to return from tree_sra. */
83 static unsigned int todoflags;
84
85 /* The set of aggregate variables that are candidates for scalarization. */
86 static bitmap sra_candidates;
87
88 /* Set of scalarizable PARM_DECLs that need copy-in operations at the
89 beginning of the function. */
90 static bitmap needs_copy_in;
91
92 /* Sets of bit pairs that cache type decomposition and instantiation. */
93 static bitmap sra_type_decomp_cache;
94 static bitmap sra_type_inst_cache;
95
96 /* One of these structures is created for each candidate aggregate and
97 each (accessed) member or group of members of such an aggregate. */
98 struct sra_elt
99 {
100 /* A tree of the elements. Used when we want to traverse everything. */
101 struct sra_elt *parent;
102 struct sra_elt *groups;
103 struct sra_elt *children;
104 struct sra_elt *sibling;
105
106 /* If this element is a root, then this is the VAR_DECL. If this is
107 a sub-element, this is some token used to identify the reference.
108 In the case of COMPONENT_REF, this is the FIELD_DECL. In the case
109 of an ARRAY_REF, this is the (constant) index. In the case of an
110 ARRAY_RANGE_REF, this is the (constant) RANGE_EXPR. In the case
111 of a complex number, this is a zero or one. */
112 tree element;
113
114 /* The type of the element. */
115 tree type;
116
117 /* A VAR_DECL, for any sub-element we've decided to replace. */
118 tree replacement;
119
120 /* The number of times the element is referenced as a whole. I.e.
121 given "a.b.c", this would be incremented for C, but not for A or B. */
122 unsigned int n_uses;
123
124 /* The number of times the element is copied to or from another
125 scalarizable element. */
126 unsigned int n_copies;
127
128 /* True if TYPE is scalar. */
129 bool is_scalar;
130
131 /* True if this element is a group of members of its parent. */
132 bool is_group;
133
134 /* True if we saw something about this element that prevents scalarization,
135 such as non-constant indexing. */
136 bool cannot_scalarize;
137
138 /* True if we've decided that structure-to-structure assignment
139 should happen via memcpy and not per-element. */
140 bool use_block_copy;
141
142 /* True if everything under this element has been marked TREE_NO_WARNING. */
143 bool all_no_warning;
144
145 /* A flag for use with/after random access traversals. */
146 bool visited;
147
148 /* True if there is BIT_FIELD_REF on the lhs with a vector. */
149 bool is_vector_lhs;
150 };
151
152 #define IS_ELEMENT_FOR_GROUP(ELEMENT) (TREE_CODE (ELEMENT) == RANGE_EXPR)
153
154 #define FOR_EACH_ACTUAL_CHILD(CHILD, ELT) \
155 for ((CHILD) = (ELT)->is_group \
156 ? next_child_for_group (NULL, (ELT)) \
157 : (ELT)->children; \
158 (CHILD); \
159 (CHILD) = (ELT)->is_group \
160 ? next_child_for_group ((CHILD), (ELT)) \
161 : (CHILD)->sibling)
162
163 /* Helper function for above macro. Return next child in group. */
164 static struct sra_elt *
165 next_child_for_group (struct sra_elt *child, struct sra_elt *group)
166 {
167 gcc_assert (group->is_group);
168
169 /* Find the next child in the parent. */
170 if (child)
171 child = child->sibling;
172 else
173 child = group->parent->children;
174
175 /* Skip siblings that do not belong to the group. */
176 while (child)
177 {
178 tree g_elt = group->element;
179 if (TREE_CODE (g_elt) == RANGE_EXPR)
180 {
181 if (!tree_int_cst_lt (child->element, TREE_OPERAND (g_elt, 0))
182 && !tree_int_cst_lt (TREE_OPERAND (g_elt, 1), child->element))
183 break;
184 }
185 else
186 gcc_unreachable ();
187
188 child = child->sibling;
189 }
190
191 return child;
192 }
193
194 /* Random access to the child of a parent is performed by hashing.
195 This prevents quadratic behavior, and allows SRA to function
196 reasonably on larger records. */
197 static htab_t sra_map;
198
199 /* All structures are allocated out of the following obstack. */
200 static struct obstack sra_obstack;
201
202 /* Debugging functions. */
203 static void dump_sra_elt_name (FILE *, struct sra_elt *);
204 extern void debug_sra_elt_name (struct sra_elt *);
205
206 /* Forward declarations. */
207 static tree generate_element_ref (struct sra_elt *);
208 \f
209 /* Return true if DECL is an SRA candidate. */
210
211 static bool
212 is_sra_candidate_decl (tree decl)
213 {
214 return DECL_P (decl) && bitmap_bit_p (sra_candidates, DECL_UID (decl));
215 }
216
217 /* Return true if TYPE is a scalar type. */
218
219 static bool
220 is_sra_scalar_type (tree type)
221 {
222 enum tree_code code = TREE_CODE (type);
223 return (code == INTEGER_TYPE || code == REAL_TYPE || code == VECTOR_TYPE
224 || code == ENUMERAL_TYPE || code == BOOLEAN_TYPE
225 || code == POINTER_TYPE || code == OFFSET_TYPE
226 || code == REFERENCE_TYPE);
227 }
228
229 /* Return true if TYPE can be decomposed into a set of independent variables.
230
231 Note that this doesn't imply that all elements of TYPE can be
232 instantiated, just that if we decide to break up the type into
233 separate pieces that it can be done. */
234
235 bool
236 sra_type_can_be_decomposed_p (tree type)
237 {
238 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
239 tree t;
240
241 /* Avoid searching the same type twice. */
242 if (bitmap_bit_p (sra_type_decomp_cache, cache+0))
243 return true;
244 if (bitmap_bit_p (sra_type_decomp_cache, cache+1))
245 return false;
246
247 /* The type must have a definite nonzero size. */
248 if (TYPE_SIZE (type) == NULL || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST
249 || integer_zerop (TYPE_SIZE (type)))
250 goto fail;
251
252 /* The type must be a non-union aggregate. */
253 switch (TREE_CODE (type))
254 {
255 case RECORD_TYPE:
256 {
257 bool saw_one_field = false;
258
259 for (t = TYPE_FIELDS (type); t ; t = TREE_CHAIN (t))
260 if (TREE_CODE (t) == FIELD_DECL)
261 {
262 /* Reject incorrectly represented bit fields. */
263 if (DECL_BIT_FIELD (t)
264 && (tree_low_cst (DECL_SIZE (t), 1)
265 != TYPE_PRECISION (TREE_TYPE (t))))
266 goto fail;
267
268 saw_one_field = true;
269 }
270
271 /* Record types must have at least one field. */
272 if (!saw_one_field)
273 goto fail;
274 }
275 break;
276
277 case ARRAY_TYPE:
278 /* Array types must have a fixed lower and upper bound. */
279 t = TYPE_DOMAIN (type);
280 if (t == NULL)
281 goto fail;
282 if (TYPE_MIN_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MIN_VALUE (t)))
283 goto fail;
284 if (TYPE_MAX_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MAX_VALUE (t)))
285 goto fail;
286 break;
287
288 case COMPLEX_TYPE:
289 break;
290
291 default:
292 goto fail;
293 }
294
295 bitmap_set_bit (sra_type_decomp_cache, cache+0);
296 return true;
297
298 fail:
299 bitmap_set_bit (sra_type_decomp_cache, cache+1);
300 return false;
301 }
302
303 /* Return true if DECL can be decomposed into a set of independent
304 (though not necessarily scalar) variables. */
305
306 static bool
307 decl_can_be_decomposed_p (tree var)
308 {
309 /* Early out for scalars. */
310 if (is_sra_scalar_type (TREE_TYPE (var)))
311 return false;
312
313 /* The variable must not be aliased. */
314 if (!is_gimple_non_addressable (var))
315 {
316 if (dump_file && (dump_flags & TDF_DETAILS))
317 {
318 fprintf (dump_file, "Cannot scalarize variable ");
319 print_generic_expr (dump_file, var, dump_flags);
320 fprintf (dump_file, " because it must live in memory\n");
321 }
322 return false;
323 }
324
325 /* The variable must not be volatile. */
326 if (TREE_THIS_VOLATILE (var))
327 {
328 if (dump_file && (dump_flags & TDF_DETAILS))
329 {
330 fprintf (dump_file, "Cannot scalarize variable ");
331 print_generic_expr (dump_file, var, dump_flags);
332 fprintf (dump_file, " because it is declared volatile\n");
333 }
334 return false;
335 }
336
337 /* We must be able to decompose the variable's type. */
338 if (!sra_type_can_be_decomposed_p (TREE_TYPE (var)))
339 {
340 if (dump_file && (dump_flags & TDF_DETAILS))
341 {
342 fprintf (dump_file, "Cannot scalarize variable ");
343 print_generic_expr (dump_file, var, dump_flags);
344 fprintf (dump_file, " because its type cannot be decomposed\n");
345 }
346 return false;
347 }
348
349 /* HACK: if we decompose a va_list_type_node before inlining, then we'll
350 confuse tree-stdarg.c, and we won't be able to figure out which and
351 how many arguments are accessed. This really should be improved in
352 tree-stdarg.c, as the decomposition is truely a win. This could also
353 be fixed if the stdarg pass ran early, but this can't be done until
354 we've aliasing information early too. See PR 30791. */
355 if (early_sra
356 && TYPE_MAIN_VARIANT (TREE_TYPE (var))
357 == TYPE_MAIN_VARIANT (va_list_type_node))
358 return false;
359
360 return true;
361 }
362
363 /* Return true if TYPE can be *completely* decomposed into scalars. */
364
365 static bool
366 type_can_instantiate_all_elements (tree type)
367 {
368 if (is_sra_scalar_type (type))
369 return true;
370 if (!sra_type_can_be_decomposed_p (type))
371 return false;
372
373 switch (TREE_CODE (type))
374 {
375 case RECORD_TYPE:
376 {
377 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
378 tree f;
379
380 if (bitmap_bit_p (sra_type_inst_cache, cache+0))
381 return true;
382 if (bitmap_bit_p (sra_type_inst_cache, cache+1))
383 return false;
384
385 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
386 if (TREE_CODE (f) == FIELD_DECL)
387 {
388 if (!type_can_instantiate_all_elements (TREE_TYPE (f)))
389 {
390 bitmap_set_bit (sra_type_inst_cache, cache+1);
391 return false;
392 }
393 }
394
395 bitmap_set_bit (sra_type_inst_cache, cache+0);
396 return true;
397 }
398
399 case ARRAY_TYPE:
400 return type_can_instantiate_all_elements (TREE_TYPE (type));
401
402 case COMPLEX_TYPE:
403 return true;
404
405 default:
406 gcc_unreachable ();
407 }
408 }
409
410 /* Test whether ELT or some sub-element cannot be scalarized. */
411
412 static bool
413 can_completely_scalarize_p (struct sra_elt *elt)
414 {
415 struct sra_elt *c;
416
417 if (elt->cannot_scalarize)
418 return false;
419
420 for (c = elt->children; c; c = c->sibling)
421 if (!can_completely_scalarize_p (c))
422 return false;
423
424 for (c = elt->groups; c; c = c->sibling)
425 if (!can_completely_scalarize_p (c))
426 return false;
427
428 return true;
429 }
430
431 \f
432 /* A simplified tree hashing algorithm that only handles the types of
433 trees we expect to find in sra_elt->element. */
434
435 static hashval_t
436 sra_hash_tree (tree t)
437 {
438 hashval_t h;
439
440 switch (TREE_CODE (t))
441 {
442 case VAR_DECL:
443 case PARM_DECL:
444 case RESULT_DECL:
445 h = DECL_UID (t);
446 break;
447
448 case INTEGER_CST:
449 h = TREE_INT_CST_LOW (t) ^ TREE_INT_CST_HIGH (t);
450 break;
451
452 case RANGE_EXPR:
453 h = iterative_hash_expr (TREE_OPERAND (t, 0), 0);
454 h = iterative_hash_expr (TREE_OPERAND (t, 1), h);
455 break;
456
457 case FIELD_DECL:
458 /* We can have types that are compatible, but have different member
459 lists, so we can't hash fields by ID. Use offsets instead. */
460 h = iterative_hash_expr (DECL_FIELD_OFFSET (t), 0);
461 h = iterative_hash_expr (DECL_FIELD_BIT_OFFSET (t), h);
462 break;
463
464 default:
465 gcc_unreachable ();
466 }
467
468 return h;
469 }
470
471 /* Hash function for type SRA_PAIR. */
472
473 static hashval_t
474 sra_elt_hash (const void *x)
475 {
476 const struct sra_elt *e = x;
477 const struct sra_elt *p;
478 hashval_t h;
479
480 h = sra_hash_tree (e->element);
481
482 /* Take into account everything back up the chain. Given that chain
483 lengths are rarely very long, this should be acceptable. If we
484 truly identify this as a performance problem, it should work to
485 hash the pointer value "e->parent". */
486 for (p = e->parent; p ; p = p->parent)
487 h = (h * 65521) ^ sra_hash_tree (p->element);
488
489 return h;
490 }
491
492 /* Equality function for type SRA_PAIR. */
493
494 static int
495 sra_elt_eq (const void *x, const void *y)
496 {
497 const struct sra_elt *a = x;
498 const struct sra_elt *b = y;
499 tree ae, be;
500
501 if (a->parent != b->parent)
502 return false;
503
504 ae = a->element;
505 be = b->element;
506
507 if (ae == be)
508 return true;
509 if (TREE_CODE (ae) != TREE_CODE (be))
510 return false;
511
512 switch (TREE_CODE (ae))
513 {
514 case VAR_DECL:
515 case PARM_DECL:
516 case RESULT_DECL:
517 /* These are all pointer unique. */
518 return false;
519
520 case INTEGER_CST:
521 /* Integers are not pointer unique, so compare their values. */
522 return tree_int_cst_equal (ae, be);
523
524 case RANGE_EXPR:
525 return
526 tree_int_cst_equal (TREE_OPERAND (ae, 0), TREE_OPERAND (be, 0))
527 && tree_int_cst_equal (TREE_OPERAND (ae, 1), TREE_OPERAND (be, 1));
528
529 case FIELD_DECL:
530 /* Fields are unique within a record, but not between
531 compatible records. */
532 if (DECL_FIELD_CONTEXT (ae) == DECL_FIELD_CONTEXT (be))
533 return false;
534 return fields_compatible_p (ae, be);
535
536 default:
537 gcc_unreachable ();
538 }
539 }
540
541 /* Create or return the SRA_ELT structure for CHILD in PARENT. PARENT
542 may be null, in which case CHILD must be a DECL. */
543
544 static struct sra_elt *
545 lookup_element (struct sra_elt *parent, tree child, tree type,
546 enum insert_option insert)
547 {
548 struct sra_elt dummy;
549 struct sra_elt **slot;
550 struct sra_elt *elt;
551
552 if (parent)
553 dummy.parent = parent->is_group ? parent->parent : parent;
554 else
555 dummy.parent = NULL;
556 dummy.element = child;
557
558 slot = (struct sra_elt **) htab_find_slot (sra_map, &dummy, insert);
559 if (!slot && insert == NO_INSERT)
560 return NULL;
561
562 elt = *slot;
563 if (!elt && insert == INSERT)
564 {
565 *slot = elt = obstack_alloc (&sra_obstack, sizeof (*elt));
566 memset (elt, 0, sizeof (*elt));
567
568 elt->parent = parent;
569 elt->element = child;
570 elt->type = type;
571 elt->is_scalar = is_sra_scalar_type (type);
572
573 if (parent)
574 {
575 if (IS_ELEMENT_FOR_GROUP (elt->element))
576 {
577 elt->is_group = true;
578 elt->sibling = parent->groups;
579 parent->groups = elt;
580 }
581 else
582 {
583 elt->sibling = parent->children;
584 parent->children = elt;
585 }
586 }
587
588 /* If this is a parameter, then if we want to scalarize, we have
589 one copy from the true function parameter. Count it now. */
590 if (TREE_CODE (child) == PARM_DECL)
591 {
592 elt->n_copies = 1;
593 bitmap_set_bit (needs_copy_in, DECL_UID (child));
594 }
595 }
596
597 return elt;
598 }
599
600 /* Create or return the SRA_ELT structure for EXPR if the expression
601 refers to a scalarizable variable. */
602
603 static struct sra_elt *
604 maybe_lookup_element_for_expr (tree expr)
605 {
606 struct sra_elt *elt;
607 tree child;
608
609 switch (TREE_CODE (expr))
610 {
611 case VAR_DECL:
612 case PARM_DECL:
613 case RESULT_DECL:
614 if (is_sra_candidate_decl (expr))
615 return lookup_element (NULL, expr, TREE_TYPE (expr), INSERT);
616 return NULL;
617
618 case ARRAY_REF:
619 /* We can't scalarize variable array indices. */
620 if (in_array_bounds_p (expr))
621 child = TREE_OPERAND (expr, 1);
622 else
623 return NULL;
624 break;
625
626 case ARRAY_RANGE_REF:
627 /* We can't scalarize variable array indices. */
628 if (range_in_array_bounds_p (expr))
629 {
630 tree domain = TYPE_DOMAIN (TREE_TYPE (expr));
631 child = build2 (RANGE_EXPR, integer_type_node,
632 TYPE_MIN_VALUE (domain), TYPE_MAX_VALUE (domain));
633 }
634 else
635 return NULL;
636 break;
637
638 case COMPONENT_REF:
639 /* Don't look through unions. */
640 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (expr, 0))) != RECORD_TYPE)
641 return NULL;
642 child = TREE_OPERAND (expr, 1);
643 break;
644
645 case REALPART_EXPR:
646 child = integer_zero_node;
647 break;
648 case IMAGPART_EXPR:
649 child = integer_one_node;
650 break;
651
652 default:
653 return NULL;
654 }
655
656 elt = maybe_lookup_element_for_expr (TREE_OPERAND (expr, 0));
657 if (elt)
658 return lookup_element (elt, child, TREE_TYPE (expr), INSERT);
659 return NULL;
660 }
661
662 \f
663 /* Functions to walk just enough of the tree to see all scalarizable
664 references, and categorize them. */
665
666 /* A set of callbacks for phases 2 and 4. They'll be invoked for the
667 various kinds of references seen. In all cases, *BSI is an iterator
668 pointing to the statement being processed. */
669 struct sra_walk_fns
670 {
671 /* Invoked when ELT is required as a unit. Note that ELT might refer to
672 a leaf node, in which case this is a simple scalar reference. *EXPR_P
673 points to the location of the expression. IS_OUTPUT is true if this
674 is a left-hand-side reference. USE_ALL is true if we saw something we
675 couldn't quite identify and had to force the use of the entire object. */
676 void (*use) (struct sra_elt *elt, tree *expr_p,
677 block_stmt_iterator *bsi, bool is_output, bool use_all);
678
679 /* Invoked when we have a copy between two scalarizable references. */
680 void (*copy) (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
681 block_stmt_iterator *bsi);
682
683 /* Invoked when ELT is initialized from a constant. VALUE may be NULL,
684 in which case it should be treated as an empty CONSTRUCTOR. */
685 void (*init) (struct sra_elt *elt, tree value, block_stmt_iterator *bsi);
686
687 /* Invoked when we have a copy between one scalarizable reference ELT
688 and one non-scalarizable reference OTHER. IS_OUTPUT is true if ELT
689 is on the left-hand side. */
690 void (*ldst) (struct sra_elt *elt, tree other,
691 block_stmt_iterator *bsi, bool is_output);
692
693 /* True during phase 2, false during phase 4. */
694 /* ??? This is a hack. */
695 bool initial_scan;
696 };
697
698 #ifdef ENABLE_CHECKING
699 /* Invoked via walk_tree, if *TP contains a candidate decl, return it. */
700
701 static tree
702 sra_find_candidate_decl (tree *tp, int *walk_subtrees,
703 void *data ATTRIBUTE_UNUSED)
704 {
705 tree t = *tp;
706 enum tree_code code = TREE_CODE (t);
707
708 if (code == VAR_DECL || code == PARM_DECL || code == RESULT_DECL)
709 {
710 *walk_subtrees = 0;
711 if (is_sra_candidate_decl (t))
712 return t;
713 }
714 else if (TYPE_P (t))
715 *walk_subtrees = 0;
716
717 return NULL;
718 }
719 #endif
720
721 /* Walk most expressions looking for a scalarizable aggregate.
722 If we find one, invoke FNS->USE. */
723
724 static void
725 sra_walk_expr (tree *expr_p, block_stmt_iterator *bsi, bool is_output,
726 const struct sra_walk_fns *fns)
727 {
728 tree expr = *expr_p;
729 tree inner = expr;
730 bool disable_scalarization = false;
731 bool use_all_p = false;
732
733 /* We're looking to collect a reference expression between EXPR and INNER,
734 such that INNER is a scalarizable decl and all other nodes through EXPR
735 are references that we can scalarize. If we come across something that
736 we can't scalarize, we reset EXPR. This has the effect of making it
737 appear that we're referring to the larger expression as a whole. */
738
739 while (1)
740 switch (TREE_CODE (inner))
741 {
742 case VAR_DECL:
743 case PARM_DECL:
744 case RESULT_DECL:
745 /* If there is a scalarizable decl at the bottom, then process it. */
746 if (is_sra_candidate_decl (inner))
747 {
748 struct sra_elt *elt = maybe_lookup_element_for_expr (expr);
749 if (disable_scalarization)
750 elt->cannot_scalarize = true;
751 else
752 fns->use (elt, expr_p, bsi, is_output, use_all_p);
753 }
754 return;
755
756 case ARRAY_REF:
757 /* Non-constant index means any member may be accessed. Prevent the
758 expression from being scalarized. If we were to treat this as a
759 reference to the whole array, we can wind up with a single dynamic
760 index reference inside a loop being overridden by several constant
761 index references during loop setup. It's possible that this could
762 be avoided by using dynamic usage counts based on BB trip counts
763 (based on loop analysis or profiling), but that hardly seems worth
764 the effort. */
765 /* ??? Hack. Figure out how to push this into the scan routines
766 without duplicating too much code. */
767 if (!in_array_bounds_p (inner))
768 {
769 disable_scalarization = true;
770 goto use_all;
771 }
772 /* ??? Are we assured that non-constant bounds and stride will have
773 the same value everywhere? I don't think Fortran will... */
774 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
775 goto use_all;
776 inner = TREE_OPERAND (inner, 0);
777 break;
778
779 case ARRAY_RANGE_REF:
780 if (!range_in_array_bounds_p (inner))
781 {
782 disable_scalarization = true;
783 goto use_all;
784 }
785 /* ??? See above non-constant bounds and stride . */
786 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
787 goto use_all;
788 inner = TREE_OPERAND (inner, 0);
789 break;
790
791 case COMPONENT_REF:
792 /* A reference to a union member constitutes a reference to the
793 entire union. */
794 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) != RECORD_TYPE)
795 goto use_all;
796 /* ??? See above re non-constant stride. */
797 if (TREE_OPERAND (inner, 2))
798 goto use_all;
799 inner = TREE_OPERAND (inner, 0);
800 break;
801
802 case REALPART_EXPR:
803 case IMAGPART_EXPR:
804 inner = TREE_OPERAND (inner, 0);
805 break;
806
807 case BIT_FIELD_REF:
808 /* A bit field reference to a specific vector is scalarized but for
809 ones for inputs need to be marked as used on the left hand size so
810 when we scalarize it, we can mark that variable as non renamable. */
811 if (is_output
812 && TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) == VECTOR_TYPE)
813 {
814 struct sra_elt *elt
815 = maybe_lookup_element_for_expr (TREE_OPERAND (inner, 0));
816 if (elt)
817 elt->is_vector_lhs = true;
818 }
819 /* A bit field reference (access to *multiple* fields simultaneously)
820 is not currently scalarized. Consider this an access to the
821 complete outer element, to which walk_tree will bring us next. */
822
823 goto use_all;
824
825 case VIEW_CONVERT_EXPR:
826 case NOP_EXPR:
827 /* Similarly, a view/nop explicitly wants to look at an object in a
828 type other than the one we've scalarized. */
829 goto use_all;
830
831 case WITH_SIZE_EXPR:
832 /* This is a transparent wrapper. The entire inner expression really
833 is being used. */
834 goto use_all;
835
836 use_all:
837 expr_p = &TREE_OPERAND (inner, 0);
838 inner = expr = *expr_p;
839 use_all_p = true;
840 break;
841
842 default:
843 #ifdef ENABLE_CHECKING
844 /* Validate that we're not missing any references. */
845 gcc_assert (!walk_tree (&inner, sra_find_candidate_decl, NULL, NULL));
846 #endif
847 return;
848 }
849 }
850
851 /* Walk a TREE_LIST of values looking for scalarizable aggregates.
852 If we find one, invoke FNS->USE. */
853
854 static void
855 sra_walk_tree_list (tree list, block_stmt_iterator *bsi, bool is_output,
856 const struct sra_walk_fns *fns)
857 {
858 tree op;
859 for (op = list; op ; op = TREE_CHAIN (op))
860 sra_walk_expr (&TREE_VALUE (op), bsi, is_output, fns);
861 }
862
863 /* Walk the arguments of a CALL_EXPR looking for scalarizable aggregates.
864 If we find one, invoke FNS->USE. */
865
866 static void
867 sra_walk_call_expr (tree expr, block_stmt_iterator *bsi,
868 const struct sra_walk_fns *fns)
869 {
870 int i;
871 int nargs = call_expr_nargs (expr);
872 for (i = 0; i < nargs; i++)
873 sra_walk_expr (&CALL_EXPR_ARG (expr, i), bsi, false, fns);
874 }
875
876 /* Walk the inputs and outputs of an ASM_EXPR looking for scalarizable
877 aggregates. If we find one, invoke FNS->USE. */
878
879 static void
880 sra_walk_asm_expr (tree expr, block_stmt_iterator *bsi,
881 const struct sra_walk_fns *fns)
882 {
883 sra_walk_tree_list (ASM_INPUTS (expr), bsi, false, fns);
884 sra_walk_tree_list (ASM_OUTPUTS (expr), bsi, true, fns);
885 }
886
887 /* Walk a GIMPLE_MODIFY_STMT and categorize the assignment appropriately. */
888
889 static void
890 sra_walk_gimple_modify_stmt (tree expr, block_stmt_iterator *bsi,
891 const struct sra_walk_fns *fns)
892 {
893 struct sra_elt *lhs_elt, *rhs_elt;
894 tree lhs, rhs;
895
896 lhs = GIMPLE_STMT_OPERAND (expr, 0);
897 rhs = GIMPLE_STMT_OPERAND (expr, 1);
898 lhs_elt = maybe_lookup_element_for_expr (lhs);
899 rhs_elt = maybe_lookup_element_for_expr (rhs);
900
901 /* If both sides are scalarizable, this is a COPY operation. */
902 if (lhs_elt && rhs_elt)
903 {
904 fns->copy (lhs_elt, rhs_elt, bsi);
905 return;
906 }
907
908 /* If the RHS is scalarizable, handle it. There are only two cases. */
909 if (rhs_elt)
910 {
911 if (!rhs_elt->is_scalar)
912 fns->ldst (rhs_elt, lhs, bsi, false);
913 else
914 fns->use (rhs_elt, &GIMPLE_STMT_OPERAND (expr, 1), bsi, false, false);
915 }
916
917 /* If it isn't scalarizable, there may be scalarizable variables within, so
918 check for a call or else walk the RHS to see if we need to do any
919 copy-in operations. We need to do it before the LHS is scalarized so
920 that the statements get inserted in the proper place, before any
921 copy-out operations. */
922 else
923 {
924 tree call = get_call_expr_in (rhs);
925 if (call)
926 sra_walk_call_expr (call, bsi, fns);
927 else
928 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 1), bsi, false, fns);
929 }
930
931 /* Likewise, handle the LHS being scalarizable. We have cases similar
932 to those above, but also want to handle RHS being constant. */
933 if (lhs_elt)
934 {
935 /* If this is an assignment from a constant, or constructor, then
936 we have access to all of the elements individually. Invoke INIT. */
937 if (TREE_CODE (rhs) == COMPLEX_EXPR
938 || TREE_CODE (rhs) == COMPLEX_CST
939 || TREE_CODE (rhs) == CONSTRUCTOR)
940 fns->init (lhs_elt, rhs, bsi);
941
942 /* If this is an assignment from read-only memory, treat this as if
943 we'd been passed the constructor directly. Invoke INIT. */
944 else if (TREE_CODE (rhs) == VAR_DECL
945 && TREE_STATIC (rhs)
946 && TREE_READONLY (rhs)
947 && targetm.binds_local_p (rhs))
948 fns->init (lhs_elt, DECL_INITIAL (rhs), bsi);
949
950 /* If this is a copy from a non-scalarizable lvalue, invoke LDST.
951 The lvalue requirement prevents us from trying to directly scalarize
952 the result of a function call. Which would result in trying to call
953 the function multiple times, and other evil things. */
954 else if (!lhs_elt->is_scalar && is_gimple_addressable (rhs))
955 fns->ldst (lhs_elt, rhs, bsi, true);
956
957 /* Otherwise we're being used in some context that requires the
958 aggregate to be seen as a whole. Invoke USE. */
959 else
960 fns->use (lhs_elt, &GIMPLE_STMT_OPERAND (expr, 0), bsi, true, false);
961 }
962
963 /* Similarly to above, LHS_ELT being null only means that the LHS as a
964 whole is not a scalarizable reference. There may be occurrences of
965 scalarizable variables within, which implies a USE. */
966 else
967 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 0), bsi, true, fns);
968 }
969
970 /* Entry point to the walk functions. Search the entire function,
971 invoking the callbacks in FNS on each of the references to
972 scalarizable variables. */
973
974 static void
975 sra_walk_function (const struct sra_walk_fns *fns)
976 {
977 basic_block bb;
978 block_stmt_iterator si, ni;
979
980 /* ??? Phase 4 could derive some benefit to walking the function in
981 dominator tree order. */
982
983 FOR_EACH_BB (bb)
984 for (si = bsi_start (bb); !bsi_end_p (si); si = ni)
985 {
986 tree stmt, t;
987 stmt_ann_t ann;
988
989 stmt = bsi_stmt (si);
990 ann = stmt_ann (stmt);
991
992 ni = si;
993 bsi_next (&ni);
994
995 /* If the statement has no virtual operands, then it doesn't
996 make any structure references that we care about. */
997 if (gimple_aliases_computed_p (cfun)
998 && ZERO_SSA_OPERANDS (stmt, (SSA_OP_VIRTUAL_DEFS | SSA_OP_VUSE)))
999 continue;
1000
1001 switch (TREE_CODE (stmt))
1002 {
1003 case RETURN_EXPR:
1004 /* If we have "return <retval>" then the return value is
1005 already exposed for our pleasure. Walk it as a USE to
1006 force all the components back in place for the return.
1007
1008 If we have an embedded assignment, then <retval> is of
1009 a type that gets returned in registers in this ABI, and
1010 we do not wish to extend their lifetimes. Treat this
1011 as a USE of the variable on the RHS of this assignment. */
1012
1013 t = TREE_OPERAND (stmt, 0);
1014 if (t == NULL_TREE)
1015 ;
1016 else if (TREE_CODE (t) == GIMPLE_MODIFY_STMT)
1017 sra_walk_expr (&GIMPLE_STMT_OPERAND (t, 1), &si, false, fns);
1018 else
1019 sra_walk_expr (&TREE_OPERAND (stmt, 0), &si, false, fns);
1020 break;
1021
1022 case GIMPLE_MODIFY_STMT:
1023 sra_walk_gimple_modify_stmt (stmt, &si, fns);
1024 break;
1025 case CALL_EXPR:
1026 sra_walk_call_expr (stmt, &si, fns);
1027 break;
1028 case ASM_EXPR:
1029 sra_walk_asm_expr (stmt, &si, fns);
1030 break;
1031
1032 default:
1033 break;
1034 }
1035 }
1036 }
1037 \f
1038 /* Phase One: Scan all referenced variables in the program looking for
1039 structures that could be decomposed. */
1040
1041 static bool
1042 find_candidates_for_sra (void)
1043 {
1044 bool any_set = false;
1045 tree var;
1046 referenced_var_iterator rvi;
1047
1048 FOR_EACH_REFERENCED_VAR (var, rvi)
1049 {
1050 if (decl_can_be_decomposed_p (var))
1051 {
1052 bitmap_set_bit (sra_candidates, DECL_UID (var));
1053 any_set = true;
1054 }
1055 }
1056
1057 return any_set;
1058 }
1059
1060 \f
1061 /* Phase Two: Scan all references to scalarizable variables. Count the
1062 number of times they are used or copied respectively. */
1063
1064 /* Callbacks to fill in SRA_WALK_FNS. Everything but USE is
1065 considered a copy, because we can decompose the reference such that
1066 the sub-elements needn't be contiguous. */
1067
1068 static void
1069 scan_use (struct sra_elt *elt, tree *expr_p ATTRIBUTE_UNUSED,
1070 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
1071 bool is_output ATTRIBUTE_UNUSED, bool use_all ATTRIBUTE_UNUSED)
1072 {
1073 elt->n_uses += 1;
1074 }
1075
1076 static void
1077 scan_copy (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
1078 block_stmt_iterator *bsi ATTRIBUTE_UNUSED)
1079 {
1080 lhs_elt->n_copies += 1;
1081 rhs_elt->n_copies += 1;
1082 }
1083
1084 static void
1085 scan_init (struct sra_elt *lhs_elt, tree rhs ATTRIBUTE_UNUSED,
1086 block_stmt_iterator *bsi ATTRIBUTE_UNUSED)
1087 {
1088 lhs_elt->n_copies += 1;
1089 }
1090
1091 static void
1092 scan_ldst (struct sra_elt *elt, tree other ATTRIBUTE_UNUSED,
1093 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
1094 bool is_output ATTRIBUTE_UNUSED)
1095 {
1096 elt->n_copies += 1;
1097 }
1098
1099 /* Dump the values we collected during the scanning phase. */
1100
1101 static void
1102 scan_dump (struct sra_elt *elt)
1103 {
1104 struct sra_elt *c;
1105
1106 dump_sra_elt_name (dump_file, elt);
1107 fprintf (dump_file, ": n_uses=%u n_copies=%u\n", elt->n_uses, elt->n_copies);
1108
1109 for (c = elt->children; c ; c = c->sibling)
1110 scan_dump (c);
1111
1112 for (c = elt->groups; c ; c = c->sibling)
1113 scan_dump (c);
1114 }
1115
1116 /* Entry point to phase 2. Scan the entire function, building up
1117 scalarization data structures, recording copies and uses. */
1118
1119 static void
1120 scan_function (void)
1121 {
1122 static const struct sra_walk_fns fns = {
1123 scan_use, scan_copy, scan_init, scan_ldst, true
1124 };
1125 bitmap_iterator bi;
1126
1127 sra_walk_function (&fns);
1128
1129 if (dump_file && (dump_flags & TDF_DETAILS))
1130 {
1131 unsigned i;
1132
1133 fputs ("\nScan results:\n", dump_file);
1134 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1135 {
1136 tree var = referenced_var (i);
1137 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1138 if (elt)
1139 scan_dump (elt);
1140 }
1141 fputc ('\n', dump_file);
1142 }
1143 }
1144 \f
1145 /* Phase Three: Make decisions about which variables to scalarize, if any.
1146 All elements to be scalarized have replacement variables made for them. */
1147
1148 /* A subroutine of build_element_name. Recursively build the element
1149 name on the obstack. */
1150
1151 static void
1152 build_element_name_1 (struct sra_elt *elt)
1153 {
1154 tree t;
1155 char buffer[32];
1156
1157 if (elt->parent)
1158 {
1159 build_element_name_1 (elt->parent);
1160 obstack_1grow (&sra_obstack, '$');
1161
1162 if (TREE_CODE (elt->parent->type) == COMPLEX_TYPE)
1163 {
1164 if (elt->element == integer_zero_node)
1165 obstack_grow (&sra_obstack, "real", 4);
1166 else
1167 obstack_grow (&sra_obstack, "imag", 4);
1168 return;
1169 }
1170 }
1171
1172 t = elt->element;
1173 if (TREE_CODE (t) == INTEGER_CST)
1174 {
1175 /* ??? Eh. Don't bother doing double-wide printing. */
1176 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (t));
1177 obstack_grow (&sra_obstack, buffer, strlen (buffer));
1178 }
1179 else
1180 {
1181 tree name = DECL_NAME (t);
1182 if (name)
1183 obstack_grow (&sra_obstack, IDENTIFIER_POINTER (name),
1184 IDENTIFIER_LENGTH (name));
1185 else
1186 {
1187 sprintf (buffer, "D%u", DECL_UID (t));
1188 obstack_grow (&sra_obstack, buffer, strlen (buffer));
1189 }
1190 }
1191 }
1192
1193 /* Construct a pretty variable name for an element's replacement variable.
1194 The name is built on the obstack. */
1195
1196 static char *
1197 build_element_name (struct sra_elt *elt)
1198 {
1199 build_element_name_1 (elt);
1200 obstack_1grow (&sra_obstack, '\0');
1201 return XOBFINISH (&sra_obstack, char *);
1202 }
1203
1204 /* Instantiate an element as an independent variable. */
1205
1206 static void
1207 instantiate_element (struct sra_elt *elt)
1208 {
1209 struct sra_elt *base_elt;
1210 tree var, base;
1211
1212 for (base_elt = elt; base_elt->parent; base_elt = base_elt->parent)
1213 continue;
1214 base = base_elt->element;
1215
1216 elt->replacement = var = make_rename_temp (elt->type, "SR");
1217
1218 /* For vectors, if used on the left hand side with BIT_FIELD_REF,
1219 they are not a gimple register. */
1220 if (TREE_CODE (TREE_TYPE (var)) == VECTOR_TYPE && elt->is_vector_lhs)
1221 DECL_GIMPLE_REG_P (var) = 0;
1222
1223 DECL_SOURCE_LOCATION (var) = DECL_SOURCE_LOCATION (base);
1224 DECL_ARTIFICIAL (var) = 1;
1225
1226 if (TREE_THIS_VOLATILE (elt->type))
1227 {
1228 TREE_THIS_VOLATILE (var) = 1;
1229 TREE_SIDE_EFFECTS (var) = 1;
1230 }
1231
1232 if (DECL_NAME (base) && !DECL_IGNORED_P (base))
1233 {
1234 char *pretty_name = build_element_name (elt);
1235 DECL_NAME (var) = get_identifier (pretty_name);
1236 obstack_free (&sra_obstack, pretty_name);
1237
1238 SET_DECL_DEBUG_EXPR (var, generate_element_ref (elt));
1239 DECL_DEBUG_EXPR_IS_FROM (var) = 1;
1240
1241 DECL_IGNORED_P (var) = 0;
1242 TREE_NO_WARNING (var) = TREE_NO_WARNING (base);
1243 }
1244 else
1245 {
1246 DECL_IGNORED_P (var) = 1;
1247 /* ??? We can't generate any warning that would be meaningful. */
1248 TREE_NO_WARNING (var) = 1;
1249 }
1250
1251 if (dump_file)
1252 {
1253 fputs (" ", dump_file);
1254 dump_sra_elt_name (dump_file, elt);
1255 fputs (" -> ", dump_file);
1256 print_generic_expr (dump_file, var, dump_flags);
1257 fputc ('\n', dump_file);
1258 }
1259 }
1260
1261 /* Make one pass across an element tree deciding whether or not it's
1262 profitable to instantiate individual leaf scalars.
1263
1264 PARENT_USES and PARENT_COPIES are the sum of the N_USES and N_COPIES
1265 fields all the way up the tree. */
1266
1267 static void
1268 decide_instantiation_1 (struct sra_elt *elt, unsigned int parent_uses,
1269 unsigned int parent_copies)
1270 {
1271 if (dump_file && !elt->parent)
1272 {
1273 fputs ("Initial instantiation for ", dump_file);
1274 dump_sra_elt_name (dump_file, elt);
1275 fputc ('\n', dump_file);
1276 }
1277
1278 if (elt->cannot_scalarize)
1279 return;
1280
1281 if (elt->is_scalar)
1282 {
1283 /* The decision is simple: instantiate if we're used more frequently
1284 than the parent needs to be seen as a complete unit. */
1285 if (elt->n_uses + elt->n_copies + parent_copies > parent_uses)
1286 instantiate_element (elt);
1287 }
1288 else
1289 {
1290 struct sra_elt *c, *group;
1291 unsigned int this_uses = elt->n_uses + parent_uses;
1292 unsigned int this_copies = elt->n_copies + parent_copies;
1293
1294 /* Consider groups of sub-elements as weighing in favour of
1295 instantiation whatever their size. */
1296 for (group = elt->groups; group ; group = group->sibling)
1297 FOR_EACH_ACTUAL_CHILD (c, group)
1298 {
1299 c->n_uses += group->n_uses;
1300 c->n_copies += group->n_copies;
1301 }
1302
1303 for (c = elt->children; c ; c = c->sibling)
1304 decide_instantiation_1 (c, this_uses, this_copies);
1305 }
1306 }
1307
1308 /* Compute the size and number of all instantiated elements below ELT.
1309 We will only care about this if the size of the complete structure
1310 fits in a HOST_WIDE_INT, so we don't have to worry about overflow. */
1311
1312 static unsigned int
1313 sum_instantiated_sizes (struct sra_elt *elt, unsigned HOST_WIDE_INT *sizep)
1314 {
1315 if (elt->replacement)
1316 {
1317 *sizep += TREE_INT_CST_LOW (TYPE_SIZE_UNIT (elt->type));
1318 return 1;
1319 }
1320 else
1321 {
1322 struct sra_elt *c;
1323 unsigned int count = 0;
1324
1325 for (c = elt->children; c ; c = c->sibling)
1326 count += sum_instantiated_sizes (c, sizep);
1327
1328 return count;
1329 }
1330 }
1331
1332 /* Instantiate fields in ELT->TYPE that are not currently present as
1333 children of ELT. */
1334
1335 static void instantiate_missing_elements (struct sra_elt *elt);
1336
1337 static void
1338 instantiate_missing_elements_1 (struct sra_elt *elt, tree child, tree type)
1339 {
1340 struct sra_elt *sub = lookup_element (elt, child, type, INSERT);
1341 if (sub->is_scalar)
1342 {
1343 if (sub->replacement == NULL)
1344 instantiate_element (sub);
1345 }
1346 else
1347 instantiate_missing_elements (sub);
1348 }
1349
1350 static void
1351 instantiate_missing_elements (struct sra_elt *elt)
1352 {
1353 tree type = elt->type;
1354
1355 switch (TREE_CODE (type))
1356 {
1357 case RECORD_TYPE:
1358 {
1359 tree f;
1360 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
1361 if (TREE_CODE (f) == FIELD_DECL)
1362 {
1363 tree field_type = TREE_TYPE (f);
1364
1365 /* canonicalize_component_ref() unwidens some bit-field
1366 types (not marked as DECL_BIT_FIELD in C++), so we
1367 must do the same, lest we may introduce type
1368 mismatches. */
1369 if (INTEGRAL_TYPE_P (field_type)
1370 && DECL_MODE (f) != TYPE_MODE (field_type))
1371 field_type = TREE_TYPE (get_unwidened (build3 (COMPONENT_REF,
1372 field_type,
1373 elt->element,
1374 f, NULL_TREE),
1375 NULL_TREE));
1376
1377 instantiate_missing_elements_1 (elt, f, field_type);
1378 }
1379 break;
1380 }
1381
1382 case ARRAY_TYPE:
1383 {
1384 tree i, max, subtype;
1385
1386 i = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1387 max = TYPE_MAX_VALUE (TYPE_DOMAIN (type));
1388 subtype = TREE_TYPE (type);
1389
1390 while (1)
1391 {
1392 instantiate_missing_elements_1 (elt, i, subtype);
1393 if (tree_int_cst_equal (i, max))
1394 break;
1395 i = int_const_binop (PLUS_EXPR, i, integer_one_node, true);
1396 }
1397
1398 break;
1399 }
1400
1401 case COMPLEX_TYPE:
1402 type = TREE_TYPE (type);
1403 instantiate_missing_elements_1 (elt, integer_zero_node, type);
1404 instantiate_missing_elements_1 (elt, integer_one_node, type);
1405 break;
1406
1407 default:
1408 gcc_unreachable ();
1409 }
1410 }
1411
1412 /* Return true if there is only one non aggregate field in the record, TYPE.
1413 Return false otherwise. */
1414
1415 static bool
1416 single_scalar_field_in_record_p (tree type)
1417 {
1418 int num_fields = 0;
1419 tree field;
1420 if (TREE_CODE (type) != RECORD_TYPE)
1421 return false;
1422
1423 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
1424 if (TREE_CODE (field) == FIELD_DECL)
1425 {
1426 num_fields++;
1427
1428 if (num_fields == 2)
1429 return false;
1430
1431 if (AGGREGATE_TYPE_P (TREE_TYPE (field)))
1432 return false;
1433 }
1434
1435 return true;
1436 }
1437
1438 /* Make one pass across an element tree deciding whether to perform block
1439 or element copies. If we decide on element copies, instantiate all
1440 elements. Return true if there are any instantiated sub-elements. */
1441
1442 static bool
1443 decide_block_copy (struct sra_elt *elt)
1444 {
1445 struct sra_elt *c;
1446 bool any_inst;
1447
1448 /* We shouldn't be invoked on groups of sub-elements as they must
1449 behave like their parent as far as block copy is concerned. */
1450 gcc_assert (!elt->is_group);
1451
1452 /* If scalarization is disabled, respect it. */
1453 if (elt->cannot_scalarize)
1454 {
1455 elt->use_block_copy = 1;
1456
1457 if (dump_file)
1458 {
1459 fputs ("Scalarization disabled for ", dump_file);
1460 dump_sra_elt_name (dump_file, elt);
1461 fputc ('\n', dump_file);
1462 }
1463
1464 /* Disable scalarization of sub-elements */
1465 for (c = elt->children; c; c = c->sibling)
1466 {
1467 c->cannot_scalarize = 1;
1468 decide_block_copy (c);
1469 }
1470
1471 /* Groups behave like their parent. */
1472 for (c = elt->groups; c; c = c->sibling)
1473 {
1474 c->cannot_scalarize = 1;
1475 c->use_block_copy = 1;
1476 }
1477
1478 return false;
1479 }
1480
1481 /* Don't decide if we've no uses. */
1482 if (elt->n_uses == 0 && elt->n_copies == 0)
1483 ;
1484
1485 else if (!elt->is_scalar)
1486 {
1487 tree size_tree = TYPE_SIZE_UNIT (elt->type);
1488 bool use_block_copy = true;
1489
1490 /* Tradeoffs for COMPLEX types pretty much always make it better
1491 to go ahead and split the components. */
1492 if (TREE_CODE (elt->type) == COMPLEX_TYPE)
1493 use_block_copy = false;
1494
1495 /* Don't bother trying to figure out the rest if the structure is
1496 so large we can't do easy arithmetic. This also forces block
1497 copies for variable sized structures. */
1498 else if (host_integerp (size_tree, 1))
1499 {
1500 unsigned HOST_WIDE_INT full_size, inst_size = 0;
1501 unsigned int max_size, max_count, inst_count, full_count;
1502
1503 /* If the sra-max-structure-size parameter is 0, then the
1504 user has not overridden the parameter and we can choose a
1505 sensible default. */
1506 max_size = SRA_MAX_STRUCTURE_SIZE
1507 ? SRA_MAX_STRUCTURE_SIZE
1508 : MOVE_RATIO * UNITS_PER_WORD;
1509 max_count = SRA_MAX_STRUCTURE_COUNT
1510 ? SRA_MAX_STRUCTURE_COUNT
1511 : MOVE_RATIO;
1512
1513 full_size = tree_low_cst (size_tree, 1);
1514 full_count = count_type_elements (elt->type, false);
1515 inst_count = sum_instantiated_sizes (elt, &inst_size);
1516
1517 /* If there is only one scalar field in the record, don't block copy. */
1518 if (single_scalar_field_in_record_p (elt->type))
1519 use_block_copy = false;
1520
1521 /* ??? What to do here. If there are two fields, and we've only
1522 instantiated one, then instantiating the other is clearly a win.
1523 If there are a large number of fields then the size of the copy
1524 is much more of a factor. */
1525
1526 /* If the structure is small, and we've made copies, go ahead
1527 and instantiate, hoping that the copies will go away. */
1528 if (full_size <= max_size
1529 && (full_count - inst_count) <= max_count
1530 && elt->n_copies > elt->n_uses)
1531 use_block_copy = false;
1532 else if (inst_count * 100 >= full_count * SRA_FIELD_STRUCTURE_RATIO
1533 && inst_size * 100 >= full_size * SRA_FIELD_STRUCTURE_RATIO)
1534 use_block_copy = false;
1535
1536 /* In order to avoid block copy, we have to be able to instantiate
1537 all elements of the type. See if this is possible. */
1538 if (!use_block_copy
1539 && (!can_completely_scalarize_p (elt)
1540 || !type_can_instantiate_all_elements (elt->type)))
1541 use_block_copy = true;
1542 }
1543
1544 elt->use_block_copy = use_block_copy;
1545
1546 /* Groups behave like their parent. */
1547 for (c = elt->groups; c; c = c->sibling)
1548 c->use_block_copy = use_block_copy;
1549
1550 if (dump_file)
1551 {
1552 fprintf (dump_file, "Using %s for ",
1553 use_block_copy ? "block-copy" : "element-copy");
1554 dump_sra_elt_name (dump_file, elt);
1555 fputc ('\n', dump_file);
1556 }
1557
1558 if (!use_block_copy)
1559 {
1560 instantiate_missing_elements (elt);
1561 return true;
1562 }
1563 }
1564
1565 any_inst = elt->replacement != NULL;
1566
1567 for (c = elt->children; c ; c = c->sibling)
1568 any_inst |= decide_block_copy (c);
1569
1570 return any_inst;
1571 }
1572
1573 /* Entry point to phase 3. Instantiate scalar replacement variables. */
1574
1575 static void
1576 decide_instantiations (void)
1577 {
1578 unsigned int i;
1579 bool cleared_any;
1580 bitmap_head done_head;
1581 bitmap_iterator bi;
1582
1583 /* We cannot clear bits from a bitmap we're iterating over,
1584 so save up all the bits to clear until the end. */
1585 bitmap_initialize (&done_head, &bitmap_default_obstack);
1586 cleared_any = false;
1587
1588 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1589 {
1590 tree var = referenced_var (i);
1591 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1592 if (elt)
1593 {
1594 decide_instantiation_1 (elt, 0, 0);
1595 if (!decide_block_copy (elt))
1596 elt = NULL;
1597 }
1598 if (!elt)
1599 {
1600 bitmap_set_bit (&done_head, i);
1601 cleared_any = true;
1602 }
1603 }
1604
1605 if (cleared_any)
1606 {
1607 bitmap_and_compl_into (sra_candidates, &done_head);
1608 bitmap_and_compl_into (needs_copy_in, &done_head);
1609 }
1610 bitmap_clear (&done_head);
1611
1612 if (!bitmap_empty_p (sra_candidates))
1613 todoflags |= TODO_update_smt_usage;
1614
1615 mark_set_for_renaming (sra_candidates);
1616
1617 if (dump_file)
1618 fputc ('\n', dump_file);
1619 }
1620
1621 \f
1622 /* Phase Four: Update the function to match the replacements created. */
1623
1624 /* Mark all the variables in VDEF/VUSE operators for STMT for
1625 renaming. This becomes necessary when we modify all of a
1626 non-scalar. */
1627
1628 static void
1629 mark_all_v_defs_1 (tree stmt)
1630 {
1631 tree sym;
1632 ssa_op_iter iter;
1633
1634 update_stmt_if_modified (stmt);
1635
1636 FOR_EACH_SSA_TREE_OPERAND (sym, stmt, iter, SSA_OP_ALL_VIRTUALS)
1637 {
1638 if (TREE_CODE (sym) == SSA_NAME)
1639 sym = SSA_NAME_VAR (sym);
1640 mark_sym_for_renaming (sym);
1641 }
1642 }
1643
1644
1645 /* Mark all the variables in virtual operands in all the statements in
1646 LIST for renaming. */
1647
1648 static void
1649 mark_all_v_defs (tree list)
1650 {
1651 if (TREE_CODE (list) != STATEMENT_LIST)
1652 mark_all_v_defs_1 (list);
1653 else
1654 {
1655 tree_stmt_iterator i;
1656 for (i = tsi_start (list); !tsi_end_p (i); tsi_next (&i))
1657 mark_all_v_defs_1 (tsi_stmt (i));
1658 }
1659 }
1660
1661
1662 /* Mark every replacement under ELT with TREE_NO_WARNING. */
1663
1664 static void
1665 mark_no_warning (struct sra_elt *elt)
1666 {
1667 if (!elt->all_no_warning)
1668 {
1669 if (elt->replacement)
1670 TREE_NO_WARNING (elt->replacement) = 1;
1671 else
1672 {
1673 struct sra_elt *c;
1674 FOR_EACH_ACTUAL_CHILD (c, elt)
1675 mark_no_warning (c);
1676 }
1677 elt->all_no_warning = true;
1678 }
1679 }
1680
1681 /* Build a single level component reference to ELT rooted at BASE. */
1682
1683 static tree
1684 generate_one_element_ref (struct sra_elt *elt, tree base)
1685 {
1686 switch (TREE_CODE (TREE_TYPE (base)))
1687 {
1688 case RECORD_TYPE:
1689 {
1690 tree field = elt->element;
1691
1692 /* Watch out for compatible records with differing field lists. */
1693 if (DECL_FIELD_CONTEXT (field) != TYPE_MAIN_VARIANT (TREE_TYPE (base)))
1694 field = find_compatible_field (TREE_TYPE (base), field);
1695
1696 return build3 (COMPONENT_REF, elt->type, base, field, NULL);
1697 }
1698
1699 case ARRAY_TYPE:
1700 todoflags |= TODO_update_smt_usage;
1701 if (TREE_CODE (elt->element) == RANGE_EXPR)
1702 return build4 (ARRAY_RANGE_REF, elt->type, base,
1703 TREE_OPERAND (elt->element, 0), NULL, NULL);
1704 else
1705 return build4 (ARRAY_REF, elt->type, base, elt->element, NULL, NULL);
1706
1707 case COMPLEX_TYPE:
1708 if (elt->element == integer_zero_node)
1709 return build1 (REALPART_EXPR, elt->type, base);
1710 else
1711 return build1 (IMAGPART_EXPR, elt->type, base);
1712
1713 default:
1714 gcc_unreachable ();
1715 }
1716 }
1717
1718 /* Build a full component reference to ELT rooted at its native variable. */
1719
1720 static tree
1721 generate_element_ref (struct sra_elt *elt)
1722 {
1723 if (elt->parent)
1724 return generate_one_element_ref (elt, generate_element_ref (elt->parent));
1725 else
1726 return elt->element;
1727 }
1728
1729 /* Create an assignment statement from SRC to DST. */
1730
1731 static tree
1732 sra_build_assignment (tree dst, tree src)
1733 {
1734 #if 0 /* ENABLE_CHECKING */
1735 /* This test ought to pass, but it is unfortunately too strict for
1736 now. */
1737 gcc_assert (TYPE_CANONICAL (TYPE_MAIN_VARIANT (TREE_TYPE (dst)))
1738 == TYPE_CANONICAL (TYPE_MAIN_VARIANT (TREE_TYPE (src))));
1739 #endif
1740 return build2 (GIMPLE_MODIFY_STMT, void_type_node, dst, src);
1741 }
1742
1743 /* Generate a set of assignment statements in *LIST_P to copy all
1744 instantiated elements under ELT to or from the equivalent structure
1745 rooted at EXPR. COPY_OUT controls the direction of the copy, with
1746 true meaning to copy out of EXPR into ELT. */
1747
1748 static void
1749 generate_copy_inout (struct sra_elt *elt, bool copy_out, tree expr,
1750 tree *list_p)
1751 {
1752 struct sra_elt *c;
1753 tree t;
1754
1755 if (!copy_out && TREE_CODE (expr) == SSA_NAME
1756 && TREE_CODE (TREE_TYPE (expr)) == COMPLEX_TYPE)
1757 {
1758 tree r, i;
1759
1760 c = lookup_element (elt, integer_zero_node, NULL, NO_INSERT);
1761 r = c->replacement;
1762 c = lookup_element (elt, integer_one_node, NULL, NO_INSERT);
1763 i = c->replacement;
1764
1765 t = build2 (COMPLEX_EXPR, elt->type, r, i);
1766 t = sra_build_assignment (expr, t);
1767 SSA_NAME_DEF_STMT (expr) = t;
1768 append_to_statement_list (t, list_p);
1769 }
1770 else if (elt->replacement)
1771 {
1772 if (copy_out)
1773 t = sra_build_assignment (elt->replacement, expr);
1774 else
1775 t = sra_build_assignment (expr, elt->replacement);
1776 append_to_statement_list (t, list_p);
1777 }
1778 else
1779 {
1780 FOR_EACH_ACTUAL_CHILD (c, elt)
1781 {
1782 t = generate_one_element_ref (c, unshare_expr (expr));
1783 generate_copy_inout (c, copy_out, t, list_p);
1784 }
1785 }
1786 }
1787
1788 /* Generate a set of assignment statements in *LIST_P to copy all instantiated
1789 elements under SRC to their counterparts under DST. There must be a 1-1
1790 correspondence of instantiated elements. */
1791
1792 static void
1793 generate_element_copy (struct sra_elt *dst, struct sra_elt *src, tree *list_p)
1794 {
1795 struct sra_elt *dc, *sc;
1796
1797 FOR_EACH_ACTUAL_CHILD (dc, dst)
1798 {
1799 sc = lookup_element (src, dc->element, NULL, NO_INSERT);
1800 gcc_assert (sc);
1801 generate_element_copy (dc, sc, list_p);
1802 }
1803
1804 if (dst->replacement)
1805 {
1806 tree t;
1807
1808 gcc_assert (src->replacement);
1809
1810 t = sra_build_assignment (dst->replacement, src->replacement);
1811 append_to_statement_list (t, list_p);
1812 }
1813 }
1814
1815 /* Generate a set of assignment statements in *LIST_P to zero all instantiated
1816 elements under ELT. In addition, do not assign to elements that have been
1817 marked VISITED but do reset the visited flag; this allows easy coordination
1818 with generate_element_init. */
1819
1820 static void
1821 generate_element_zero (struct sra_elt *elt, tree *list_p)
1822 {
1823 struct sra_elt *c;
1824
1825 if (elt->visited)
1826 {
1827 elt->visited = false;
1828 return;
1829 }
1830
1831 FOR_EACH_ACTUAL_CHILD (c, elt)
1832 generate_element_zero (c, list_p);
1833
1834 if (elt->replacement)
1835 {
1836 tree t;
1837
1838 gcc_assert (elt->is_scalar);
1839 t = fold_convert (elt->type, integer_zero_node);
1840
1841 t = sra_build_assignment (elt->replacement, t);
1842 append_to_statement_list (t, list_p);
1843 }
1844 }
1845
1846 /* Generate an assignment VAR = INIT, where INIT may need gimplification.
1847 Add the result to *LIST_P. */
1848
1849 static void
1850 generate_one_element_init (tree var, tree init, tree *list_p)
1851 {
1852 /* The replacement can be almost arbitrarily complex. Gimplify. */
1853 tree stmt = sra_build_assignment (var, init);
1854 gimplify_and_add (stmt, list_p);
1855 }
1856
1857 /* Generate a set of assignment statements in *LIST_P to set all instantiated
1858 elements under ELT with the contents of the initializer INIT. In addition,
1859 mark all assigned elements VISITED; this allows easy coordination with
1860 generate_element_zero. Return false if we found a case we couldn't
1861 handle. */
1862
1863 static bool
1864 generate_element_init_1 (struct sra_elt *elt, tree init, tree *list_p)
1865 {
1866 bool result = true;
1867 enum tree_code init_code;
1868 struct sra_elt *sub;
1869 tree t;
1870 unsigned HOST_WIDE_INT idx;
1871 tree value, purpose;
1872
1873 /* We can be passed DECL_INITIAL of a static variable. It might have a
1874 conversion, which we strip off here. */
1875 STRIP_USELESS_TYPE_CONVERSION (init);
1876 init_code = TREE_CODE (init);
1877
1878 if (elt->is_scalar)
1879 {
1880 if (elt->replacement)
1881 {
1882 generate_one_element_init (elt->replacement, init, list_p);
1883 elt->visited = true;
1884 }
1885 return result;
1886 }
1887
1888 switch (init_code)
1889 {
1890 case COMPLEX_CST:
1891 case COMPLEX_EXPR:
1892 FOR_EACH_ACTUAL_CHILD (sub, elt)
1893 {
1894 if (sub->element == integer_zero_node)
1895 t = (init_code == COMPLEX_EXPR
1896 ? TREE_OPERAND (init, 0) : TREE_REALPART (init));
1897 else
1898 t = (init_code == COMPLEX_EXPR
1899 ? TREE_OPERAND (init, 1) : TREE_IMAGPART (init));
1900 result &= generate_element_init_1 (sub, t, list_p);
1901 }
1902 break;
1903
1904 case CONSTRUCTOR:
1905 FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (init), idx, purpose, value)
1906 {
1907 if (TREE_CODE (purpose) == RANGE_EXPR)
1908 {
1909 tree lower = TREE_OPERAND (purpose, 0);
1910 tree upper = TREE_OPERAND (purpose, 1);
1911
1912 while (1)
1913 {
1914 sub = lookup_element (elt, lower, NULL, NO_INSERT);
1915 if (sub != NULL)
1916 result &= generate_element_init_1 (sub, value, list_p);
1917 if (tree_int_cst_equal (lower, upper))
1918 break;
1919 lower = int_const_binop (PLUS_EXPR, lower,
1920 integer_one_node, true);
1921 }
1922 }
1923 else
1924 {
1925 sub = lookup_element (elt, purpose, NULL, NO_INSERT);
1926 if (sub != NULL)
1927 result &= generate_element_init_1 (sub, value, list_p);
1928 }
1929 }
1930 break;
1931
1932 default:
1933 elt->visited = true;
1934 result = false;
1935 }
1936
1937 return result;
1938 }
1939
1940 /* A wrapper function for generate_element_init_1 that handles cleanup after
1941 gimplification. */
1942
1943 static bool
1944 generate_element_init (struct sra_elt *elt, tree init, tree *list_p)
1945 {
1946 bool ret;
1947
1948 push_gimplify_context ();
1949 ret = generate_element_init_1 (elt, init, list_p);
1950 pop_gimplify_context (NULL);
1951
1952 /* The replacement can expose previously unreferenced variables. */
1953 if (ret && *list_p)
1954 {
1955 tree_stmt_iterator i;
1956
1957 for (i = tsi_start (*list_p); !tsi_end_p (i); tsi_next (&i))
1958 find_new_referenced_vars (tsi_stmt_ptr (i));
1959 }
1960
1961 return ret;
1962 }
1963
1964 /* Insert STMT on all the outgoing edges out of BB. Note that if BB
1965 has more than one edge, STMT will be replicated for each edge. Also,
1966 abnormal edges will be ignored. */
1967
1968 void
1969 insert_edge_copies (tree stmt, basic_block bb)
1970 {
1971 edge e;
1972 edge_iterator ei;
1973 bool first_copy;
1974
1975 first_copy = true;
1976 FOR_EACH_EDGE (e, ei, bb->succs)
1977 {
1978 /* We don't need to insert copies on abnormal edges. The
1979 value of the scalar replacement is not guaranteed to
1980 be valid through an abnormal edge. */
1981 if (!(e->flags & EDGE_ABNORMAL))
1982 {
1983 if (first_copy)
1984 {
1985 bsi_insert_on_edge (e, stmt);
1986 first_copy = false;
1987 }
1988 else
1989 bsi_insert_on_edge (e, unsave_expr_now (stmt));
1990 }
1991 }
1992 }
1993
1994 /* Helper function to insert LIST before BSI, and set up line number info. */
1995
1996 void
1997 sra_insert_before (block_stmt_iterator *bsi, tree list)
1998 {
1999 tree stmt = bsi_stmt (*bsi);
2000
2001 if (EXPR_HAS_LOCATION (stmt))
2002 annotate_all_with_locus (&list, EXPR_LOCATION (stmt));
2003 bsi_insert_before (bsi, list, BSI_SAME_STMT);
2004 }
2005
2006 /* Similarly, but insert after BSI. Handles insertion onto edges as well. */
2007
2008 void
2009 sra_insert_after (block_stmt_iterator *bsi, tree list)
2010 {
2011 tree stmt = bsi_stmt (*bsi);
2012
2013 if (EXPR_HAS_LOCATION (stmt))
2014 annotate_all_with_locus (&list, EXPR_LOCATION (stmt));
2015
2016 if (stmt_ends_bb_p (stmt))
2017 insert_edge_copies (list, bsi->bb);
2018 else
2019 bsi_insert_after (bsi, list, BSI_SAME_STMT);
2020 }
2021
2022 /* Similarly, but replace the statement at BSI. */
2023
2024 static void
2025 sra_replace (block_stmt_iterator *bsi, tree list)
2026 {
2027 sra_insert_before (bsi, list);
2028 bsi_remove (bsi, false);
2029 if (bsi_end_p (*bsi))
2030 *bsi = bsi_last (bsi->bb);
2031 else
2032 bsi_prev (bsi);
2033 }
2034
2035 /* Scalarize a USE. To recap, this is either a simple reference to ELT,
2036 if elt is scalar, or some occurrence of ELT that requires a complete
2037 aggregate. IS_OUTPUT is true if ELT is being modified. */
2038
2039 static void
2040 scalarize_use (struct sra_elt *elt, tree *expr_p, block_stmt_iterator *bsi,
2041 bool is_output, bool use_all)
2042 {
2043 tree list = NULL, stmt = bsi_stmt (*bsi);
2044
2045 if (elt->replacement)
2046 {
2047 /* If we have a replacement, then updating the reference is as
2048 simple as modifying the existing statement in place. */
2049 if (is_output)
2050 mark_all_v_defs (stmt);
2051 *expr_p = elt->replacement;
2052 update_stmt (stmt);
2053 }
2054 else
2055 {
2056 /* Otherwise we need some copies. If ELT is being read, then we want
2057 to store all (modified) sub-elements back into the structure before
2058 the reference takes place. If ELT is being written, then we want to
2059 load the changed values back into our shadow variables. */
2060 /* ??? We don't check modified for reads, we just always write all of
2061 the values. We should be able to record the SSA number of the VOP
2062 for which the values were last read. If that number matches the
2063 SSA number of the VOP in the current statement, then we needn't
2064 emit an assignment. This would also eliminate double writes when
2065 a structure is passed as more than one argument to a function call.
2066 This optimization would be most effective if sra_walk_function
2067 processed the blocks in dominator order. */
2068
2069 generate_copy_inout (elt, is_output, generate_element_ref (elt), &list);
2070 if (list == NULL)
2071 return;
2072 mark_all_v_defs (list);
2073 if (is_output)
2074 sra_insert_after (bsi, list);
2075 else
2076 {
2077 sra_insert_before (bsi, list);
2078 if (use_all)
2079 mark_no_warning (elt);
2080 }
2081 }
2082 }
2083
2084 /* Scalarize a COPY. To recap, this is an assignment statement between
2085 two scalarizable references, LHS_ELT and RHS_ELT. */
2086
2087 static void
2088 scalarize_copy (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
2089 block_stmt_iterator *bsi)
2090 {
2091 tree list, stmt;
2092
2093 if (lhs_elt->replacement && rhs_elt->replacement)
2094 {
2095 /* If we have two scalar operands, modify the existing statement. */
2096 stmt = bsi_stmt (*bsi);
2097
2098 /* See the commentary in sra_walk_function concerning
2099 RETURN_EXPR, and why we should never see one here. */
2100 gcc_assert (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT);
2101
2102 GIMPLE_STMT_OPERAND (stmt, 0) = lhs_elt->replacement;
2103 GIMPLE_STMT_OPERAND (stmt, 1) = rhs_elt->replacement;
2104 update_stmt (stmt);
2105 }
2106 else if (lhs_elt->use_block_copy || rhs_elt->use_block_copy)
2107 {
2108 /* If either side requires a block copy, then sync the RHS back
2109 to the original structure, leave the original assignment
2110 statement (which will perform the block copy), then load the
2111 LHS values out of its now-updated original structure. */
2112 /* ??? Could perform a modified pair-wise element copy. That
2113 would at least allow those elements that are instantiated in
2114 both structures to be optimized well. */
2115
2116 list = NULL;
2117 generate_copy_inout (rhs_elt, false,
2118 generate_element_ref (rhs_elt), &list);
2119 if (list)
2120 {
2121 mark_all_v_defs (list);
2122 sra_insert_before (bsi, list);
2123 }
2124
2125 list = NULL;
2126 generate_copy_inout (lhs_elt, true,
2127 generate_element_ref (lhs_elt), &list);
2128 if (list)
2129 {
2130 mark_all_v_defs (list);
2131 sra_insert_after (bsi, list);
2132 }
2133 }
2134 else
2135 {
2136 /* Otherwise both sides must be fully instantiated. In which
2137 case perform pair-wise element assignments and replace the
2138 original block copy statement. */
2139
2140 stmt = bsi_stmt (*bsi);
2141 mark_all_v_defs (stmt);
2142
2143 list = NULL;
2144 generate_element_copy (lhs_elt, rhs_elt, &list);
2145 gcc_assert (list);
2146 mark_all_v_defs (list);
2147 sra_replace (bsi, list);
2148 }
2149 }
2150
2151 /* Scalarize an INIT. To recap, this is an assignment to a scalarizable
2152 reference from some form of constructor: CONSTRUCTOR, COMPLEX_CST or
2153 COMPLEX_EXPR. If RHS is NULL, it should be treated as an empty
2154 CONSTRUCTOR. */
2155
2156 static void
2157 scalarize_init (struct sra_elt *lhs_elt, tree rhs, block_stmt_iterator *bsi)
2158 {
2159 bool result = true;
2160 tree list = NULL;
2161
2162 /* Generate initialization statements for all members extant in the RHS. */
2163 if (rhs)
2164 {
2165 /* Unshare the expression just in case this is from a decl's initial. */
2166 rhs = unshare_expr (rhs);
2167 result = generate_element_init (lhs_elt, rhs, &list);
2168 }
2169
2170 /* CONSTRUCTOR is defined such that any member not mentioned is assigned
2171 a zero value. Initialize the rest of the instantiated elements. */
2172 generate_element_zero (lhs_elt, &list);
2173
2174 if (!result)
2175 {
2176 /* If we failed to convert the entire initializer, then we must
2177 leave the structure assignment in place and must load values
2178 from the structure into the slots for which we did not find
2179 constants. The easiest way to do this is to generate a complete
2180 copy-out, and then follow that with the constant assignments
2181 that we were able to build. DCE will clean things up. */
2182 tree list0 = NULL;
2183 generate_copy_inout (lhs_elt, true, generate_element_ref (lhs_elt),
2184 &list0);
2185 append_to_statement_list (list, &list0);
2186 list = list0;
2187 }
2188
2189 if (lhs_elt->use_block_copy || !result)
2190 {
2191 /* Since LHS is not fully instantiated, we must leave the structure
2192 assignment in place. Treating this case differently from a USE
2193 exposes constants to later optimizations. */
2194 if (list)
2195 {
2196 mark_all_v_defs (list);
2197 sra_insert_after (bsi, list);
2198 }
2199 }
2200 else
2201 {
2202 /* The LHS is fully instantiated. The list of initializations
2203 replaces the original structure assignment. */
2204 gcc_assert (list);
2205 mark_all_v_defs (bsi_stmt (*bsi));
2206 mark_all_v_defs (list);
2207 sra_replace (bsi, list);
2208 }
2209 }
2210
2211 /* A subroutine of scalarize_ldst called via walk_tree. Set TREE_NO_TRAP
2212 on all INDIRECT_REFs. */
2213
2214 static tree
2215 mark_notrap (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
2216 {
2217 tree t = *tp;
2218
2219 if (TREE_CODE (t) == INDIRECT_REF)
2220 {
2221 TREE_THIS_NOTRAP (t) = 1;
2222 *walk_subtrees = 0;
2223 }
2224 else if (IS_TYPE_OR_DECL_P (t))
2225 *walk_subtrees = 0;
2226
2227 return NULL;
2228 }
2229
2230 /* Scalarize a LDST. To recap, this is an assignment between one scalarizable
2231 reference ELT and one non-scalarizable reference OTHER. IS_OUTPUT is true
2232 if ELT is on the left-hand side. */
2233
2234 static void
2235 scalarize_ldst (struct sra_elt *elt, tree other,
2236 block_stmt_iterator *bsi, bool is_output)
2237 {
2238 /* Shouldn't have gotten called for a scalar. */
2239 gcc_assert (!elt->replacement);
2240
2241 if (elt->use_block_copy)
2242 {
2243 /* Since ELT is not fully instantiated, we have to leave the
2244 block copy in place. Treat this as a USE. */
2245 scalarize_use (elt, NULL, bsi, is_output, false);
2246 }
2247 else
2248 {
2249 /* The interesting case is when ELT is fully instantiated. In this
2250 case we can have each element stored/loaded directly to/from the
2251 corresponding slot in OTHER. This avoids a block copy. */
2252
2253 tree list = NULL, stmt = bsi_stmt (*bsi);
2254
2255 mark_all_v_defs (stmt);
2256 generate_copy_inout (elt, is_output, other, &list);
2257 mark_all_v_defs (list);
2258 gcc_assert (list);
2259
2260 /* Preserve EH semantics. */
2261 if (stmt_ends_bb_p (stmt))
2262 {
2263 tree_stmt_iterator tsi;
2264 tree first;
2265
2266 /* Extract the first statement from LIST. */
2267 tsi = tsi_start (list);
2268 first = tsi_stmt (tsi);
2269 tsi_delink (&tsi);
2270
2271 /* Replace the old statement with this new representative. */
2272 bsi_replace (bsi, first, true);
2273
2274 if (!tsi_end_p (tsi))
2275 {
2276 /* If any reference would trap, then they all would. And more
2277 to the point, the first would. Therefore none of the rest
2278 will trap since the first didn't. Indicate this by
2279 iterating over the remaining statements and set
2280 TREE_THIS_NOTRAP in all INDIRECT_REFs. */
2281 do
2282 {
2283 walk_tree (tsi_stmt_ptr (tsi), mark_notrap, NULL, NULL);
2284 tsi_next (&tsi);
2285 }
2286 while (!tsi_end_p (tsi));
2287
2288 insert_edge_copies (list, bsi->bb);
2289 }
2290 }
2291 else
2292 sra_replace (bsi, list);
2293 }
2294 }
2295
2296 /* Generate initializations for all scalarizable parameters. */
2297
2298 static void
2299 scalarize_parms (void)
2300 {
2301 tree list = NULL;
2302 unsigned i;
2303 bitmap_iterator bi;
2304
2305 EXECUTE_IF_SET_IN_BITMAP (needs_copy_in, 0, i, bi)
2306 {
2307 tree var = referenced_var (i);
2308 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
2309 generate_copy_inout (elt, true, var, &list);
2310 }
2311
2312 if (list)
2313 {
2314 insert_edge_copies (list, ENTRY_BLOCK_PTR);
2315 mark_all_v_defs (list);
2316 }
2317 }
2318
2319 /* Entry point to phase 4. Update the function to match replacements. */
2320
2321 static void
2322 scalarize_function (void)
2323 {
2324 static const struct sra_walk_fns fns = {
2325 scalarize_use, scalarize_copy, scalarize_init, scalarize_ldst, false
2326 };
2327
2328 sra_walk_function (&fns);
2329 scalarize_parms ();
2330 bsi_commit_edge_inserts ();
2331 }
2332
2333 \f
2334 /* Debug helper function. Print ELT in a nice human-readable format. */
2335
2336 static void
2337 dump_sra_elt_name (FILE *f, struct sra_elt *elt)
2338 {
2339 if (elt->parent && TREE_CODE (elt->parent->type) == COMPLEX_TYPE)
2340 {
2341 fputs (elt->element == integer_zero_node ? "__real__ " : "__imag__ ", f);
2342 dump_sra_elt_name (f, elt->parent);
2343 }
2344 else
2345 {
2346 if (elt->parent)
2347 dump_sra_elt_name (f, elt->parent);
2348 if (DECL_P (elt->element))
2349 {
2350 if (TREE_CODE (elt->element) == FIELD_DECL)
2351 fputc ('.', f);
2352 print_generic_expr (f, elt->element, dump_flags);
2353 }
2354 else if (TREE_CODE (elt->element) == RANGE_EXPR)
2355 fprintf (f, "["HOST_WIDE_INT_PRINT_DEC".."HOST_WIDE_INT_PRINT_DEC"]",
2356 TREE_INT_CST_LOW (TREE_OPERAND (elt->element, 0)),
2357 TREE_INT_CST_LOW (TREE_OPERAND (elt->element, 1)));
2358 else
2359 fprintf (f, "[" HOST_WIDE_INT_PRINT_DEC "]",
2360 TREE_INT_CST_LOW (elt->element));
2361 }
2362 }
2363
2364 /* Likewise, but callable from the debugger. */
2365
2366 void
2367 debug_sra_elt_name (struct sra_elt *elt)
2368 {
2369 dump_sra_elt_name (stderr, elt);
2370 fputc ('\n', stderr);
2371 }
2372
2373 void
2374 sra_init_cache (void)
2375 {
2376 if (sra_type_decomp_cache)
2377 return;
2378
2379 sra_type_decomp_cache = BITMAP_ALLOC (NULL);
2380 sra_type_inst_cache = BITMAP_ALLOC (NULL);
2381 }
2382
2383 /* Main entry point. */
2384
2385 static unsigned int
2386 tree_sra (void)
2387 {
2388 /* Initialize local variables. */
2389 todoflags = 0;
2390 gcc_obstack_init (&sra_obstack);
2391 sra_candidates = BITMAP_ALLOC (NULL);
2392 needs_copy_in = BITMAP_ALLOC (NULL);
2393 sra_init_cache ();
2394 sra_map = htab_create (101, sra_elt_hash, sra_elt_eq, NULL);
2395
2396 /* Scan. If we find anything, instantiate and scalarize. */
2397 if (find_candidates_for_sra ())
2398 {
2399 scan_function ();
2400 decide_instantiations ();
2401 scalarize_function ();
2402 }
2403
2404 /* Free allocated memory. */
2405 htab_delete (sra_map);
2406 sra_map = NULL;
2407 BITMAP_FREE (sra_candidates);
2408 BITMAP_FREE (needs_copy_in);
2409 BITMAP_FREE (sra_type_decomp_cache);
2410 BITMAP_FREE (sra_type_inst_cache);
2411 obstack_free (&sra_obstack, NULL);
2412 return todoflags;
2413 }
2414
2415 static unsigned int
2416 tree_sra_early (void)
2417 {
2418 unsigned int ret;
2419
2420 early_sra = true;
2421 ret = tree_sra ();
2422 early_sra = false;
2423
2424 return ret;
2425 }
2426
2427 static bool
2428 gate_sra (void)
2429 {
2430 return flag_tree_sra != 0;
2431 }
2432
2433 struct tree_opt_pass pass_sra_early =
2434 {
2435 "esra", /* name */
2436 gate_sra, /* gate */
2437 tree_sra_early, /* execute */
2438 NULL, /* sub */
2439 NULL, /* next */
2440 0, /* static_pass_number */
2441 TV_TREE_SRA, /* tv_id */
2442 PROP_cfg | PROP_ssa, /* properties_required */
2443 0, /* properties_provided */
2444 0, /* properties_destroyed */
2445 0, /* todo_flags_start */
2446 TODO_dump_func
2447 | TODO_update_ssa
2448 | TODO_ggc_collect
2449 | TODO_verify_ssa, /* todo_flags_finish */
2450 0 /* letter */
2451 };
2452
2453 struct tree_opt_pass pass_sra =
2454 {
2455 "sra", /* name */
2456 gate_sra, /* gate */
2457 tree_sra, /* execute */
2458 NULL, /* sub */
2459 NULL, /* next */
2460 0, /* static_pass_number */
2461 TV_TREE_SRA, /* tv_id */
2462 PROP_cfg | PROP_ssa, /* properties_required */
2463 0, /* properties_provided */
2464 0, /* properties_destroyed */
2465 0, /* todo_flags_start */
2466 TODO_dump_func
2467 | TODO_update_ssa
2468 | TODO_ggc_collect
2469 | TODO_verify_ssa, /* todo_flags_finish */
2470 0 /* letter */
2471 };