1 /* Scalar Replacement of Aggregates (SRA) converts some structure
2 references into scalar references, exposing them to the scalar
4 Copyright (C) 2003, 2004, 2005, 2006, 2007
5 Free Software Foundation, Inc.
6 Contributed by Diego Novillo <dnovillo@redhat.com>
8 This file is part of GCC.
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
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
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/>. */
26 #include "coretypes.h"
31 /* These RTL headers are needed for basic-block.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"
48 /* expr.h is needed for MOVE_RATIO. */
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.
59 This substitution is done globally. More localized substitutions would
60 be the purvey of a load-store motion pass.
62 The optimization proceeds in phases:
64 (1) Identify variables that have types that are candidates for
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.
72 (3) Based on the usage profile, instantiate substitution variables.
74 (4) Scan the function making replacements.
78 /* True if this is the "early" pass, before inlining. */
79 static bool early_sra;
81 /* The set of todo flags to return from tree_sra. */
82 static unsigned int todoflags;
84 /* The set of aggregate variables that are candidates for scalarization. */
85 static bitmap sra_candidates;
87 /* Set of scalarizable PARM_DECLs that need copy-in operations at the
88 beginning of the function. */
89 static bitmap needs_copy_in;
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;
95 /* One of these structures is created for each candidate aggregate and
96 each (accessed) member or group of members of such an aggregate. */
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;
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. */
113 /* The type of the element. */
116 /* A VAR_DECL, for any sub-element we've decided to replace. */
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. */
123 /* The number of times the element is copied to or from another
124 scalarizable element. */
125 unsigned int n_copies;
127 /* True if TYPE is scalar. */
130 /* True if this element is a group of members of its parent. */
133 /* True if we saw something about this element that prevents scalarization,
134 such as non-constant indexing. */
135 bool cannot_scalarize;
137 /* True if we've decided that structure-to-structure assignment
138 should happen via memcpy and not per-element. */
141 /* True if everything under this element has been marked TREE_NO_WARNING. */
144 /* A flag for use with/after random access traversals. */
147 /* True if there is BIT_FIELD_REF on the lhs with a vector. */
151 #define IS_ELEMENT_FOR_GROUP(ELEMENT) (TREE_CODE (ELEMENT) == RANGE_EXPR)
153 #define FOR_EACH_ACTUAL_CHILD(CHILD, ELT) \
154 for ((CHILD) = (ELT)->is_group \
155 ? next_child_for_group (NULL, (ELT)) \
158 (CHILD) = (ELT)->is_group \
159 ? next_child_for_group ((CHILD), (ELT)) \
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)
166 gcc_assert (group->is_group);
168 /* Find the next child in the parent. */
170 child = child->sibling;
172 child = group->parent->children;
174 /* Skip siblings that do not belong to the group. */
177 tree g_elt = group->element;
178 if (TREE_CODE (g_elt) == RANGE_EXPR)
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))
187 child = child->sibling;
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;
198 /* All structures are allocated out of the following obstack. */
199 static struct obstack sra_obstack;
201 /* Debugging functions. */
202 static void dump_sra_elt_name (FILE *, struct sra_elt *);
203 extern void debug_sra_elt_name (struct sra_elt *);
205 /* Forward declarations. */
206 static tree generate_element_ref (struct sra_elt *);
208 /* Return true if DECL is an SRA candidate. */
211 is_sra_candidate_decl (tree decl)
213 return DECL_P (decl) && bitmap_bit_p (sra_candidates, DECL_UID (decl));
216 /* Return true if TYPE is a scalar type. */
219 is_sra_scalar_type (tree type)
221 enum tree_code code = TREE_CODE (type);
222 return (code == INTEGER_TYPE || code == REAL_TYPE || code == VECTOR_TYPE
223 || code == FIXED_POINT_TYPE
224 || code == ENUMERAL_TYPE || code == BOOLEAN_TYPE
225 || code == POINTER_TYPE || code == OFFSET_TYPE
226 || code == REFERENCE_TYPE);
229 /* Return true if TYPE can be decomposed into a set of independent variables.
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. */
236 sra_type_can_be_decomposed_p (tree type)
238 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
241 /* Avoid searching the same type twice. */
242 if (bitmap_bit_p (sra_type_decomp_cache, cache+0))
244 if (bitmap_bit_p (sra_type_decomp_cache, cache+1))
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)))
252 /* The type must be a non-union aggregate. */
253 switch (TREE_CODE (type))
257 bool saw_one_field = false;
259 for (t = TYPE_FIELDS (type); t ; t = TREE_CHAIN (t))
260 if (TREE_CODE (t) == FIELD_DECL)
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))))
268 saw_one_field = true;
271 /* Record types must have at least one field. */
278 /* Array types must have a fixed lower and upper bound. */
279 t = TYPE_DOMAIN (type);
282 if (TYPE_MIN_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MIN_VALUE (t)))
284 if (TYPE_MAX_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MAX_VALUE (t)))
295 bitmap_set_bit (sra_type_decomp_cache, cache+0);
299 bitmap_set_bit (sra_type_decomp_cache, cache+1);
303 /* Return true if DECL can be decomposed into a set of independent
304 (though not necessarily scalar) variables. */
307 decl_can_be_decomposed_p (tree var)
309 /* Early out for scalars. */
310 if (is_sra_scalar_type (TREE_TYPE (var)))
313 /* The variable must not be aliased. */
314 if (!is_gimple_non_addressable (var))
316 if (dump_file && (dump_flags & TDF_DETAILS))
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");
325 /* The variable must not be volatile. */
326 if (TREE_THIS_VOLATILE (var))
328 if (dump_file && (dump_flags & TDF_DETAILS))
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");
337 /* We must be able to decompose the variable's type. */
338 if (!sra_type_can_be_decomposed_p (TREE_TYPE (var)))
340 if (dump_file && (dump_flags & TDF_DETAILS))
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");
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. */
356 && TYPE_MAIN_VARIANT (TREE_TYPE (var))
357 == TYPE_MAIN_VARIANT (va_list_type_node))
363 /* Return true if TYPE can be *completely* decomposed into scalars. */
366 type_can_instantiate_all_elements (tree type)
368 if (is_sra_scalar_type (type))
370 if (!sra_type_can_be_decomposed_p (type))
373 switch (TREE_CODE (type))
377 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
380 if (bitmap_bit_p (sra_type_inst_cache, cache+0))
382 if (bitmap_bit_p (sra_type_inst_cache, cache+1))
385 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
386 if (TREE_CODE (f) == FIELD_DECL)
388 if (!type_can_instantiate_all_elements (TREE_TYPE (f)))
390 bitmap_set_bit (sra_type_inst_cache, cache+1);
395 bitmap_set_bit (sra_type_inst_cache, cache+0);
400 return type_can_instantiate_all_elements (TREE_TYPE (type));
410 /* Test whether ELT or some sub-element cannot be scalarized. */
413 can_completely_scalarize_p (struct sra_elt *elt)
417 if (elt->cannot_scalarize)
420 for (c = elt->children; c; c = c->sibling)
421 if (!can_completely_scalarize_p (c))
424 for (c = elt->groups; c; c = c->sibling)
425 if (!can_completely_scalarize_p (c))
432 /* A simplified tree hashing algorithm that only handles the types of
433 trees we expect to find in sra_elt->element. */
436 sra_hash_tree (tree t)
440 switch (TREE_CODE (t))
449 h = TREE_INT_CST_LOW (t) ^ TREE_INT_CST_HIGH (t);
453 h = iterative_hash_expr (TREE_OPERAND (t, 0), 0);
454 h = iterative_hash_expr (TREE_OPERAND (t, 1), h);
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);
471 /* Hash function for type SRA_PAIR. */
474 sra_elt_hash (const void *x)
476 const struct sra_elt *e = x;
477 const struct sra_elt *p;
480 h = sra_hash_tree (e->element);
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);
492 /* Equality function for type SRA_PAIR. */
495 sra_elt_eq (const void *x, const void *y)
497 const struct sra_elt *a = x;
498 const struct sra_elt *b = y;
501 if (a->parent != b->parent)
509 if (TREE_CODE (ae) != TREE_CODE (be))
512 switch (TREE_CODE (ae))
517 /* These are all pointer unique. */
521 /* Integers are not pointer unique, so compare their values. */
522 return tree_int_cst_equal (ae, be);
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));
530 /* Fields are unique within a record, but not between
531 compatible records. */
532 if (DECL_FIELD_CONTEXT (ae) == DECL_FIELD_CONTEXT (be))
534 return fields_compatible_p (ae, be);
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. */
544 static struct sra_elt *
545 lookup_element (struct sra_elt *parent, tree child, tree type,
546 enum insert_option insert)
548 struct sra_elt dummy;
549 struct sra_elt **slot;
553 dummy.parent = parent->is_group ? parent->parent : parent;
556 dummy.element = child;
558 slot = (struct sra_elt **) htab_find_slot (sra_map, &dummy, insert);
559 if (!slot && insert == NO_INSERT)
563 if (!elt && insert == INSERT)
565 *slot = elt = obstack_alloc (&sra_obstack, sizeof (*elt));
566 memset (elt, 0, sizeof (*elt));
568 elt->parent = parent;
569 elt->element = child;
571 elt->is_scalar = is_sra_scalar_type (type);
575 if (IS_ELEMENT_FOR_GROUP (elt->element))
577 elt->is_group = true;
578 elt->sibling = parent->groups;
579 parent->groups = elt;
583 elt->sibling = parent->children;
584 parent->children = elt;
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)
593 bitmap_set_bit (needs_copy_in, DECL_UID (child));
600 /* Create or return the SRA_ELT structure for EXPR if the expression
601 refers to a scalarizable variable. */
603 static struct sra_elt *
604 maybe_lookup_element_for_expr (tree expr)
609 switch (TREE_CODE (expr))
614 if (is_sra_candidate_decl (expr))
615 return lookup_element (NULL, expr, TREE_TYPE (expr), INSERT);
619 /* We can't scalarize variable array indices. */
620 if (in_array_bounds_p (expr))
621 child = TREE_OPERAND (expr, 1);
626 case ARRAY_RANGE_REF:
627 /* We can't scalarize variable array indices. */
628 if (range_in_array_bounds_p (expr))
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));
640 tree type = TREE_TYPE (TREE_OPERAND (expr, 0));
641 /* Don't look through unions. */
642 if (TREE_CODE (type) != RECORD_TYPE)
644 /* Neither through variable-sized records. */
645 if (TYPE_SIZE (type) == NULL_TREE
646 || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
648 child = TREE_OPERAND (expr, 1);
653 child = integer_zero_node;
656 child = integer_one_node;
663 elt = maybe_lookup_element_for_expr (TREE_OPERAND (expr, 0));
665 return lookup_element (elt, child, TREE_TYPE (expr), INSERT);
670 /* Functions to walk just enough of the tree to see all scalarizable
671 references, and categorize them. */
673 /* A set of callbacks for phases 2 and 4. They'll be invoked for the
674 various kinds of references seen. In all cases, *BSI is an iterator
675 pointing to the statement being processed. */
678 /* Invoked when ELT is required as a unit. Note that ELT might refer to
679 a leaf node, in which case this is a simple scalar reference. *EXPR_P
680 points to the location of the expression. IS_OUTPUT is true if this
681 is a left-hand-side reference. USE_ALL is true if we saw something we
682 couldn't quite identify and had to force the use of the entire object. */
683 void (*use) (struct sra_elt *elt, tree *expr_p,
684 block_stmt_iterator *bsi, bool is_output, bool use_all);
686 /* Invoked when we have a copy between two scalarizable references. */
687 void (*copy) (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
688 block_stmt_iterator *bsi);
690 /* Invoked when ELT is initialized from a constant. VALUE may be NULL,
691 in which case it should be treated as an empty CONSTRUCTOR. */
692 void (*init) (struct sra_elt *elt, tree value, block_stmt_iterator *bsi);
694 /* Invoked when we have a copy between one scalarizable reference ELT
695 and one non-scalarizable reference OTHER without side-effects.
696 IS_OUTPUT is true if ELT is on the left-hand side. */
697 void (*ldst) (struct sra_elt *elt, tree other,
698 block_stmt_iterator *bsi, bool is_output);
700 /* True during phase 2, false during phase 4. */
701 /* ??? This is a hack. */
705 #ifdef ENABLE_CHECKING
706 /* Invoked via walk_tree, if *TP contains a candidate decl, return it. */
709 sra_find_candidate_decl (tree *tp, int *walk_subtrees,
710 void *data ATTRIBUTE_UNUSED)
713 enum tree_code code = TREE_CODE (t);
715 if (code == VAR_DECL || code == PARM_DECL || code == RESULT_DECL)
718 if (is_sra_candidate_decl (t))
728 /* Walk most expressions looking for a scalarizable aggregate.
729 If we find one, invoke FNS->USE. */
732 sra_walk_expr (tree *expr_p, block_stmt_iterator *bsi, bool is_output,
733 const struct sra_walk_fns *fns)
737 bool disable_scalarization = false;
738 bool use_all_p = false;
740 /* We're looking to collect a reference expression between EXPR and INNER,
741 such that INNER is a scalarizable decl and all other nodes through EXPR
742 are references that we can scalarize. If we come across something that
743 we can't scalarize, we reset EXPR. This has the effect of making it
744 appear that we're referring to the larger expression as a whole. */
747 switch (TREE_CODE (inner))
752 /* If there is a scalarizable decl at the bottom, then process it. */
753 if (is_sra_candidate_decl (inner))
755 struct sra_elt *elt = maybe_lookup_element_for_expr (expr);
756 if (disable_scalarization)
757 elt->cannot_scalarize = true;
759 fns->use (elt, expr_p, bsi, is_output, use_all_p);
764 /* Non-constant index means any member may be accessed. Prevent the
765 expression from being scalarized. If we were to treat this as a
766 reference to the whole array, we can wind up with a single dynamic
767 index reference inside a loop being overridden by several constant
768 index references during loop setup. It's possible that this could
769 be avoided by using dynamic usage counts based on BB trip counts
770 (based on loop analysis or profiling), but that hardly seems worth
772 /* ??? Hack. Figure out how to push this into the scan routines
773 without duplicating too much code. */
774 if (!in_array_bounds_p (inner))
776 disable_scalarization = true;
779 /* ??? Are we assured that non-constant bounds and stride will have
780 the same value everywhere? I don't think Fortran will... */
781 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
783 inner = TREE_OPERAND (inner, 0);
786 case ARRAY_RANGE_REF:
787 if (!range_in_array_bounds_p (inner))
789 disable_scalarization = true;
792 /* ??? See above non-constant bounds and stride . */
793 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
795 inner = TREE_OPERAND (inner, 0);
800 tree type = TREE_TYPE (TREE_OPERAND (inner, 0));
801 /* Don't look through unions. */
802 if (TREE_CODE (type) != RECORD_TYPE)
804 /* Neither through variable-sized records. */
805 if (TYPE_SIZE (type) == NULL_TREE
806 || TREE_CODE (TYPE_SIZE (type)) != INTEGER_CST)
808 inner = TREE_OPERAND (inner, 0);
814 inner = TREE_OPERAND (inner, 0);
818 /* A bit field reference to a specific vector is scalarized but for
819 ones for inputs need to be marked as used on the left hand size so
820 when we scalarize it, we can mark that variable as non renamable. */
822 && TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) == VECTOR_TYPE)
825 = maybe_lookup_element_for_expr (TREE_OPERAND (inner, 0));
827 elt->is_vector_lhs = true;
829 /* A bit field reference (access to *multiple* fields simultaneously)
830 is not currently scalarized. Consider this an access to the
831 complete outer element, to which walk_tree will bring us next. */
835 case VIEW_CONVERT_EXPR:
837 /* Similarly, a view/nop explicitly wants to look at an object in a
838 type other than the one we've scalarized. */
842 /* This is a transparent wrapper. The entire inner expression really
847 expr_p = &TREE_OPERAND (inner, 0);
848 inner = expr = *expr_p;
853 #ifdef ENABLE_CHECKING
854 /* Validate that we're not missing any references. */
855 gcc_assert (!walk_tree (&inner, sra_find_candidate_decl, NULL, NULL));
861 /* Walk a TREE_LIST of values looking for scalarizable aggregates.
862 If we find one, invoke FNS->USE. */
865 sra_walk_tree_list (tree list, block_stmt_iterator *bsi, bool is_output,
866 const struct sra_walk_fns *fns)
869 for (op = list; op ; op = TREE_CHAIN (op))
870 sra_walk_expr (&TREE_VALUE (op), bsi, is_output, fns);
873 /* Walk the arguments of a CALL_EXPR looking for scalarizable aggregates.
874 If we find one, invoke FNS->USE. */
877 sra_walk_call_expr (tree expr, block_stmt_iterator *bsi,
878 const struct sra_walk_fns *fns)
881 int nargs = call_expr_nargs (expr);
882 for (i = 0; i < nargs; i++)
883 sra_walk_expr (&CALL_EXPR_ARG (expr, i), bsi, false, fns);
886 /* Walk the inputs and outputs of an ASM_EXPR looking for scalarizable
887 aggregates. If we find one, invoke FNS->USE. */
890 sra_walk_asm_expr (tree expr, block_stmt_iterator *bsi,
891 const struct sra_walk_fns *fns)
893 sra_walk_tree_list (ASM_INPUTS (expr), bsi, false, fns);
894 sra_walk_tree_list (ASM_OUTPUTS (expr), bsi, true, fns);
897 /* Walk a GIMPLE_MODIFY_STMT and categorize the assignment appropriately. */
900 sra_walk_gimple_modify_stmt (tree expr, block_stmt_iterator *bsi,
901 const struct sra_walk_fns *fns)
903 struct sra_elt *lhs_elt, *rhs_elt;
906 lhs = GIMPLE_STMT_OPERAND (expr, 0);
907 rhs = GIMPLE_STMT_OPERAND (expr, 1);
908 lhs_elt = maybe_lookup_element_for_expr (lhs);
909 rhs_elt = maybe_lookup_element_for_expr (rhs);
911 /* If both sides are scalarizable, this is a COPY operation. */
912 if (lhs_elt && rhs_elt)
914 fns->copy (lhs_elt, rhs_elt, bsi);
918 /* If the RHS is scalarizable, handle it. There are only two cases. */
921 if (!rhs_elt->is_scalar && !TREE_SIDE_EFFECTS (lhs))
922 fns->ldst (rhs_elt, lhs, bsi, false);
924 fns->use (rhs_elt, &GIMPLE_STMT_OPERAND (expr, 1), bsi, false, false);
927 /* If it isn't scalarizable, there may be scalarizable variables within, so
928 check for a call or else walk the RHS to see if we need to do any
929 copy-in operations. We need to do it before the LHS is scalarized so
930 that the statements get inserted in the proper place, before any
931 copy-out operations. */
934 tree call = get_call_expr_in (rhs);
936 sra_walk_call_expr (call, bsi, fns);
938 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 1), bsi, false, fns);
941 /* Likewise, handle the LHS being scalarizable. We have cases similar
942 to those above, but also want to handle RHS being constant. */
945 /* If this is an assignment from a constant, or constructor, then
946 we have access to all of the elements individually. Invoke INIT. */
947 if (TREE_CODE (rhs) == COMPLEX_EXPR
948 || TREE_CODE (rhs) == COMPLEX_CST
949 || TREE_CODE (rhs) == CONSTRUCTOR)
950 fns->init (lhs_elt, rhs, bsi);
952 /* If this is an assignment from read-only memory, treat this as if
953 we'd been passed the constructor directly. Invoke INIT. */
954 else if (TREE_CODE (rhs) == VAR_DECL
956 && TREE_READONLY (rhs)
957 && targetm.binds_local_p (rhs))
958 fns->init (lhs_elt, DECL_INITIAL (rhs), bsi);
960 /* If this is a copy from a non-scalarizable lvalue, invoke LDST.
961 The lvalue requirement prevents us from trying to directly scalarize
962 the result of a function call. Which would result in trying to call
963 the function multiple times, and other evil things. */
964 else if (!lhs_elt->is_scalar
965 && !TREE_SIDE_EFFECTS (rhs) && is_gimple_addressable (rhs))
966 fns->ldst (lhs_elt, rhs, bsi, true);
968 /* Otherwise we're being used in some context that requires the
969 aggregate to be seen as a whole. Invoke USE. */
971 fns->use (lhs_elt, &GIMPLE_STMT_OPERAND (expr, 0), bsi, true, false);
974 /* Similarly to above, LHS_ELT being null only means that the LHS as a
975 whole is not a scalarizable reference. There may be occurrences of
976 scalarizable variables within, which implies a USE. */
978 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 0), bsi, true, fns);
981 /* Entry point to the walk functions. Search the entire function,
982 invoking the callbacks in FNS on each of the references to
983 scalarizable variables. */
986 sra_walk_function (const struct sra_walk_fns *fns)
989 block_stmt_iterator si, ni;
991 /* ??? Phase 4 could derive some benefit to walking the function in
992 dominator tree order. */
995 for (si = bsi_start (bb); !bsi_end_p (si); si = ni)
1000 stmt = bsi_stmt (si);
1001 ann = stmt_ann (stmt);
1006 /* If the statement has no virtual operands, then it doesn't
1007 make any structure references that we care about. */
1008 if (gimple_aliases_computed_p (cfun)
1009 && ZERO_SSA_OPERANDS (stmt, (SSA_OP_VIRTUAL_DEFS | SSA_OP_VUSE)))
1012 switch (TREE_CODE (stmt))
1015 /* If we have "return <retval>" then the return value is
1016 already exposed for our pleasure. Walk it as a USE to
1017 force all the components back in place for the return.
1019 If we have an embedded assignment, then <retval> is of
1020 a type that gets returned in registers in this ABI, and
1021 we do not wish to extend their lifetimes. Treat this
1022 as a USE of the variable on the RHS of this assignment. */
1024 t = TREE_OPERAND (stmt, 0);
1027 else if (TREE_CODE (t) == GIMPLE_MODIFY_STMT)
1028 sra_walk_expr (&GIMPLE_STMT_OPERAND (t, 1), &si, false, fns);
1030 sra_walk_expr (&TREE_OPERAND (stmt, 0), &si, false, fns);
1033 case GIMPLE_MODIFY_STMT:
1034 sra_walk_gimple_modify_stmt (stmt, &si, fns);
1037 sra_walk_call_expr (stmt, &si, fns);
1040 sra_walk_asm_expr (stmt, &si, fns);
1049 /* Phase One: Scan all referenced variables in the program looking for
1050 structures that could be decomposed. */
1053 find_candidates_for_sra (void)
1055 bool any_set = false;
1057 referenced_var_iterator rvi;
1059 FOR_EACH_REFERENCED_VAR (var, rvi)
1061 if (decl_can_be_decomposed_p (var))
1063 bitmap_set_bit (sra_candidates, DECL_UID (var));
1072 /* Phase Two: Scan all references to scalarizable variables. Count the
1073 number of times they are used or copied respectively. */
1075 /* Callbacks to fill in SRA_WALK_FNS. Everything but USE is
1076 considered a copy, because we can decompose the reference such that
1077 the sub-elements needn't be contiguous. */
1080 scan_use (struct sra_elt *elt, tree *expr_p ATTRIBUTE_UNUSED,
1081 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
1082 bool is_output ATTRIBUTE_UNUSED, bool use_all ATTRIBUTE_UNUSED)
1088 scan_copy (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
1089 block_stmt_iterator *bsi ATTRIBUTE_UNUSED)
1091 lhs_elt->n_copies += 1;
1092 rhs_elt->n_copies += 1;
1096 scan_init (struct sra_elt *lhs_elt, tree rhs ATTRIBUTE_UNUSED,
1097 block_stmt_iterator *bsi ATTRIBUTE_UNUSED)
1099 lhs_elt->n_copies += 1;
1103 scan_ldst (struct sra_elt *elt, tree other ATTRIBUTE_UNUSED,
1104 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
1105 bool is_output ATTRIBUTE_UNUSED)
1110 /* Dump the values we collected during the scanning phase. */
1113 scan_dump (struct sra_elt *elt)
1117 dump_sra_elt_name (dump_file, elt);
1118 fprintf (dump_file, ": n_uses=%u n_copies=%u\n", elt->n_uses, elt->n_copies);
1120 for (c = elt->children; c ; c = c->sibling)
1123 for (c = elt->groups; c ; c = c->sibling)
1127 /* Entry point to phase 2. Scan the entire function, building up
1128 scalarization data structures, recording copies and uses. */
1131 scan_function (void)
1133 static const struct sra_walk_fns fns = {
1134 scan_use, scan_copy, scan_init, scan_ldst, true
1138 sra_walk_function (&fns);
1140 if (dump_file && (dump_flags & TDF_DETAILS))
1144 fputs ("\nScan results:\n", dump_file);
1145 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1147 tree var = referenced_var (i);
1148 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1152 fputc ('\n', dump_file);
1156 /* Phase Three: Make decisions about which variables to scalarize, if any.
1157 All elements to be scalarized have replacement variables made for them. */
1159 /* A subroutine of build_element_name. Recursively build the element
1160 name on the obstack. */
1163 build_element_name_1 (struct sra_elt *elt)
1170 build_element_name_1 (elt->parent);
1171 obstack_1grow (&sra_obstack, '$');
1173 if (TREE_CODE (elt->parent->type) == COMPLEX_TYPE)
1175 if (elt->element == integer_zero_node)
1176 obstack_grow (&sra_obstack, "real", 4);
1178 obstack_grow (&sra_obstack, "imag", 4);
1184 if (TREE_CODE (t) == INTEGER_CST)
1186 /* ??? Eh. Don't bother doing double-wide printing. */
1187 sprintf (buffer, HOST_WIDE_INT_PRINT_DEC, TREE_INT_CST_LOW (t));
1188 obstack_grow (&sra_obstack, buffer, strlen (buffer));
1192 tree name = DECL_NAME (t);
1194 obstack_grow (&sra_obstack, IDENTIFIER_POINTER (name),
1195 IDENTIFIER_LENGTH (name));
1198 sprintf (buffer, "D%u", DECL_UID (t));
1199 obstack_grow (&sra_obstack, buffer, strlen (buffer));
1204 /* Construct a pretty variable name for an element's replacement variable.
1205 The name is built on the obstack. */
1208 build_element_name (struct sra_elt *elt)
1210 build_element_name_1 (elt);
1211 obstack_1grow (&sra_obstack, '\0');
1212 return XOBFINISH (&sra_obstack, char *);
1215 /* Instantiate an element as an independent variable. */
1218 instantiate_element (struct sra_elt *elt)
1220 struct sra_elt *base_elt;
1223 for (base_elt = elt; base_elt->parent; base_elt = base_elt->parent)
1225 base = base_elt->element;
1227 elt->replacement = var = make_rename_temp (elt->type, "SR");
1229 /* For vectors, if used on the left hand side with BIT_FIELD_REF,
1230 they are not a gimple register. */
1231 if (TREE_CODE (TREE_TYPE (var)) == VECTOR_TYPE && elt->is_vector_lhs)
1232 DECL_GIMPLE_REG_P (var) = 0;
1234 DECL_SOURCE_LOCATION (var) = DECL_SOURCE_LOCATION (base);
1235 DECL_ARTIFICIAL (var) = 1;
1237 if (TREE_THIS_VOLATILE (elt->type))
1239 TREE_THIS_VOLATILE (var) = 1;
1240 TREE_SIDE_EFFECTS (var) = 1;
1243 if (DECL_NAME (base) && !DECL_IGNORED_P (base))
1245 char *pretty_name = build_element_name (elt);
1246 DECL_NAME (var) = get_identifier (pretty_name);
1247 obstack_free (&sra_obstack, pretty_name);
1249 SET_DECL_DEBUG_EXPR (var, generate_element_ref (elt));
1250 DECL_DEBUG_EXPR_IS_FROM (var) = 1;
1252 DECL_IGNORED_P (var) = 0;
1253 TREE_NO_WARNING (var) = TREE_NO_WARNING (base);
1254 if (elt->element && TREE_NO_WARNING (elt->element))
1255 TREE_NO_WARNING (var) = 1;
1259 DECL_IGNORED_P (var) = 1;
1260 /* ??? We can't generate any warning that would be meaningful. */
1261 TREE_NO_WARNING (var) = 1;
1266 fputs (" ", dump_file);
1267 dump_sra_elt_name (dump_file, elt);
1268 fputs (" -> ", dump_file);
1269 print_generic_expr (dump_file, var, dump_flags);
1270 fputc ('\n', dump_file);
1274 /* Make one pass across an element tree deciding whether or not it's
1275 profitable to instantiate individual leaf scalars.
1277 PARENT_USES and PARENT_COPIES are the sum of the N_USES and N_COPIES
1278 fields all the way up the tree. */
1281 decide_instantiation_1 (struct sra_elt *elt, unsigned int parent_uses,
1282 unsigned int parent_copies)
1284 if (dump_file && !elt->parent)
1286 fputs ("Initial instantiation for ", dump_file);
1287 dump_sra_elt_name (dump_file, elt);
1288 fputc ('\n', dump_file);
1291 if (elt->cannot_scalarize)
1296 /* The decision is simple: instantiate if we're used more frequently
1297 than the parent needs to be seen as a complete unit. */
1298 if (elt->n_uses + elt->n_copies + parent_copies > parent_uses)
1299 instantiate_element (elt);
1303 struct sra_elt *c, *group;
1304 unsigned int this_uses = elt->n_uses + parent_uses;
1305 unsigned int this_copies = elt->n_copies + parent_copies;
1307 /* Consider groups of sub-elements as weighing in favour of
1308 instantiation whatever their size. */
1309 for (group = elt->groups; group ; group = group->sibling)
1310 FOR_EACH_ACTUAL_CHILD (c, group)
1312 c->n_uses += group->n_uses;
1313 c->n_copies += group->n_copies;
1316 for (c = elt->children; c ; c = c->sibling)
1317 decide_instantiation_1 (c, this_uses, this_copies);
1321 /* Compute the size and number of all instantiated elements below ELT.
1322 We will only care about this if the size of the complete structure
1323 fits in a HOST_WIDE_INT, so we don't have to worry about overflow. */
1326 sum_instantiated_sizes (struct sra_elt *elt, unsigned HOST_WIDE_INT *sizep)
1328 if (elt->replacement)
1330 *sizep += TREE_INT_CST_LOW (TYPE_SIZE_UNIT (elt->type));
1336 unsigned int count = 0;
1338 for (c = elt->children; c ; c = c->sibling)
1339 count += sum_instantiated_sizes (c, sizep);
1345 /* Instantiate fields in ELT->TYPE that are not currently present as
1348 static void instantiate_missing_elements (struct sra_elt *elt);
1351 instantiate_missing_elements_1 (struct sra_elt *elt, tree child, tree type)
1353 struct sra_elt *sub = lookup_element (elt, child, type, INSERT);
1356 if (sub->replacement == NULL)
1357 instantiate_element (sub);
1360 instantiate_missing_elements (sub);
1364 instantiate_missing_elements (struct sra_elt *elt)
1366 tree type = elt->type;
1368 switch (TREE_CODE (type))
1373 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
1374 if (TREE_CODE (f) == FIELD_DECL)
1376 tree field_type = TREE_TYPE (f);
1378 /* canonicalize_component_ref() unwidens some bit-field
1379 types (not marked as DECL_BIT_FIELD in C++), so we
1380 must do the same, lest we may introduce type
1382 if (INTEGRAL_TYPE_P (field_type)
1383 && DECL_MODE (f) != TYPE_MODE (field_type))
1384 field_type = TREE_TYPE (get_unwidened (build3 (COMPONENT_REF,
1390 instantiate_missing_elements_1 (elt, f, field_type);
1397 tree i, max, subtype;
1399 i = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1400 max = TYPE_MAX_VALUE (TYPE_DOMAIN (type));
1401 subtype = TREE_TYPE (type);
1405 instantiate_missing_elements_1 (elt, i, subtype);
1406 if (tree_int_cst_equal (i, max))
1408 i = int_const_binop (PLUS_EXPR, i, integer_one_node, true);
1415 type = TREE_TYPE (type);
1416 instantiate_missing_elements_1 (elt, integer_zero_node, type);
1417 instantiate_missing_elements_1 (elt, integer_one_node, type);
1425 /* Return true if there is only one non aggregate field in the record, TYPE.
1426 Return false otherwise. */
1429 single_scalar_field_in_record_p (tree type)
1433 if (TREE_CODE (type) != RECORD_TYPE)
1436 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
1437 if (TREE_CODE (field) == FIELD_DECL)
1441 if (num_fields == 2)
1444 if (AGGREGATE_TYPE_P (TREE_TYPE (field)))
1451 /* Make one pass across an element tree deciding whether to perform block
1452 or element copies. If we decide on element copies, instantiate all
1453 elements. Return true if there are any instantiated sub-elements. */
1456 decide_block_copy (struct sra_elt *elt)
1461 /* We shouldn't be invoked on groups of sub-elements as they must
1462 behave like their parent as far as block copy is concerned. */
1463 gcc_assert (!elt->is_group);
1465 /* If scalarization is disabled, respect it. */
1466 if (elt->cannot_scalarize)
1468 elt->use_block_copy = 1;
1472 fputs ("Scalarization disabled for ", dump_file);
1473 dump_sra_elt_name (dump_file, elt);
1474 fputc ('\n', dump_file);
1477 /* Disable scalarization of sub-elements */
1478 for (c = elt->children; c; c = c->sibling)
1480 c->cannot_scalarize = 1;
1481 decide_block_copy (c);
1484 /* Groups behave like their parent. */
1485 for (c = elt->groups; c; c = c->sibling)
1487 c->cannot_scalarize = 1;
1488 c->use_block_copy = 1;
1494 /* Don't decide if we've no uses and no groups. */
1495 if (elt->n_uses == 0 && elt->n_copies == 0 && elt->groups == NULL)
1498 else if (!elt->is_scalar)
1500 tree size_tree = TYPE_SIZE_UNIT (elt->type);
1501 bool use_block_copy = true;
1503 /* Tradeoffs for COMPLEX types pretty much always make it better
1504 to go ahead and split the components. */
1505 if (TREE_CODE (elt->type) == COMPLEX_TYPE)
1506 use_block_copy = false;
1508 /* Don't bother trying to figure out the rest if the structure is
1509 so large we can't do easy arithmetic. This also forces block
1510 copies for variable sized structures. */
1511 else if (host_integerp (size_tree, 1))
1513 unsigned HOST_WIDE_INT full_size, inst_size = 0;
1514 unsigned int max_size, max_count, inst_count, full_count;
1516 /* If the sra-max-structure-size parameter is 0, then the
1517 user has not overridden the parameter and we can choose a
1518 sensible default. */
1519 max_size = SRA_MAX_STRUCTURE_SIZE
1520 ? SRA_MAX_STRUCTURE_SIZE
1521 : MOVE_RATIO * UNITS_PER_WORD;
1522 max_count = SRA_MAX_STRUCTURE_COUNT
1523 ? SRA_MAX_STRUCTURE_COUNT
1526 full_size = tree_low_cst (size_tree, 1);
1527 full_count = count_type_elements (elt->type, false);
1528 inst_count = sum_instantiated_sizes (elt, &inst_size);
1530 /* If there is only one scalar field in the record, don't block copy. */
1531 if (single_scalar_field_in_record_p (elt->type))
1532 use_block_copy = false;
1534 /* ??? What to do here. If there are two fields, and we've only
1535 instantiated one, then instantiating the other is clearly a win.
1536 If there are a large number of fields then the size of the copy
1537 is much more of a factor. */
1539 /* If the structure is small, and we've made copies, go ahead
1540 and instantiate, hoping that the copies will go away. */
1541 if (full_size <= max_size
1542 && (full_count - inst_count) <= max_count
1543 && elt->n_copies > elt->n_uses)
1544 use_block_copy = false;
1545 else if (inst_count * 100 >= full_count * SRA_FIELD_STRUCTURE_RATIO
1546 && inst_size * 100 >= full_size * SRA_FIELD_STRUCTURE_RATIO)
1547 use_block_copy = false;
1549 /* In order to avoid block copy, we have to be able to instantiate
1550 all elements of the type. See if this is possible. */
1552 && (!can_completely_scalarize_p (elt)
1553 || !type_can_instantiate_all_elements (elt->type)))
1554 use_block_copy = true;
1557 elt->use_block_copy = use_block_copy;
1559 /* Groups behave like their parent. */
1560 for (c = elt->groups; c; c = c->sibling)
1561 c->use_block_copy = use_block_copy;
1565 fprintf (dump_file, "Using %s for ",
1566 use_block_copy ? "block-copy" : "element-copy");
1567 dump_sra_elt_name (dump_file, elt);
1568 fputc ('\n', dump_file);
1571 if (!use_block_copy)
1573 instantiate_missing_elements (elt);
1578 any_inst = elt->replacement != NULL;
1580 for (c = elt->children; c ; c = c->sibling)
1581 any_inst |= decide_block_copy (c);
1586 /* Entry point to phase 3. Instantiate scalar replacement variables. */
1589 decide_instantiations (void)
1593 bitmap_head done_head;
1596 /* We cannot clear bits from a bitmap we're iterating over,
1597 so save up all the bits to clear until the end. */
1598 bitmap_initialize (&done_head, &bitmap_default_obstack);
1599 cleared_any = false;
1601 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1603 tree var = referenced_var (i);
1604 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1607 decide_instantiation_1 (elt, 0, 0);
1608 if (!decide_block_copy (elt))
1613 bitmap_set_bit (&done_head, i);
1620 bitmap_and_compl_into (sra_candidates, &done_head);
1621 bitmap_and_compl_into (needs_copy_in, &done_head);
1623 bitmap_clear (&done_head);
1625 mark_set_for_renaming (sra_candidates);
1628 fputc ('\n', dump_file);
1632 /* Phase Four: Update the function to match the replacements created. */
1634 /* Mark all the variables in VDEF/VUSE operators for STMT for
1635 renaming. This becomes necessary when we modify all of a
1639 mark_all_v_defs_1 (tree stmt)
1644 update_stmt_if_modified (stmt);
1646 FOR_EACH_SSA_TREE_OPERAND (sym, stmt, iter, SSA_OP_ALL_VIRTUALS)
1648 if (TREE_CODE (sym) == SSA_NAME)
1649 sym = SSA_NAME_VAR (sym);
1650 mark_sym_for_renaming (sym);
1655 /* Mark all the variables in virtual operands in all the statements in
1656 LIST for renaming. */
1659 mark_all_v_defs (tree list)
1661 if (TREE_CODE (list) != STATEMENT_LIST)
1662 mark_all_v_defs_1 (list);
1665 tree_stmt_iterator i;
1666 for (i = tsi_start (list); !tsi_end_p (i); tsi_next (&i))
1667 mark_all_v_defs_1 (tsi_stmt (i));
1672 /* Mark every replacement under ELT with TREE_NO_WARNING. */
1675 mark_no_warning (struct sra_elt *elt)
1677 if (!elt->all_no_warning)
1679 if (elt->replacement)
1680 TREE_NO_WARNING (elt->replacement) = 1;
1684 FOR_EACH_ACTUAL_CHILD (c, elt)
1685 mark_no_warning (c);
1687 elt->all_no_warning = true;
1691 /* Build a single level component reference to ELT rooted at BASE. */
1694 generate_one_element_ref (struct sra_elt *elt, tree base)
1696 switch (TREE_CODE (TREE_TYPE (base)))
1700 tree field = elt->element;
1702 /* Watch out for compatible records with differing field lists. */
1703 if (DECL_FIELD_CONTEXT (field) != TYPE_MAIN_VARIANT (TREE_TYPE (base)))
1704 field = find_compatible_field (TREE_TYPE (base), field);
1706 return build3 (COMPONENT_REF, elt->type, base, field, NULL);
1710 if (TREE_CODE (elt->element) == RANGE_EXPR)
1711 return build4 (ARRAY_RANGE_REF, elt->type, base,
1712 TREE_OPERAND (elt->element, 0), NULL, NULL);
1714 return build4 (ARRAY_REF, elt->type, base, elt->element, NULL, NULL);
1717 if (elt->element == integer_zero_node)
1718 return build1 (REALPART_EXPR, elt->type, base);
1720 return build1 (IMAGPART_EXPR, elt->type, base);
1727 /* Build a full component reference to ELT rooted at its native variable. */
1730 generate_element_ref (struct sra_elt *elt)
1733 return generate_one_element_ref (elt, generate_element_ref (elt->parent));
1735 return elt->element;
1738 /* Create an assignment statement from SRC to DST. */
1741 sra_build_assignment (tree dst, tree src)
1743 /* It was hoped that we could perform some type sanity checking
1744 here, but since front-ends can emit accesses of fields in types
1745 different from their nominal types and copy structures containing
1746 them as a whole, we'd have to handle such differences here.
1747 Since such accesses under different types require compatibility
1748 anyway, there's little point in making tests and/or adding
1749 conversions to ensure the types of src and dst are the same.
1750 So we just assume type differences at this point are ok. */
1751 return build_gimple_modify_stmt (dst, src);
1754 /* Generate a set of assignment statements in *LIST_P to copy all
1755 instantiated elements under ELT to or from the equivalent structure
1756 rooted at EXPR. COPY_OUT controls the direction of the copy, with
1757 true meaning to copy out of EXPR into ELT. */
1760 generate_copy_inout (struct sra_elt *elt, bool copy_out, tree expr,
1766 if (!copy_out && TREE_CODE (expr) == SSA_NAME
1767 && TREE_CODE (TREE_TYPE (expr)) == COMPLEX_TYPE)
1771 c = lookup_element (elt, integer_zero_node, NULL, NO_INSERT);
1773 c = lookup_element (elt, integer_one_node, NULL, NO_INSERT);
1776 t = build2 (COMPLEX_EXPR, elt->type, r, i);
1777 t = sra_build_assignment (expr, t);
1778 SSA_NAME_DEF_STMT (expr) = t;
1779 append_to_statement_list (t, list_p);
1781 else if (elt->replacement)
1784 t = sra_build_assignment (elt->replacement, expr);
1786 t = sra_build_assignment (expr, elt->replacement);
1787 append_to_statement_list (t, list_p);
1791 FOR_EACH_ACTUAL_CHILD (c, elt)
1793 t = generate_one_element_ref (c, unshare_expr (expr));
1794 generate_copy_inout (c, copy_out, t, list_p);
1799 /* Generate a set of assignment statements in *LIST_P to copy all instantiated
1800 elements under SRC to their counterparts under DST. There must be a 1-1
1801 correspondence of instantiated elements. */
1804 generate_element_copy (struct sra_elt *dst, struct sra_elt *src, tree *list_p)
1806 struct sra_elt *dc, *sc;
1808 FOR_EACH_ACTUAL_CHILD (dc, dst)
1810 sc = lookup_element (src, dc->element, NULL, NO_INSERT);
1812 generate_element_copy (dc, sc, list_p);
1815 if (dst->replacement)
1819 gcc_assert (src->replacement);
1821 t = sra_build_assignment (dst->replacement, src->replacement);
1822 append_to_statement_list (t, list_p);
1826 /* Generate a set of assignment statements in *LIST_P to zero all instantiated
1827 elements under ELT. In addition, do not assign to elements that have been
1828 marked VISITED but do reset the visited flag; this allows easy coordination
1829 with generate_element_init. */
1832 generate_element_zero (struct sra_elt *elt, tree *list_p)
1838 elt->visited = false;
1842 FOR_EACH_ACTUAL_CHILD (c, elt)
1843 generate_element_zero (c, list_p);
1845 if (elt->replacement)
1849 gcc_assert (elt->is_scalar);
1850 t = fold_convert (elt->type, integer_zero_node);
1852 t = sra_build_assignment (elt->replacement, t);
1853 append_to_statement_list (t, list_p);
1857 /* Generate an assignment VAR = INIT, where INIT may need gimplification.
1858 Add the result to *LIST_P. */
1861 generate_one_element_init (tree var, tree init, tree *list_p)
1863 /* The replacement can be almost arbitrarily complex. Gimplify. */
1864 tree stmt = sra_build_assignment (var, init);
1865 gimplify_and_add (stmt, list_p);
1868 /* Generate a set of assignment statements in *LIST_P to set all instantiated
1869 elements under ELT with the contents of the initializer INIT. In addition,
1870 mark all assigned elements VISITED; this allows easy coordination with
1871 generate_element_zero. Return false if we found a case we couldn't
1875 generate_element_init_1 (struct sra_elt *elt, tree init, tree *list_p)
1878 enum tree_code init_code;
1879 struct sra_elt *sub;
1881 unsigned HOST_WIDE_INT idx;
1882 tree value, purpose;
1884 /* We can be passed DECL_INITIAL of a static variable. It might have a
1885 conversion, which we strip off here. */
1886 STRIP_USELESS_TYPE_CONVERSION (init);
1887 init_code = TREE_CODE (init);
1891 if (elt->replacement)
1893 generate_one_element_init (elt->replacement, init, list_p);
1894 elt->visited = true;
1903 FOR_EACH_ACTUAL_CHILD (sub, elt)
1905 if (sub->element == integer_zero_node)
1906 t = (init_code == COMPLEX_EXPR
1907 ? TREE_OPERAND (init, 0) : TREE_REALPART (init));
1909 t = (init_code == COMPLEX_EXPR
1910 ? TREE_OPERAND (init, 1) : TREE_IMAGPART (init));
1911 result &= generate_element_init_1 (sub, t, list_p);
1916 FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (init), idx, purpose, value)
1918 if (TREE_CODE (purpose) == RANGE_EXPR)
1920 tree lower = TREE_OPERAND (purpose, 0);
1921 tree upper = TREE_OPERAND (purpose, 1);
1925 sub = lookup_element (elt, lower, NULL, NO_INSERT);
1927 result &= generate_element_init_1 (sub, value, list_p);
1928 if (tree_int_cst_equal (lower, upper))
1930 lower = int_const_binop (PLUS_EXPR, lower,
1931 integer_one_node, true);
1936 sub = lookup_element (elt, purpose, NULL, NO_INSERT);
1938 result &= generate_element_init_1 (sub, value, list_p);
1944 elt->visited = true;
1951 /* A wrapper function for generate_element_init_1 that handles cleanup after
1955 generate_element_init (struct sra_elt *elt, tree init, tree *list_p)
1959 push_gimplify_context ();
1960 ret = generate_element_init_1 (elt, init, list_p);
1961 pop_gimplify_context (NULL);
1963 /* The replacement can expose previously unreferenced variables. */
1966 tree_stmt_iterator i;
1968 for (i = tsi_start (*list_p); !tsi_end_p (i); tsi_next (&i))
1969 find_new_referenced_vars (tsi_stmt_ptr (i));
1975 /* Insert STMT on all the outgoing edges out of BB. Note that if BB
1976 has more than one edge, STMT will be replicated for each edge. Also,
1977 abnormal edges will be ignored. */
1980 insert_edge_copies (tree stmt, basic_block bb)
1987 FOR_EACH_EDGE (e, ei, bb->succs)
1989 /* We don't need to insert copies on abnormal edges. The
1990 value of the scalar replacement is not guaranteed to
1991 be valid through an abnormal edge. */
1992 if (!(e->flags & EDGE_ABNORMAL))
1996 bsi_insert_on_edge (e, stmt);
2000 bsi_insert_on_edge (e, unsave_expr_now (stmt));
2005 /* Helper function to insert LIST before BSI, and set up line number info. */
2008 sra_insert_before (block_stmt_iterator *bsi, tree list)
2010 tree stmt = bsi_stmt (*bsi);
2012 if (EXPR_HAS_LOCATION (stmt))
2013 annotate_all_with_locus (&list, EXPR_LOCATION (stmt));
2014 bsi_insert_before (bsi, list, BSI_SAME_STMT);
2017 /* Similarly, but insert after BSI. Handles insertion onto edges as well. */
2020 sra_insert_after (block_stmt_iterator *bsi, tree list)
2022 tree stmt = bsi_stmt (*bsi);
2024 if (EXPR_HAS_LOCATION (stmt))
2025 annotate_all_with_locus (&list, EXPR_LOCATION (stmt));
2027 if (stmt_ends_bb_p (stmt))
2028 insert_edge_copies (list, bsi->bb);
2030 bsi_insert_after (bsi, list, BSI_SAME_STMT);
2033 /* Similarly, but replace the statement at BSI. */
2036 sra_replace (block_stmt_iterator *bsi, tree list)
2038 sra_insert_before (bsi, list);
2039 bsi_remove (bsi, false);
2040 if (bsi_end_p (*bsi))
2041 *bsi = bsi_last (bsi->bb);
2046 /* Scalarize a USE. To recap, this is either a simple reference to ELT,
2047 if elt is scalar, or some occurrence of ELT that requires a complete
2048 aggregate. IS_OUTPUT is true if ELT is being modified. */
2051 scalarize_use (struct sra_elt *elt, tree *expr_p, block_stmt_iterator *bsi,
2052 bool is_output, bool use_all)
2054 tree list = NULL, stmt = bsi_stmt (*bsi);
2056 if (elt->replacement)
2058 /* If we have a replacement, then updating the reference is as
2059 simple as modifying the existing statement in place. */
2061 mark_all_v_defs (stmt);
2062 *expr_p = elt->replacement;
2067 /* Otherwise we need some copies. If ELT is being read, then we want
2068 to store all (modified) sub-elements back into the structure before
2069 the reference takes place. If ELT is being written, then we want to
2070 load the changed values back into our shadow variables. */
2071 /* ??? We don't check modified for reads, we just always write all of
2072 the values. We should be able to record the SSA number of the VOP
2073 for which the values were last read. If that number matches the
2074 SSA number of the VOP in the current statement, then we needn't
2075 emit an assignment. This would also eliminate double writes when
2076 a structure is passed as more than one argument to a function call.
2077 This optimization would be most effective if sra_walk_function
2078 processed the blocks in dominator order. */
2080 generate_copy_inout (elt, is_output, generate_element_ref (elt), &list);
2083 mark_all_v_defs (list);
2085 sra_insert_after (bsi, list);
2088 sra_insert_before (bsi, list);
2090 mark_no_warning (elt);
2095 /* Scalarize a COPY. To recap, this is an assignment statement between
2096 two scalarizable references, LHS_ELT and RHS_ELT. */
2099 scalarize_copy (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
2100 block_stmt_iterator *bsi)
2104 if (lhs_elt->replacement && rhs_elt->replacement)
2106 /* If we have two scalar operands, modify the existing statement. */
2107 stmt = bsi_stmt (*bsi);
2109 /* See the commentary in sra_walk_function concerning
2110 RETURN_EXPR, and why we should never see one here. */
2111 gcc_assert (TREE_CODE (stmt) == GIMPLE_MODIFY_STMT);
2113 GIMPLE_STMT_OPERAND (stmt, 0) = lhs_elt->replacement;
2114 GIMPLE_STMT_OPERAND (stmt, 1) = rhs_elt->replacement;
2117 else if (lhs_elt->use_block_copy || rhs_elt->use_block_copy)
2119 /* If either side requires a block copy, then sync the RHS back
2120 to the original structure, leave the original assignment
2121 statement (which will perform the block copy), then load the
2122 LHS values out of its now-updated original structure. */
2123 /* ??? Could perform a modified pair-wise element copy. That
2124 would at least allow those elements that are instantiated in
2125 both structures to be optimized well. */
2128 generate_copy_inout (rhs_elt, false,
2129 generate_element_ref (rhs_elt), &list);
2132 mark_all_v_defs (list);
2133 sra_insert_before (bsi, list);
2137 generate_copy_inout (lhs_elt, true,
2138 generate_element_ref (lhs_elt), &list);
2141 mark_all_v_defs (list);
2142 sra_insert_after (bsi, list);
2147 /* Otherwise both sides must be fully instantiated. In which
2148 case perform pair-wise element assignments and replace the
2149 original block copy statement. */
2151 stmt = bsi_stmt (*bsi);
2152 mark_all_v_defs (stmt);
2155 generate_element_copy (lhs_elt, rhs_elt, &list);
2157 mark_all_v_defs (list);
2158 sra_replace (bsi, list);
2162 /* Scalarize an INIT. To recap, this is an assignment to a scalarizable
2163 reference from some form of constructor: CONSTRUCTOR, COMPLEX_CST or
2164 COMPLEX_EXPR. If RHS is NULL, it should be treated as an empty
2168 scalarize_init (struct sra_elt *lhs_elt, tree rhs, block_stmt_iterator *bsi)
2173 /* Generate initialization statements for all members extant in the RHS. */
2176 /* Unshare the expression just in case this is from a decl's initial. */
2177 rhs = unshare_expr (rhs);
2178 result = generate_element_init (lhs_elt, rhs, &list);
2181 /* CONSTRUCTOR is defined such that any member not mentioned is assigned
2182 a zero value. Initialize the rest of the instantiated elements. */
2183 generate_element_zero (lhs_elt, &list);
2187 /* If we failed to convert the entire initializer, then we must
2188 leave the structure assignment in place and must load values
2189 from the structure into the slots for which we did not find
2190 constants. The easiest way to do this is to generate a complete
2191 copy-out, and then follow that with the constant assignments
2192 that we were able to build. DCE will clean things up. */
2194 generate_copy_inout (lhs_elt, true, generate_element_ref (lhs_elt),
2196 append_to_statement_list (list, &list0);
2200 if (lhs_elt->use_block_copy || !result)
2202 /* Since LHS is not fully instantiated, we must leave the structure
2203 assignment in place. Treating this case differently from a USE
2204 exposes constants to later optimizations. */
2207 mark_all_v_defs (list);
2208 sra_insert_after (bsi, list);
2213 /* The LHS is fully instantiated. The list of initializations
2214 replaces the original structure assignment. */
2216 mark_all_v_defs (bsi_stmt (*bsi));
2217 mark_all_v_defs (list);
2218 sra_replace (bsi, list);
2222 /* A subroutine of scalarize_ldst called via walk_tree. Set TREE_NO_TRAP
2223 on all INDIRECT_REFs. */
2226 mark_notrap (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
2230 if (TREE_CODE (t) == INDIRECT_REF)
2232 TREE_THIS_NOTRAP (t) = 1;
2235 else if (IS_TYPE_OR_DECL_P (t))
2241 /* Scalarize a LDST. To recap, this is an assignment between one scalarizable
2242 reference ELT and one non-scalarizable reference OTHER. IS_OUTPUT is true
2243 if ELT is on the left-hand side. */
2246 scalarize_ldst (struct sra_elt *elt, tree other,
2247 block_stmt_iterator *bsi, bool is_output)
2249 /* Shouldn't have gotten called for a scalar. */
2250 gcc_assert (!elt->replacement);
2252 if (elt->use_block_copy)
2254 /* Since ELT is not fully instantiated, we have to leave the
2255 block copy in place. Treat this as a USE. */
2256 scalarize_use (elt, NULL, bsi, is_output, false);
2260 /* The interesting case is when ELT is fully instantiated. In this
2261 case we can have each element stored/loaded directly to/from the
2262 corresponding slot in OTHER. This avoids a block copy. */
2264 tree list = NULL, stmt = bsi_stmt (*bsi);
2266 mark_all_v_defs (stmt);
2267 generate_copy_inout (elt, is_output, other, &list);
2268 mark_all_v_defs (list);
2271 /* Preserve EH semantics. */
2272 if (stmt_ends_bb_p (stmt))
2274 tree_stmt_iterator tsi;
2277 /* Extract the first statement from LIST. */
2278 tsi = tsi_start (list);
2279 first = tsi_stmt (tsi);
2282 /* Replace the old statement with this new representative. */
2283 bsi_replace (bsi, first, true);
2285 if (!tsi_end_p (tsi))
2287 /* If any reference would trap, then they all would. And more
2288 to the point, the first would. Therefore none of the rest
2289 will trap since the first didn't. Indicate this by
2290 iterating over the remaining statements and set
2291 TREE_THIS_NOTRAP in all INDIRECT_REFs. */
2294 walk_tree (tsi_stmt_ptr (tsi), mark_notrap, NULL, NULL);
2297 while (!tsi_end_p (tsi));
2299 insert_edge_copies (list, bsi->bb);
2303 sra_replace (bsi, list);
2307 /* Generate initializations for all scalarizable parameters. */
2310 scalarize_parms (void)
2316 EXECUTE_IF_SET_IN_BITMAP (needs_copy_in, 0, i, bi)
2318 tree var = referenced_var (i);
2319 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
2320 generate_copy_inout (elt, true, var, &list);
2325 insert_edge_copies (list, ENTRY_BLOCK_PTR);
2326 mark_all_v_defs (list);
2330 /* Entry point to phase 4. Update the function to match replacements. */
2333 scalarize_function (void)
2335 static const struct sra_walk_fns fns = {
2336 scalarize_use, scalarize_copy, scalarize_init, scalarize_ldst, false
2339 sra_walk_function (&fns);
2341 bsi_commit_edge_inserts ();
2345 /* Debug helper function. Print ELT in a nice human-readable format. */
2348 dump_sra_elt_name (FILE *f, struct sra_elt *elt)
2350 if (elt->parent && TREE_CODE (elt->parent->type) == COMPLEX_TYPE)
2352 fputs (elt->element == integer_zero_node ? "__real__ " : "__imag__ ", f);
2353 dump_sra_elt_name (f, elt->parent);
2358 dump_sra_elt_name (f, elt->parent);
2359 if (DECL_P (elt->element))
2361 if (TREE_CODE (elt->element) == FIELD_DECL)
2363 print_generic_expr (f, elt->element, dump_flags);
2365 else if (TREE_CODE (elt->element) == RANGE_EXPR)
2366 fprintf (f, "["HOST_WIDE_INT_PRINT_DEC".."HOST_WIDE_INT_PRINT_DEC"]",
2367 TREE_INT_CST_LOW (TREE_OPERAND (elt->element, 0)),
2368 TREE_INT_CST_LOW (TREE_OPERAND (elt->element, 1)));
2370 fprintf (f, "[" HOST_WIDE_INT_PRINT_DEC "]",
2371 TREE_INT_CST_LOW (elt->element));
2375 /* Likewise, but callable from the debugger. */
2378 debug_sra_elt_name (struct sra_elt *elt)
2380 dump_sra_elt_name (stderr, elt);
2381 fputc ('\n', stderr);
2385 sra_init_cache (void)
2387 if (sra_type_decomp_cache)
2390 sra_type_decomp_cache = BITMAP_ALLOC (NULL);
2391 sra_type_inst_cache = BITMAP_ALLOC (NULL);
2394 /* Main entry point. */
2399 /* Initialize local variables. */
2401 gcc_obstack_init (&sra_obstack);
2402 sra_candidates = BITMAP_ALLOC (NULL);
2403 needs_copy_in = BITMAP_ALLOC (NULL);
2405 sra_map = htab_create (101, sra_elt_hash, sra_elt_eq, NULL);
2407 /* Scan. If we find anything, instantiate and scalarize. */
2408 if (find_candidates_for_sra ())
2411 decide_instantiations ();
2412 scalarize_function ();
2413 if (!bitmap_empty_p (sra_candidates))
2414 todoflags |= TODO_rebuild_alias;
2417 /* Free allocated memory. */
2418 htab_delete (sra_map);
2420 BITMAP_FREE (sra_candidates);
2421 BITMAP_FREE (needs_copy_in);
2422 BITMAP_FREE (sra_type_decomp_cache);
2423 BITMAP_FREE (sra_type_inst_cache);
2424 obstack_free (&sra_obstack, NULL);
2429 tree_sra_early (void)
2437 return ret & ~TODO_rebuild_alias;
2443 return flag_tree_sra != 0;
2446 struct tree_opt_pass pass_sra_early =
2449 gate_sra, /* gate */
2450 tree_sra_early, /* execute */
2453 0, /* static_pass_number */
2454 TV_TREE_SRA, /* tv_id */
2455 PROP_cfg | PROP_ssa, /* properties_required */
2456 0, /* properties_provided */
2457 0, /* properties_destroyed */
2458 0, /* todo_flags_start */
2462 | TODO_verify_ssa, /* todo_flags_finish */
2466 struct tree_opt_pass pass_sra =
2469 gate_sra, /* gate */
2470 tree_sra, /* execute */
2473 0, /* static_pass_number */
2474 TV_TREE_SRA, /* tv_id */
2475 PROP_cfg | PROP_ssa, /* properties_required */
2476 0, /* properties_provided */
2477 0, /* properties_destroyed */
2478 0, /* todo_flags_start */
2482 | TODO_verify_ssa, /* todo_flags_finish */