]> git.ipfire.org Git - thirdparty/gcc.git/blob - gcc/tree-sra.c
Change copyright header to refer to version 3 of the GNU General Public License and...
[thirdparty/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 3, 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 COPYING3. If not see
22 <http://www.gnu.org/licenses/>. */
23
24 #include "config.h"
25 #include "system.h"
26 #include "coretypes.h"
27 #include "tm.h"
28 #include "ggc.h"
29 #include "tree.h"
30
31 /* These RTL headers are needed for basic-block.h. */
32 #include "rtl.h"
33 #include "tm_p.h"
34 #include "hard-reg-set.h"
35 #include "basic-block.h"
36 #include "diagnostic.h"
37 #include "langhooks.h"
38 #include "tree-inline.h"
39 #include "tree-flow.h"
40 #include "tree-gimple.h"
41 #include "tree-dump.h"
42 #include "tree-pass.h"
43 #include "timevar.h"
44 #include "flags.h"
45 #include "bitmap.h"
46 #include "obstack.h"
47 #include "target.h"
48 /* expr.h is needed for MOVE_RATIO. */
49 #include "expr.h"
50 #include "params.h"
51
52
53 /* This object of this pass is to replace a non-addressable aggregate with a
54 set of independent variables. Most of the time, all of these variables
55 will be scalars. But a secondary objective is to break up larger
56 aggregates into smaller aggregates. In the process we may find that some
57 bits of the larger aggregate can be deleted as unreferenced.
58
59 This substitution is done globally. More localized substitutions would
60 be the purvey of a load-store motion pass.
61
62 The optimization proceeds in phases:
63
64 (1) Identify variables that have types that are candidates for
65 decomposition.
66
67 (2) Scan the function looking for the ways these variables are used.
68 In particular we're interested in the number of times a variable
69 (or member) is needed as a complete unit, and the number of times
70 a variable (or member) is copied.
71
72 (3) Based on the usage profile, instantiate substitution variables.
73
74 (4) Scan the function making replacements.
75 */
76
77
78 /* True if this is the "early" pass, before inlining. */
79 static bool early_sra;
80
81 /* The set of todo flags to return from tree_sra. */
82 static unsigned int todoflags;
83
84 /* The set of aggregate variables that are candidates for scalarization. */
85 static bitmap sra_candidates;
86
87 /* Set of scalarizable PARM_DECLs that need copy-in operations at the
88 beginning of the function. */
89 static bitmap needs_copy_in;
90
91 /* Sets of bit pairs that cache type decomposition and instantiation. */
92 static bitmap sra_type_decomp_cache;
93 static bitmap sra_type_inst_cache;
94
95 /* One of these structures is created for each candidate aggregate and
96 each (accessed) member or group of members of such an aggregate. */
97 struct sra_elt
98 {
99 /* A tree of the elements. Used when we want to traverse everything. */
100 struct sra_elt *parent;
101 struct sra_elt *groups;
102 struct sra_elt *children;
103 struct sra_elt *sibling;
104
105 /* If this element is a root, then this is the VAR_DECL. If this is
106 a sub-element, this is some token used to identify the reference.
107 In the case of COMPONENT_REF, this is the FIELD_DECL. In the case
108 of an ARRAY_REF, this is the (constant) index. In the case of an
109 ARRAY_RANGE_REF, this is the (constant) RANGE_EXPR. In the case
110 of a complex number, this is a zero or one. */
111 tree element;
112
113 /* The type of the element. */
114 tree type;
115
116 /* A VAR_DECL, for any sub-element we've decided to replace. */
117 tree replacement;
118
119 /* The number of times the element is referenced as a whole. I.e.
120 given "a.b.c", this would be incremented for C, but not for A or B. */
121 unsigned int n_uses;
122
123 /* The number of times the element is copied to or from another
124 scalarizable element. */
125 unsigned int n_copies;
126
127 /* True if TYPE is scalar. */
128 bool is_scalar;
129
130 /* True if this element is a group of members of its parent. */
131 bool is_group;
132
133 /* True if we saw something about this element that prevents scalarization,
134 such as non-constant indexing. */
135 bool cannot_scalarize;
136
137 /* True if we've decided that structure-to-structure assignment
138 should happen via memcpy and not per-element. */
139 bool use_block_copy;
140
141 /* True if everything under this element has been marked TREE_NO_WARNING. */
142 bool all_no_warning;
143
144 /* A flag for use with/after random access traversals. */
145 bool visited;
146
147 /* True if there is BIT_FIELD_REF on the lhs with a vector. */
148 bool is_vector_lhs;
149 };
150
151 #define IS_ELEMENT_FOR_GROUP(ELEMENT) (TREE_CODE (ELEMENT) == RANGE_EXPR)
152
153 #define FOR_EACH_ACTUAL_CHILD(CHILD, ELT) \
154 for ((CHILD) = (ELT)->is_group \
155 ? next_child_for_group (NULL, (ELT)) \
156 : (ELT)->children; \
157 (CHILD); \
158 (CHILD) = (ELT)->is_group \
159 ? next_child_for_group ((CHILD), (ELT)) \
160 : (CHILD)->sibling)
161
162 /* Helper function for above macro. Return next child in group. */
163 static struct sra_elt *
164 next_child_for_group (struct sra_elt *child, struct sra_elt *group)
165 {
166 gcc_assert (group->is_group);
167
168 /* Find the next child in the parent. */
169 if (child)
170 child = child->sibling;
171 else
172 child = group->parent->children;
173
174 /* Skip siblings that do not belong to the group. */
175 while (child)
176 {
177 tree g_elt = group->element;
178 if (TREE_CODE (g_elt) == RANGE_EXPR)
179 {
180 if (!tree_int_cst_lt (child->element, TREE_OPERAND (g_elt, 0))
181 && !tree_int_cst_lt (TREE_OPERAND (g_elt, 1), child->element))
182 break;
183 }
184 else
185 gcc_unreachable ();
186
187 child = child->sibling;
188 }
189
190 return child;
191 }
192
193 /* Random access to the child of a parent is performed by hashing.
194 This prevents quadratic behavior, and allows SRA to function
195 reasonably on larger records. */
196 static htab_t sra_map;
197
198 /* All structures are allocated out of the following obstack. */
199 static struct obstack sra_obstack;
200
201 /* Debugging functions. */
202 static void dump_sra_elt_name (FILE *, struct sra_elt *);
203 extern void debug_sra_elt_name (struct sra_elt *);
204
205 /* Forward declarations. */
206 static tree generate_element_ref (struct sra_elt *);
207 \f
208 /* Return true if DECL is an SRA candidate. */
209
210 static bool
211 is_sra_candidate_decl (tree decl)
212 {
213 return DECL_P (decl) && bitmap_bit_p (sra_candidates, DECL_UID (decl));
214 }
215
216 /* Return true if TYPE is a scalar type. */
217
218 static bool
219 is_sra_scalar_type (tree type)
220 {
221 enum tree_code code = TREE_CODE (type);
222 return (code == INTEGER_TYPE || code == REAL_TYPE || code == VECTOR_TYPE
223 || code == ENUMERAL_TYPE || code == BOOLEAN_TYPE
224 || code == POINTER_TYPE || code == OFFSET_TYPE
225 || code == REFERENCE_TYPE);
226 }
227
228 /* Return true if TYPE can be decomposed into a set of independent variables.
229
230 Note that this doesn't imply that all elements of TYPE can be
231 instantiated, just that if we decide to break up the type into
232 separate pieces that it can be done. */
233
234 bool
235 sra_type_can_be_decomposed_p (tree type)
236 {
237 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
238 tree t;
239
240 /* Avoid searching the same type twice. */
241 if (bitmap_bit_p (sra_type_decomp_cache, cache+0))
242 return true;
243 if (bitmap_bit_p (sra_type_decomp_cache, cache+1))
244 return false;
245
246 /* The type must have a definite nonzero size. */
247 if (TYPE_SIZE (type) == NULL || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST
248 || integer_zerop (TYPE_SIZE (type)))
249 goto fail;
250
251 /* The type must be a non-union aggregate. */
252 switch (TREE_CODE (type))
253 {
254 case RECORD_TYPE:
255 {
256 bool saw_one_field = false;
257
258 for (t = TYPE_FIELDS (type); t ; t = TREE_CHAIN (t))
259 if (TREE_CODE (t) == FIELD_DECL)
260 {
261 /* Reject incorrectly represented bit fields. */
262 if (DECL_BIT_FIELD (t)
263 && (tree_low_cst (DECL_SIZE (t), 1)
264 != TYPE_PRECISION (TREE_TYPE (t))))
265 goto fail;
266
267 saw_one_field = true;
268 }
269
270 /* Record types must have at least one field. */
271 if (!saw_one_field)
272 goto fail;
273 }
274 break;
275
276 case ARRAY_TYPE:
277 /* Array types must have a fixed lower and upper bound. */
278 t = TYPE_DOMAIN (type);
279 if (t == NULL)
280 goto fail;
281 if (TYPE_MIN_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MIN_VALUE (t)))
282 goto fail;
283 if (TYPE_MAX_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MAX_VALUE (t)))
284 goto fail;
285 break;
286
287 case COMPLEX_TYPE:
288 break;
289
290 default:
291 goto fail;
292 }
293
294 bitmap_set_bit (sra_type_decomp_cache, cache+0);
295 return true;
296
297 fail:
298 bitmap_set_bit (sra_type_decomp_cache, cache+1);
299 return false;
300 }
301
302 /* Return true if DECL can be decomposed into a set of independent
303 (though not necessarily scalar) variables. */
304
305 static bool
306 decl_can_be_decomposed_p (tree var)
307 {
308 /* Early out for scalars. */
309 if (is_sra_scalar_type (TREE_TYPE (var)))
310 return false;
311
312 /* The variable must not be aliased. */
313 if (!is_gimple_non_addressable (var))
314 {
315 if (dump_file && (dump_flags & TDF_DETAILS))
316 {
317 fprintf (dump_file, "Cannot scalarize variable ");
318 print_generic_expr (dump_file, var, dump_flags);
319 fprintf (dump_file, " because it must live in memory\n");
320 }
321 return false;
322 }
323
324 /* The variable must not be volatile. */
325 if (TREE_THIS_VOLATILE (var))
326 {
327 if (dump_file && (dump_flags & TDF_DETAILS))
328 {
329 fprintf (dump_file, "Cannot scalarize variable ");
330 print_generic_expr (dump_file, var, dump_flags);
331 fprintf (dump_file, " because it is declared volatile\n");
332 }
333 return false;
334 }
335
336 /* We must be able to decompose the variable's type. */
337 if (!sra_type_can_be_decomposed_p (TREE_TYPE (var)))
338 {
339 if (dump_file && (dump_flags & TDF_DETAILS))
340 {
341 fprintf (dump_file, "Cannot scalarize variable ");
342 print_generic_expr (dump_file, var, dump_flags);
343 fprintf (dump_file, " because its type cannot be decomposed\n");
344 }
345 return false;
346 }
347
348 /* HACK: if we decompose a va_list_type_node before inlining, then we'll
349 confuse tree-stdarg.c, and we won't be able to figure out which and
350 how many arguments are accessed. This really should be improved in
351 tree-stdarg.c, as the decomposition is truely a win. This could also
352 be fixed if the stdarg pass ran early, but this can't be done until
353 we've aliasing information early too. See PR 30791. */
354 if (early_sra
355 && TYPE_MAIN_VARIANT (TREE_TYPE (var))
356 == TYPE_MAIN_VARIANT (va_list_type_node))
357 return false;
358
359 return true;
360 }
361
362 /* Return true if TYPE can be *completely* decomposed into scalars. */
363
364 static bool
365 type_can_instantiate_all_elements (tree type)
366 {
367 if (is_sra_scalar_type (type))
368 return true;
369 if (!sra_type_can_be_decomposed_p (type))
370 return false;
371
372 switch (TREE_CODE (type))
373 {
374 case RECORD_TYPE:
375 {
376 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
377 tree f;
378
379 if (bitmap_bit_p (sra_type_inst_cache, cache+0))
380 return true;
381 if (bitmap_bit_p (sra_type_inst_cache, cache+1))
382 return false;
383
384 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
385 if (TREE_CODE (f) == FIELD_DECL)
386 {
387 if (!type_can_instantiate_all_elements (TREE_TYPE (f)))
388 {
389 bitmap_set_bit (sra_type_inst_cache, cache+1);
390 return false;
391 }
392 }
393
394 bitmap_set_bit (sra_type_inst_cache, cache+0);
395 return true;
396 }
397
398 case ARRAY_TYPE:
399 return type_can_instantiate_all_elements (TREE_TYPE (type));
400
401 case COMPLEX_TYPE:
402 return true;
403
404 default:
405 gcc_unreachable ();
406 }
407 }
408
409 /* Test whether ELT or some sub-element cannot be scalarized. */
410
411 static bool
412 can_completely_scalarize_p (struct sra_elt *elt)
413 {
414 struct sra_elt *c;
415
416 if (elt->cannot_scalarize)
417 return false;
418
419 for (c = elt->children; c; c = c->sibling)
420 if (!can_completely_scalarize_p (c))
421 return false;
422
423 for (c = elt->groups; c; c = c->sibling)
424 if (!can_completely_scalarize_p (c))
425 return false;
426
427 return true;
428 }
429
430 \f
431 /* A simplified tree hashing algorithm that only handles the types of
432 trees we expect to find in sra_elt->element. */
433
434 static hashval_t
435 sra_hash_tree (tree t)
436 {
437 hashval_t h;
438
439 switch (TREE_CODE (t))
440 {
441 case VAR_DECL:
442 case PARM_DECL:
443 case RESULT_DECL:
444 h = DECL_UID (t);
445 break;
446
447 case INTEGER_CST:
448 h = TREE_INT_CST_LOW (t) ^ TREE_INT_CST_HIGH (t);
449 break;
450
451 case RANGE_EXPR:
452 h = iterative_hash_expr (TREE_OPERAND (t, 0), 0);
453 h = iterative_hash_expr (TREE_OPERAND (t, 1), h);
454 break;
455
456 case FIELD_DECL:
457 /* We can have types that are compatible, but have different member
458 lists, so we can't hash fields by ID. Use offsets instead. */
459 h = iterative_hash_expr (DECL_FIELD_OFFSET (t), 0);
460 h = iterative_hash_expr (DECL_FIELD_BIT_OFFSET (t), h);
461 break;
462
463 default:
464 gcc_unreachable ();
465 }
466
467 return h;
468 }
469
470 /* Hash function for type SRA_PAIR. */
471
472 static hashval_t
473 sra_elt_hash (const void *x)
474 {
475 const struct sra_elt *e = x;
476 const struct sra_elt *p;
477 hashval_t h;
478
479 h = sra_hash_tree (e->element);
480
481 /* Take into account everything back up the chain. Given that chain
482 lengths are rarely very long, this should be acceptable. If we
483 truly identify this as a performance problem, it should work to
484 hash the pointer value "e->parent". */
485 for (p = e->parent; p ; p = p->parent)
486 h = (h * 65521) ^ sra_hash_tree (p->element);
487
488 return h;
489 }
490
491 /* Equality function for type SRA_PAIR. */
492
493 static int
494 sra_elt_eq (const void *x, const void *y)
495 {
496 const struct sra_elt *a = x;
497 const struct sra_elt *b = y;
498 tree ae, be;
499
500 if (a->parent != b->parent)
501 return false;
502
503 ae = a->element;
504 be = b->element;
505
506 if (ae == be)
507 return true;
508 if (TREE_CODE (ae) != TREE_CODE (be))
509 return false;
510
511 switch (TREE_CODE (ae))
512 {
513 case VAR_DECL:
514 case PARM_DECL:
515 case RESULT_DECL:
516 /* These are all pointer unique. */
517 return false;
518
519 case INTEGER_CST:
520 /* Integers are not pointer unique, so compare their values. */
521 return tree_int_cst_equal (ae, be);
522
523 case RANGE_EXPR:
524 return
525 tree_int_cst_equal (TREE_OPERAND (ae, 0), TREE_OPERAND (be, 0))
526 && tree_int_cst_equal (TREE_OPERAND (ae, 1), TREE_OPERAND (be, 1));
527
528 case FIELD_DECL:
529 /* Fields are unique within a record, but not between
530 compatible records. */
531 if (DECL_FIELD_CONTEXT (ae) == DECL_FIELD_CONTEXT (be))
532 return false;
533 return fields_compatible_p (ae, be);
534
535 default:
536 gcc_unreachable ();
537 }
538 }
539
540 /* Create or return the SRA_ELT structure for CHILD in PARENT. PARENT
541 may be null, in which case CHILD must be a DECL. */
542
543 static struct sra_elt *
544 lookup_element (struct sra_elt *parent, tree child, tree type,
545 enum insert_option insert)
546 {
547 struct sra_elt dummy;
548 struct sra_elt **slot;
549 struct sra_elt *elt;
550
551 if (parent)
552 dummy.parent = parent->is_group ? parent->parent : parent;
553 else
554 dummy.parent = NULL;
555 dummy.element = child;
556
557 slot = (struct sra_elt **) htab_find_slot (sra_map, &dummy, insert);
558 if (!slot && insert == NO_INSERT)
559 return NULL;
560
561 elt = *slot;
562 if (!elt && insert == INSERT)
563 {
564 *slot = elt = obstack_alloc (&sra_obstack, sizeof (*elt));
565 memset (elt, 0, sizeof (*elt));
566
567 elt->parent = parent;
568 elt->element = child;
569 elt->type = type;
570 elt->is_scalar = is_sra_scalar_type (type);
571
572 if (parent)
573 {
574 if (IS_ELEMENT_FOR_GROUP (elt->element))
575 {
576 elt->is_group = true;
577 elt->sibling = parent->groups;
578 parent->groups = elt;
579 }
580 else
581 {
582 elt->sibling = parent->children;
583 parent->children = elt;
584 }
585 }
586
587 /* If this is a parameter, then if we want to scalarize, we have
588 one copy from the true function parameter. Count it now. */
589 if (TREE_CODE (child) == PARM_DECL)
590 {
591 elt->n_copies = 1;
592 bitmap_set_bit (needs_copy_in, DECL_UID (child));
593 }
594 }
595
596 return elt;
597 }
598
599 /* Create or return the SRA_ELT structure for EXPR if the expression
600 refers to a scalarizable variable. */
601
602 static struct sra_elt *
603 maybe_lookup_element_for_expr (tree expr)
604 {
605 struct sra_elt *elt;
606 tree child;
607
608 switch (TREE_CODE (expr))
609 {
610 case VAR_DECL:
611 case PARM_DECL:
612 case RESULT_DECL:
613 if (is_sra_candidate_decl (expr))
614 return lookup_element (NULL, expr, TREE_TYPE (expr), INSERT);
615 return NULL;
616
617 case ARRAY_REF:
618 /* We can't scalarize variable array indices. */
619 if (in_array_bounds_p (expr))
620 child = TREE_OPERAND (expr, 1);
621 else
622 return NULL;
623 break;
624
625 case ARRAY_RANGE_REF:
626 /* We can't scalarize variable array indices. */
627 if (range_in_array_bounds_p (expr))
628 {
629 tree domain = TYPE_DOMAIN (TREE_TYPE (expr));
630 child = build2 (RANGE_EXPR, integer_type_node,
631 TYPE_MIN_VALUE (domain), TYPE_MAX_VALUE (domain));
632 }
633 else
634 return NULL;
635 break;
636
637 case COMPONENT_REF:
638 /* Don't look through unions. */
639 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (expr, 0))) != RECORD_TYPE)
640 return NULL;
641 child = TREE_OPERAND (expr, 1);
642 break;
643
644 case REALPART_EXPR:
645 child = integer_zero_node;
646 break;
647 case IMAGPART_EXPR:
648 child = integer_one_node;
649 break;
650
651 default:
652 return NULL;
653 }
654
655 elt = maybe_lookup_element_for_expr (TREE_OPERAND (expr, 0));
656 if (elt)
657 return lookup_element (elt, child, TREE_TYPE (expr), INSERT);
658 return NULL;
659 }
660
661 \f
662 /* Functions to walk just enough of the tree to see all scalarizable
663 references, and categorize them. */
664
665 /* A set of callbacks for phases 2 and 4. They'll be invoked for the
666 various kinds of references seen. In all cases, *BSI is an iterator
667 pointing to the statement being processed. */
668 struct sra_walk_fns
669 {
670 /* Invoked when ELT is required as a unit. Note that ELT might refer to
671 a leaf node, in which case this is a simple scalar reference. *EXPR_P
672 points to the location of the expression. IS_OUTPUT is true if this
673 is a left-hand-side reference. USE_ALL is true if we saw something we
674 couldn't quite identify and had to force the use of the entire object. */
675 void (*use) (struct sra_elt *elt, tree *expr_p,
676 block_stmt_iterator *bsi, bool is_output, bool use_all);
677
678 /* Invoked when we have a copy between two scalarizable references. */
679 void (*copy) (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
680 block_stmt_iterator *bsi);
681
682 /* Invoked when ELT is initialized from a constant. VALUE may be NULL,
683 in which case it should be treated as an empty CONSTRUCTOR. */
684 void (*init) (struct sra_elt *elt, tree value, block_stmt_iterator *bsi);
685
686 /* Invoked when we have a copy between one scalarizable reference ELT
687 and one non-scalarizable reference OTHER without side-effects.
688 IS_OUTPUT is true if ELT is on the left-hand side. */
689 void (*ldst) (struct sra_elt *elt, tree other,
690 block_stmt_iterator *bsi, bool is_output);
691
692 /* True during phase 2, false during phase 4. */
693 /* ??? This is a hack. */
694 bool initial_scan;
695 };
696
697 #ifdef ENABLE_CHECKING
698 /* Invoked via walk_tree, if *TP contains a candidate decl, return it. */
699
700 static tree
701 sra_find_candidate_decl (tree *tp, int *walk_subtrees,
702 void *data ATTRIBUTE_UNUSED)
703 {
704 tree t = *tp;
705 enum tree_code code = TREE_CODE (t);
706
707 if (code == VAR_DECL || code == PARM_DECL || code == RESULT_DECL)
708 {
709 *walk_subtrees = 0;
710 if (is_sra_candidate_decl (t))
711 return t;
712 }
713 else if (TYPE_P (t))
714 *walk_subtrees = 0;
715
716 return NULL;
717 }
718 #endif
719
720 /* Walk most expressions looking for a scalarizable aggregate.
721 If we find one, invoke FNS->USE. */
722
723 static void
724 sra_walk_expr (tree *expr_p, block_stmt_iterator *bsi, bool is_output,
725 const struct sra_walk_fns *fns)
726 {
727 tree expr = *expr_p;
728 tree inner = expr;
729 bool disable_scalarization = false;
730 bool use_all_p = false;
731
732 /* We're looking to collect a reference expression between EXPR and INNER,
733 such that INNER is a scalarizable decl and all other nodes through EXPR
734 are references that we can scalarize. If we come across something that
735 we can't scalarize, we reset EXPR. This has the effect of making it
736 appear that we're referring to the larger expression as a whole. */
737
738 while (1)
739 switch (TREE_CODE (inner))
740 {
741 case VAR_DECL:
742 case PARM_DECL:
743 case RESULT_DECL:
744 /* If there is a scalarizable decl at the bottom, then process it. */
745 if (is_sra_candidate_decl (inner))
746 {
747 struct sra_elt *elt = maybe_lookup_element_for_expr (expr);
748 if (disable_scalarization)
749 elt->cannot_scalarize = true;
750 else
751 fns->use (elt, expr_p, bsi, is_output, use_all_p);
752 }
753 return;
754
755 case ARRAY_REF:
756 /* Non-constant index means any member may be accessed. Prevent the
757 expression from being scalarized. If we were to treat this as a
758 reference to the whole array, we can wind up with a single dynamic
759 index reference inside a loop being overridden by several constant
760 index references during loop setup. It's possible that this could
761 be avoided by using dynamic usage counts based on BB trip counts
762 (based on loop analysis or profiling), but that hardly seems worth
763 the effort. */
764 /* ??? Hack. Figure out how to push this into the scan routines
765 without duplicating too much code. */
766 if (!in_array_bounds_p (inner))
767 {
768 disable_scalarization = true;
769 goto use_all;
770 }
771 /* ??? Are we assured that non-constant bounds and stride will have
772 the same value everywhere? I don't think Fortran will... */
773 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
774 goto use_all;
775 inner = TREE_OPERAND (inner, 0);
776 break;
777
778 case ARRAY_RANGE_REF:
779 if (!range_in_array_bounds_p (inner))
780 {
781 disable_scalarization = true;
782 goto use_all;
783 }
784 /* ??? See above non-constant bounds and stride . */
785 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
786 goto use_all;
787 inner = TREE_OPERAND (inner, 0);
788 break;
789
790 case COMPONENT_REF:
791 /* A reference to a union member constitutes a reference to the
792 entire union. */
793 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) != RECORD_TYPE)
794 goto use_all;
795 /* ??? See above re non-constant stride. */
796 if (TREE_OPERAND (inner, 2))
797 goto use_all;
798 inner = TREE_OPERAND (inner, 0);
799 break;
800
801 case REALPART_EXPR:
802 case IMAGPART_EXPR:
803 inner = TREE_OPERAND (inner, 0);
804 break;
805
806 case BIT_FIELD_REF:
807 /* A bit field reference to a specific vector is scalarized but for
808 ones for inputs need to be marked as used on the left hand size so
809 when we scalarize it, we can mark that variable as non renamable. */
810 if (is_output
811 && TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) == VECTOR_TYPE)
812 {
813 struct sra_elt *elt
814 = maybe_lookup_element_for_expr (TREE_OPERAND (inner, 0));
815 if (elt)
816 elt->is_vector_lhs = true;
817 }
818 /* A bit field reference (access to *multiple* fields simultaneously)
819 is not currently scalarized. Consider this an access to the
820 complete outer element, to which walk_tree will bring us next. */
821
822 goto use_all;
823
824 case VIEW_CONVERT_EXPR:
825 case NOP_EXPR:
826 /* Similarly, a view/nop explicitly wants to look at an object in a
827 type other than the one we've scalarized. */
828 goto use_all;
829
830 case WITH_SIZE_EXPR:
831 /* This is a transparent wrapper. The entire inner expression really
832 is being used. */
833 goto use_all;
834
835 use_all:
836 expr_p = &TREE_OPERAND (inner, 0);
837 inner = expr = *expr_p;
838 use_all_p = true;
839 break;
840
841 default:
842 #ifdef ENABLE_CHECKING
843 /* Validate that we're not missing any references. */
844 gcc_assert (!walk_tree (&inner, sra_find_candidate_decl, NULL, NULL));
845 #endif
846 return;
847 }
848 }
849
850 /* Walk a TREE_LIST of values looking for scalarizable aggregates.
851 If we find one, invoke FNS->USE. */
852
853 static void
854 sra_walk_tree_list (tree list, block_stmt_iterator *bsi, bool is_output,
855 const struct sra_walk_fns *fns)
856 {
857 tree op;
858 for (op = list; op ; op = TREE_CHAIN (op))
859 sra_walk_expr (&TREE_VALUE (op), bsi, is_output, fns);
860 }
861
862 /* Walk the arguments of a CALL_EXPR looking for scalarizable aggregates.
863 If we find one, invoke FNS->USE. */
864
865 static void
866 sra_walk_call_expr (tree expr, block_stmt_iterator *bsi,
867 const struct sra_walk_fns *fns)
868 {
869 int i;
870 int nargs = call_expr_nargs (expr);
871 for (i = 0; i < nargs; i++)
872 sra_walk_expr (&CALL_EXPR_ARG (expr, i), bsi, false, fns);
873 }
874
875 /* Walk the inputs and outputs of an ASM_EXPR looking for scalarizable
876 aggregates. If we find one, invoke FNS->USE. */
877
878 static void
879 sra_walk_asm_expr (tree expr, block_stmt_iterator *bsi,
880 const struct sra_walk_fns *fns)
881 {
882 sra_walk_tree_list (ASM_INPUTS (expr), bsi, false, fns);
883 sra_walk_tree_list (ASM_OUTPUTS (expr), bsi, true, fns);
884 }
885
886 /* Walk a GIMPLE_MODIFY_STMT and categorize the assignment appropriately. */
887
888 static void
889 sra_walk_gimple_modify_stmt (tree expr, block_stmt_iterator *bsi,
890 const struct sra_walk_fns *fns)
891 {
892 struct sra_elt *lhs_elt, *rhs_elt;
893 tree lhs, rhs;
894
895 lhs = GIMPLE_STMT_OPERAND (expr, 0);
896 rhs = GIMPLE_STMT_OPERAND (expr, 1);
897 lhs_elt = maybe_lookup_element_for_expr (lhs);
898 rhs_elt = maybe_lookup_element_for_expr (rhs);
899
900 /* If both sides are scalarizable, this is a COPY operation. */
901 if (lhs_elt && rhs_elt)
902 {
903 fns->copy (lhs_elt, rhs_elt, bsi);
904 return;
905 }
906
907 /* If the RHS is scalarizable, handle it. There are only two cases. */
908 if (rhs_elt)
909 {
910 if (!rhs_elt->is_scalar && !TREE_SIDE_EFFECTS (lhs))
911 fns->ldst (rhs_elt, lhs, bsi, false);
912 else
913 fns->use (rhs_elt, &GIMPLE_STMT_OPERAND (expr, 1), bsi, false, false);
914 }
915
916 /* If it isn't scalarizable, there may be scalarizable variables within, so
917 check for a call or else walk the RHS to see if we need to do any
918 copy-in operations. We need to do it before the LHS is scalarized so
919 that the statements get inserted in the proper place, before any
920 copy-out operations. */
921 else
922 {
923 tree call = get_call_expr_in (rhs);
924 if (call)
925 sra_walk_call_expr (call, bsi, fns);
926 else
927 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 1), bsi, false, fns);
928 }
929
930 /* Likewise, handle the LHS being scalarizable. We have cases similar
931 to those above, but also want to handle RHS being constant. */
932 if (lhs_elt)
933 {
934 /* If this is an assignment from a constant, or constructor, then
935 we have access to all of the elements individually. Invoke INIT. */
936 if (TREE_CODE (rhs) == COMPLEX_EXPR
937 || TREE_CODE (rhs) == COMPLEX_CST
938 || TREE_CODE (rhs) == CONSTRUCTOR)
939 fns->init (lhs_elt, rhs, bsi);
940
941 /* If this is an assignment from read-only memory, treat this as if
942 we'd been passed the constructor directly. Invoke INIT. */
943 else if (TREE_CODE (rhs) == VAR_DECL
944 && TREE_STATIC (rhs)
945 && TREE_READONLY (rhs)
946 && targetm.binds_local_p (rhs))
947 fns->init (lhs_elt, DECL_INITIAL (rhs), bsi);
948
949 /* If this is a copy from a non-scalarizable lvalue, invoke LDST.
950 The lvalue requirement prevents us from trying to directly scalarize
951 the result of a function call. Which would result in trying to call
952 the function multiple times, and other evil things. */
953 else if (!lhs_elt->is_scalar
954 && !TREE_SIDE_EFFECTS (rhs) && 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 if (elt->element && TREE_NO_WARNING (elt->element))
1244 TREE_NO_WARNING (var) = 1;
1245 }
1246 else
1247 {
1248 DECL_IGNORED_P (var) = 1;
1249 /* ??? We can't generate any warning that would be meaningful. */
1250 TREE_NO_WARNING (var) = 1;
1251 }
1252
1253 if (dump_file)
1254 {
1255 fputs (" ", dump_file);
1256 dump_sra_elt_name (dump_file, elt);
1257 fputs (" -> ", dump_file);
1258 print_generic_expr (dump_file, var, dump_flags);
1259 fputc ('\n', dump_file);
1260 }
1261 }
1262
1263 /* Make one pass across an element tree deciding whether or not it's
1264 profitable to instantiate individual leaf scalars.
1265
1266 PARENT_USES and PARENT_COPIES are the sum of the N_USES and N_COPIES
1267 fields all the way up the tree. */
1268
1269 static void
1270 decide_instantiation_1 (struct sra_elt *elt, unsigned int parent_uses,
1271 unsigned int parent_copies)
1272 {
1273 if (dump_file && !elt->parent)
1274 {
1275 fputs ("Initial instantiation for ", dump_file);
1276 dump_sra_elt_name (dump_file, elt);
1277 fputc ('\n', dump_file);
1278 }
1279
1280 if (elt->cannot_scalarize)
1281 return;
1282
1283 if (elt->is_scalar)
1284 {
1285 /* The decision is simple: instantiate if we're used more frequently
1286 than the parent needs to be seen as a complete unit. */
1287 if (elt->n_uses + elt->n_copies + parent_copies > parent_uses)
1288 instantiate_element (elt);
1289 }
1290 else
1291 {
1292 struct sra_elt *c, *group;
1293 unsigned int this_uses = elt->n_uses + parent_uses;
1294 unsigned int this_copies = elt->n_copies + parent_copies;
1295
1296 /* Consider groups of sub-elements as weighing in favour of
1297 instantiation whatever their size. */
1298 for (group = elt->groups; group ; group = group->sibling)
1299 FOR_EACH_ACTUAL_CHILD (c, group)
1300 {
1301 c->n_uses += group->n_uses;
1302 c->n_copies += group->n_copies;
1303 }
1304
1305 for (c = elt->children; c ; c = c->sibling)
1306 decide_instantiation_1 (c, this_uses, this_copies);
1307 }
1308 }
1309
1310 /* Compute the size and number of all instantiated elements below ELT.
1311 We will only care about this if the size of the complete structure
1312 fits in a HOST_WIDE_INT, so we don't have to worry about overflow. */
1313
1314 static unsigned int
1315 sum_instantiated_sizes (struct sra_elt *elt, unsigned HOST_WIDE_INT *sizep)
1316 {
1317 if (elt->replacement)
1318 {
1319 *sizep += TREE_INT_CST_LOW (TYPE_SIZE_UNIT (elt->type));
1320 return 1;
1321 }
1322 else
1323 {
1324 struct sra_elt *c;
1325 unsigned int count = 0;
1326
1327 for (c = elt->children; c ; c = c->sibling)
1328 count += sum_instantiated_sizes (c, sizep);
1329
1330 return count;
1331 }
1332 }
1333
1334 /* Instantiate fields in ELT->TYPE that are not currently present as
1335 children of ELT. */
1336
1337 static void instantiate_missing_elements (struct sra_elt *elt);
1338
1339 static void
1340 instantiate_missing_elements_1 (struct sra_elt *elt, tree child, tree type)
1341 {
1342 struct sra_elt *sub = lookup_element (elt, child, type, INSERT);
1343 if (sub->is_scalar)
1344 {
1345 if (sub->replacement == NULL)
1346 instantiate_element (sub);
1347 }
1348 else
1349 instantiate_missing_elements (sub);
1350 }
1351
1352 static void
1353 instantiate_missing_elements (struct sra_elt *elt)
1354 {
1355 tree type = elt->type;
1356
1357 switch (TREE_CODE (type))
1358 {
1359 case RECORD_TYPE:
1360 {
1361 tree f;
1362 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
1363 if (TREE_CODE (f) == FIELD_DECL)
1364 {
1365 tree field_type = TREE_TYPE (f);
1366
1367 /* canonicalize_component_ref() unwidens some bit-field
1368 types (not marked as DECL_BIT_FIELD in C++), so we
1369 must do the same, lest we may introduce type
1370 mismatches. */
1371 if (INTEGRAL_TYPE_P (field_type)
1372 && DECL_MODE (f) != TYPE_MODE (field_type))
1373 field_type = TREE_TYPE (get_unwidened (build3 (COMPONENT_REF,
1374 field_type,
1375 elt->element,
1376 f, NULL_TREE),
1377 NULL_TREE));
1378
1379 instantiate_missing_elements_1 (elt, f, field_type);
1380 }
1381 break;
1382 }
1383
1384 case ARRAY_TYPE:
1385 {
1386 tree i, max, subtype;
1387
1388 i = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1389 max = TYPE_MAX_VALUE (TYPE_DOMAIN (type));
1390 subtype = TREE_TYPE (type);
1391
1392 while (1)
1393 {
1394 instantiate_missing_elements_1 (elt, i, subtype);
1395 if (tree_int_cst_equal (i, max))
1396 break;
1397 i = int_const_binop (PLUS_EXPR, i, integer_one_node, true);
1398 }
1399
1400 break;
1401 }
1402
1403 case COMPLEX_TYPE:
1404 type = TREE_TYPE (type);
1405 instantiate_missing_elements_1 (elt, integer_zero_node, type);
1406 instantiate_missing_elements_1 (elt, integer_one_node, type);
1407 break;
1408
1409 default:
1410 gcc_unreachable ();
1411 }
1412 }
1413
1414 /* Return true if there is only one non aggregate field in the record, TYPE.
1415 Return false otherwise. */
1416
1417 static bool
1418 single_scalar_field_in_record_p (tree type)
1419 {
1420 int num_fields = 0;
1421 tree field;
1422 if (TREE_CODE (type) != RECORD_TYPE)
1423 return false;
1424
1425 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
1426 if (TREE_CODE (field) == FIELD_DECL)
1427 {
1428 num_fields++;
1429
1430 if (num_fields == 2)
1431 return false;
1432
1433 if (AGGREGATE_TYPE_P (TREE_TYPE (field)))
1434 return false;
1435 }
1436
1437 return true;
1438 }
1439
1440 /* Make one pass across an element tree deciding whether to perform block
1441 or element copies. If we decide on element copies, instantiate all
1442 elements. Return true if there are any instantiated sub-elements. */
1443
1444 static bool
1445 decide_block_copy (struct sra_elt *elt)
1446 {
1447 struct sra_elt *c;
1448 bool any_inst;
1449
1450 /* We shouldn't be invoked on groups of sub-elements as they must
1451 behave like their parent as far as block copy is concerned. */
1452 gcc_assert (!elt->is_group);
1453
1454 /* If scalarization is disabled, respect it. */
1455 if (elt->cannot_scalarize)
1456 {
1457 elt->use_block_copy = 1;
1458
1459 if (dump_file)
1460 {
1461 fputs ("Scalarization disabled for ", dump_file);
1462 dump_sra_elt_name (dump_file, elt);
1463 fputc ('\n', dump_file);
1464 }
1465
1466 /* Disable scalarization of sub-elements */
1467 for (c = elt->children; c; c = c->sibling)
1468 {
1469 c->cannot_scalarize = 1;
1470 decide_block_copy (c);
1471 }
1472
1473 /* Groups behave like their parent. */
1474 for (c = elt->groups; c; c = c->sibling)
1475 {
1476 c->cannot_scalarize = 1;
1477 c->use_block_copy = 1;
1478 }
1479
1480 return false;
1481 }
1482
1483 /* Don't decide if we've no uses. */
1484 if (elt->n_uses == 0 && elt->n_copies == 0)
1485 ;
1486
1487 else if (!elt->is_scalar)
1488 {
1489 tree size_tree = TYPE_SIZE_UNIT (elt->type);
1490 bool use_block_copy = true;
1491
1492 /* Tradeoffs for COMPLEX types pretty much always make it better
1493 to go ahead and split the components. */
1494 if (TREE_CODE (elt->type) == COMPLEX_TYPE)
1495 use_block_copy = false;
1496
1497 /* Don't bother trying to figure out the rest if the structure is
1498 so large we can't do easy arithmetic. This also forces block
1499 copies for variable sized structures. */
1500 else if (host_integerp (size_tree, 1))
1501 {
1502 unsigned HOST_WIDE_INT full_size, inst_size = 0;
1503 unsigned int max_size, max_count, inst_count, full_count;
1504
1505 /* If the sra-max-structure-size parameter is 0, then the
1506 user has not overridden the parameter and we can choose a
1507 sensible default. */
1508 max_size = SRA_MAX_STRUCTURE_SIZE
1509 ? SRA_MAX_STRUCTURE_SIZE
1510 : MOVE_RATIO * UNITS_PER_WORD;
1511 max_count = SRA_MAX_STRUCTURE_COUNT
1512 ? SRA_MAX_STRUCTURE_COUNT
1513 : MOVE_RATIO;
1514
1515 full_size = tree_low_cst (size_tree, 1);
1516 full_count = count_type_elements (elt->type, false);
1517 inst_count = sum_instantiated_sizes (elt, &inst_size);
1518
1519 /* If there is only one scalar field in the record, don't block copy. */
1520 if (single_scalar_field_in_record_p (elt->type))
1521 use_block_copy = false;
1522
1523 /* ??? What to do here. If there are two fields, and we've only
1524 instantiated one, then instantiating the other is clearly a win.
1525 If there are a large number of fields then the size of the copy
1526 is much more of a factor. */
1527
1528 /* If the structure is small, and we've made copies, go ahead
1529 and instantiate, hoping that the copies will go away. */
1530 if (full_size <= max_size
1531 && (full_count - inst_count) <= max_count
1532 && elt->n_copies > elt->n_uses)
1533 use_block_copy = false;
1534 else if (inst_count * 100 >= full_count * SRA_FIELD_STRUCTURE_RATIO
1535 && inst_size * 100 >= full_size * SRA_FIELD_STRUCTURE_RATIO)
1536 use_block_copy = false;
1537
1538 /* In order to avoid block copy, we have to be able to instantiate
1539 all elements of the type. See if this is possible. */
1540 if (!use_block_copy
1541 && (!can_completely_scalarize_p (elt)
1542 || !type_can_instantiate_all_elements (elt->type)))
1543 use_block_copy = true;
1544 }
1545
1546 elt->use_block_copy = use_block_copy;
1547
1548 /* Groups behave like their parent. */
1549 for (c = elt->groups; c; c = c->sibling)
1550 c->use_block_copy = use_block_copy;
1551
1552 if (dump_file)
1553 {
1554 fprintf (dump_file, "Using %s for ",
1555 use_block_copy ? "block-copy" : "element-copy");
1556 dump_sra_elt_name (dump_file, elt);
1557 fputc ('\n', dump_file);
1558 }
1559
1560 if (!use_block_copy)
1561 {
1562 instantiate_missing_elements (elt);
1563 return true;
1564 }
1565 }
1566
1567 any_inst = elt->replacement != NULL;
1568
1569 for (c = elt->children; c ; c = c->sibling)
1570 any_inst |= decide_block_copy (c);
1571
1572 return any_inst;
1573 }
1574
1575 /* Entry point to phase 3. Instantiate scalar replacement variables. */
1576
1577 static void
1578 decide_instantiations (void)
1579 {
1580 unsigned int i;
1581 bool cleared_any;
1582 bitmap_head done_head;
1583 bitmap_iterator bi;
1584
1585 /* We cannot clear bits from a bitmap we're iterating over,
1586 so save up all the bits to clear until the end. */
1587 bitmap_initialize (&done_head, &bitmap_default_obstack);
1588 cleared_any = false;
1589
1590 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1591 {
1592 tree var = referenced_var (i);
1593 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1594 if (elt)
1595 {
1596 decide_instantiation_1 (elt, 0, 0);
1597 if (!decide_block_copy (elt))
1598 elt = NULL;
1599 }
1600 if (!elt)
1601 {
1602 bitmap_set_bit (&done_head, i);
1603 cleared_any = true;
1604 }
1605 }
1606
1607 if (cleared_any)
1608 {
1609 bitmap_and_compl_into (sra_candidates, &done_head);
1610 bitmap_and_compl_into (needs_copy_in, &done_head);
1611 }
1612 bitmap_clear (&done_head);
1613
1614 mark_set_for_renaming (sra_candidates);
1615
1616 if (dump_file)
1617 fputc ('\n', dump_file);
1618 }
1619
1620 \f
1621 /* Phase Four: Update the function to match the replacements created. */
1622
1623 /* Mark all the variables in VDEF/VUSE operators for STMT for
1624 renaming. This becomes necessary when we modify all of a
1625 non-scalar. */
1626
1627 static void
1628 mark_all_v_defs_1 (tree stmt)
1629 {
1630 tree sym;
1631 ssa_op_iter iter;
1632
1633 update_stmt_if_modified (stmt);
1634
1635 FOR_EACH_SSA_TREE_OPERAND (sym, stmt, iter, SSA_OP_ALL_VIRTUALS)
1636 {
1637 if (TREE_CODE (sym) == SSA_NAME)
1638 sym = SSA_NAME_VAR (sym);
1639 mark_sym_for_renaming (sym);
1640 }
1641 }
1642
1643
1644 /* Mark all the variables in virtual operands in all the statements in
1645 LIST for renaming. */
1646
1647 static void
1648 mark_all_v_defs (tree list)
1649 {
1650 if (TREE_CODE (list) != STATEMENT_LIST)
1651 mark_all_v_defs_1 (list);
1652 else
1653 {
1654 tree_stmt_iterator i;
1655 for (i = tsi_start (list); !tsi_end_p (i); tsi_next (&i))
1656 mark_all_v_defs_1 (tsi_stmt (i));
1657 }
1658 }
1659
1660
1661 /* Mark every replacement under ELT with TREE_NO_WARNING. */
1662
1663 static void
1664 mark_no_warning (struct sra_elt *elt)
1665 {
1666 if (!elt->all_no_warning)
1667 {
1668 if (elt->replacement)
1669 TREE_NO_WARNING (elt->replacement) = 1;
1670 else
1671 {
1672 struct sra_elt *c;
1673 FOR_EACH_ACTUAL_CHILD (c, elt)
1674 mark_no_warning (c);
1675 }
1676 elt->all_no_warning = true;
1677 }
1678 }
1679
1680 /* Build a single level component reference to ELT rooted at BASE. */
1681
1682 static tree
1683 generate_one_element_ref (struct sra_elt *elt, tree base)
1684 {
1685 switch (TREE_CODE (TREE_TYPE (base)))
1686 {
1687 case RECORD_TYPE:
1688 {
1689 tree field = elt->element;
1690
1691 /* Watch out for compatible records with differing field lists. */
1692 if (DECL_FIELD_CONTEXT (field) != TYPE_MAIN_VARIANT (TREE_TYPE (base)))
1693 field = find_compatible_field (TREE_TYPE (base), field);
1694
1695 return build3 (COMPONENT_REF, elt->type, base, field, NULL);
1696 }
1697
1698 case ARRAY_TYPE:
1699 if (TREE_CODE (elt->element) == RANGE_EXPR)
1700 return build4 (ARRAY_RANGE_REF, elt->type, base,
1701 TREE_OPERAND (elt->element, 0), NULL, NULL);
1702 else
1703 return build4 (ARRAY_REF, elt->type, base, elt->element, NULL, NULL);
1704
1705 case COMPLEX_TYPE:
1706 if (elt->element == integer_zero_node)
1707 return build1 (REALPART_EXPR, elt->type, base);
1708 else
1709 return build1 (IMAGPART_EXPR, elt->type, base);
1710
1711 default:
1712 gcc_unreachable ();
1713 }
1714 }
1715
1716 /* Build a full component reference to ELT rooted at its native variable. */
1717
1718 static tree
1719 generate_element_ref (struct sra_elt *elt)
1720 {
1721 if (elt->parent)
1722 return generate_one_element_ref (elt, generate_element_ref (elt->parent));
1723 else
1724 return elt->element;
1725 }
1726
1727 /* Create an assignment statement from SRC to DST. */
1728
1729 static tree
1730 sra_build_assignment (tree dst, tree src)
1731 {
1732 /* It was hoped that we could perform some type sanity checking
1733 here, but since front-ends can emit accesses of fields in types
1734 different from their nominal types and copy structures containing
1735 them as a whole, we'd have to handle such differences here.
1736 Since such accesses under different types require compatibility
1737 anyway, there's little point in making tests and/or adding
1738 conversions to ensure the types of src and dst are the same.
1739 So we just assume type differences at this point are ok. */
1740 return build_gimple_modify_stmt (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 };