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 == ENUMERAL_TYPE || code == BOOLEAN_TYPE
224 || code == POINTER_TYPE || code == OFFSET_TYPE
225 || code == REFERENCE_TYPE);
228 /* Return true if TYPE can be decomposed into a set of independent variables.
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. */
235 sra_type_can_be_decomposed_p (tree type)
237 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
240 /* Avoid searching the same type twice. */
241 if (bitmap_bit_p (sra_type_decomp_cache, cache+0))
243 if (bitmap_bit_p (sra_type_decomp_cache, cache+1))
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)))
251 /* The type must be a non-union aggregate. */
252 switch (TREE_CODE (type))
256 bool saw_one_field = false;
258 for (t = TYPE_FIELDS (type); t ; t = TREE_CHAIN (t))
259 if (TREE_CODE (t) == FIELD_DECL)
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))))
267 saw_one_field = true;
270 /* Record types must have at least one field. */
277 /* Array types must have a fixed lower and upper bound. */
278 t = TYPE_DOMAIN (type);
281 if (TYPE_MIN_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MIN_VALUE (t)))
283 if (TYPE_MAX_VALUE (t) == NULL || !TREE_CONSTANT (TYPE_MAX_VALUE (t)))
294 bitmap_set_bit (sra_type_decomp_cache, cache+0);
298 bitmap_set_bit (sra_type_decomp_cache, cache+1);
302 /* Return true if DECL can be decomposed into a set of independent
303 (though not necessarily scalar) variables. */
306 decl_can_be_decomposed_p (tree var)
308 /* Early out for scalars. */
309 if (is_sra_scalar_type (TREE_TYPE (var)))
312 /* The variable must not be aliased. */
313 if (!is_gimple_non_addressable (var))
315 if (dump_file && (dump_flags & TDF_DETAILS))
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");
324 /* The variable must not be volatile. */
325 if (TREE_THIS_VOLATILE (var))
327 if (dump_file && (dump_flags & TDF_DETAILS))
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");
336 /* We must be able to decompose the variable's type. */
337 if (!sra_type_can_be_decomposed_p (TREE_TYPE (var)))
339 if (dump_file && (dump_flags & TDF_DETAILS))
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");
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. */
355 && TYPE_MAIN_VARIANT (TREE_TYPE (var))
356 == TYPE_MAIN_VARIANT (va_list_type_node))
362 /* Return true if TYPE can be *completely* decomposed into scalars. */
365 type_can_instantiate_all_elements (tree type)
367 if (is_sra_scalar_type (type))
369 if (!sra_type_can_be_decomposed_p (type))
372 switch (TREE_CODE (type))
376 unsigned int cache = TYPE_UID (TYPE_MAIN_VARIANT (type)) * 2;
379 if (bitmap_bit_p (sra_type_inst_cache, cache+0))
381 if (bitmap_bit_p (sra_type_inst_cache, cache+1))
384 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
385 if (TREE_CODE (f) == FIELD_DECL)
387 if (!type_can_instantiate_all_elements (TREE_TYPE (f)))
389 bitmap_set_bit (sra_type_inst_cache, cache+1);
394 bitmap_set_bit (sra_type_inst_cache, cache+0);
399 return type_can_instantiate_all_elements (TREE_TYPE (type));
409 /* Test whether ELT or some sub-element cannot be scalarized. */
412 can_completely_scalarize_p (struct sra_elt *elt)
416 if (elt->cannot_scalarize)
419 for (c = elt->children; c; c = c->sibling)
420 if (!can_completely_scalarize_p (c))
423 for (c = elt->groups; c; c = c->sibling)
424 if (!can_completely_scalarize_p (c))
431 /* A simplified tree hashing algorithm that only handles the types of
432 trees we expect to find in sra_elt->element. */
435 sra_hash_tree (tree t)
439 switch (TREE_CODE (t))
448 h = TREE_INT_CST_LOW (t) ^ TREE_INT_CST_HIGH (t);
452 h = iterative_hash_expr (TREE_OPERAND (t, 0), 0);
453 h = iterative_hash_expr (TREE_OPERAND (t, 1), h);
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);
470 /* Hash function for type SRA_PAIR. */
473 sra_elt_hash (const void *x)
475 const struct sra_elt *e = x;
476 const struct sra_elt *p;
479 h = sra_hash_tree (e->element);
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);
491 /* Equality function for type SRA_PAIR. */
494 sra_elt_eq (const void *x, const void *y)
496 const struct sra_elt *a = x;
497 const struct sra_elt *b = y;
500 if (a->parent != b->parent)
508 if (TREE_CODE (ae) != TREE_CODE (be))
511 switch (TREE_CODE (ae))
516 /* These are all pointer unique. */
520 /* Integers are not pointer unique, so compare their values. */
521 return tree_int_cst_equal (ae, be);
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));
529 /* Fields are unique within a record, but not between
530 compatible records. */
531 if (DECL_FIELD_CONTEXT (ae) == DECL_FIELD_CONTEXT (be))
533 return fields_compatible_p (ae, be);
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. */
543 static struct sra_elt *
544 lookup_element (struct sra_elt *parent, tree child, tree type,
545 enum insert_option insert)
547 struct sra_elt dummy;
548 struct sra_elt **slot;
552 dummy.parent = parent->is_group ? parent->parent : parent;
555 dummy.element = child;
557 slot = (struct sra_elt **) htab_find_slot (sra_map, &dummy, insert);
558 if (!slot && insert == NO_INSERT)
562 if (!elt && insert == INSERT)
564 *slot = elt = obstack_alloc (&sra_obstack, sizeof (*elt));
565 memset (elt, 0, sizeof (*elt));
567 elt->parent = parent;
568 elt->element = child;
570 elt->is_scalar = is_sra_scalar_type (type);
574 if (IS_ELEMENT_FOR_GROUP (elt->element))
576 elt->is_group = true;
577 elt->sibling = parent->groups;
578 parent->groups = elt;
582 elt->sibling = parent->children;
583 parent->children = elt;
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)
592 bitmap_set_bit (needs_copy_in, DECL_UID (child));
599 /* Create or return the SRA_ELT structure for EXPR if the expression
600 refers to a scalarizable variable. */
602 static struct sra_elt *
603 maybe_lookup_element_for_expr (tree expr)
608 switch (TREE_CODE (expr))
613 if (is_sra_candidate_decl (expr))
614 return lookup_element (NULL, expr, TREE_TYPE (expr), INSERT);
618 /* We can't scalarize variable array indices. */
619 if (in_array_bounds_p (expr))
620 child = TREE_OPERAND (expr, 1);
625 case ARRAY_RANGE_REF:
626 /* We can't scalarize variable array indices. */
627 if (range_in_array_bounds_p (expr))
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));
638 /* Don't look through unions. */
639 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (expr, 0))) != RECORD_TYPE)
641 child = TREE_OPERAND (expr, 1);
645 child = integer_zero_node;
648 child = integer_one_node;
655 elt = maybe_lookup_element_for_expr (TREE_OPERAND (expr, 0));
657 return lookup_element (elt, child, TREE_TYPE (expr), INSERT);
662 /* Functions to walk just enough of the tree to see all scalarizable
663 references, and categorize them. */
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. */
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);
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);
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);
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);
692 /* True during phase 2, false during phase 4. */
693 /* ??? This is a hack. */
697 #ifdef ENABLE_CHECKING
698 /* Invoked via walk_tree, if *TP contains a candidate decl, return it. */
701 sra_find_candidate_decl (tree *tp, int *walk_subtrees,
702 void *data ATTRIBUTE_UNUSED)
705 enum tree_code code = TREE_CODE (t);
707 if (code == VAR_DECL || code == PARM_DECL || code == RESULT_DECL)
710 if (is_sra_candidate_decl (t))
720 /* Walk most expressions looking for a scalarizable aggregate.
721 If we find one, invoke FNS->USE. */
724 sra_walk_expr (tree *expr_p, block_stmt_iterator *bsi, bool is_output,
725 const struct sra_walk_fns *fns)
729 bool disable_scalarization = false;
730 bool use_all_p = false;
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. */
739 switch (TREE_CODE (inner))
744 /* If there is a scalarizable decl at the bottom, then process it. */
745 if (is_sra_candidate_decl (inner))
747 struct sra_elt *elt = maybe_lookup_element_for_expr (expr);
748 if (disable_scalarization)
749 elt->cannot_scalarize = true;
751 fns->use (elt, expr_p, bsi, is_output, use_all_p);
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
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))
768 disable_scalarization = true;
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))
775 inner = TREE_OPERAND (inner, 0);
778 case ARRAY_RANGE_REF:
779 if (!range_in_array_bounds_p (inner))
781 disable_scalarization = true;
784 /* ??? See above non-constant bounds and stride . */
785 if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
787 inner = TREE_OPERAND (inner, 0);
791 /* A reference to a union member constitutes a reference to the
793 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) != RECORD_TYPE)
795 /* ??? See above re non-constant stride. */
796 if (TREE_OPERAND (inner, 2))
798 inner = TREE_OPERAND (inner, 0);
803 inner = TREE_OPERAND (inner, 0);
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. */
811 && TREE_CODE (TREE_TYPE (TREE_OPERAND (inner, 0))) == VECTOR_TYPE)
814 = maybe_lookup_element_for_expr (TREE_OPERAND (inner, 0));
816 elt->is_vector_lhs = true;
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. */
824 case VIEW_CONVERT_EXPR:
826 /* Similarly, a view/nop explicitly wants to look at an object in a
827 type other than the one we've scalarized. */
831 /* This is a transparent wrapper. The entire inner expression really
836 expr_p = &TREE_OPERAND (inner, 0);
837 inner = expr = *expr_p;
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));
850 /* Walk a TREE_LIST of values looking for scalarizable aggregates.
851 If we find one, invoke FNS->USE. */
854 sra_walk_tree_list (tree list, block_stmt_iterator *bsi, bool is_output,
855 const struct sra_walk_fns *fns)
858 for (op = list; op ; op = TREE_CHAIN (op))
859 sra_walk_expr (&TREE_VALUE (op), bsi, is_output, fns);
862 /* Walk the arguments of a CALL_EXPR looking for scalarizable aggregates.
863 If we find one, invoke FNS->USE. */
866 sra_walk_call_expr (tree expr, block_stmt_iterator *bsi,
867 const struct sra_walk_fns *fns)
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);
875 /* Walk the inputs and outputs of an ASM_EXPR looking for scalarizable
876 aggregates. If we find one, invoke FNS->USE. */
879 sra_walk_asm_expr (tree expr, block_stmt_iterator *bsi,
880 const struct sra_walk_fns *fns)
882 sra_walk_tree_list (ASM_INPUTS (expr), bsi, false, fns);
883 sra_walk_tree_list (ASM_OUTPUTS (expr), bsi, true, fns);
886 /* Walk a GIMPLE_MODIFY_STMT and categorize the assignment appropriately. */
889 sra_walk_gimple_modify_stmt (tree expr, block_stmt_iterator *bsi,
890 const struct sra_walk_fns *fns)
892 struct sra_elt *lhs_elt, *rhs_elt;
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);
900 /* If both sides are scalarizable, this is a COPY operation. */
901 if (lhs_elt && rhs_elt)
903 fns->copy (lhs_elt, rhs_elt, bsi);
907 /* If the RHS is scalarizable, handle it. There are only two cases. */
910 if (!rhs_elt->is_scalar && !TREE_SIDE_EFFECTS (lhs))
911 fns->ldst (rhs_elt, lhs, bsi, false);
913 fns->use (rhs_elt, &GIMPLE_STMT_OPERAND (expr, 1), bsi, false, false);
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. */
923 tree call = get_call_expr_in (rhs);
925 sra_walk_call_expr (call, bsi, fns);
927 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 1), bsi, false, fns);
930 /* Likewise, handle the LHS being scalarizable. We have cases similar
931 to those above, but also want to handle RHS being constant. */
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);
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
945 && TREE_READONLY (rhs)
946 && targetm.binds_local_p (rhs))
947 fns->init (lhs_elt, DECL_INITIAL (rhs), bsi);
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);
957 /* Otherwise we're being used in some context that requires the
958 aggregate to be seen as a whole. Invoke USE. */
960 fns->use (lhs_elt, &GIMPLE_STMT_OPERAND (expr, 0), bsi, true, false);
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. */
967 sra_walk_expr (&GIMPLE_STMT_OPERAND (expr, 0), bsi, true, fns);
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. */
975 sra_walk_function (const struct sra_walk_fns *fns)
978 block_stmt_iterator si, ni;
980 /* ??? Phase 4 could derive some benefit to walking the function in
981 dominator tree order. */
984 for (si = bsi_start (bb); !bsi_end_p (si); si = ni)
989 stmt = bsi_stmt (si);
990 ann = stmt_ann (stmt);
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)))
1001 switch (TREE_CODE (stmt))
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.
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. */
1013 t = TREE_OPERAND (stmt, 0);
1016 else if (TREE_CODE (t) == GIMPLE_MODIFY_STMT)
1017 sra_walk_expr (&GIMPLE_STMT_OPERAND (t, 1), &si, false, fns);
1019 sra_walk_expr (&TREE_OPERAND (stmt, 0), &si, false, fns);
1022 case GIMPLE_MODIFY_STMT:
1023 sra_walk_gimple_modify_stmt (stmt, &si, fns);
1026 sra_walk_call_expr (stmt, &si, fns);
1029 sra_walk_asm_expr (stmt, &si, fns);
1038 /* Phase One: Scan all referenced variables in the program looking for
1039 structures that could be decomposed. */
1042 find_candidates_for_sra (void)
1044 bool any_set = false;
1046 referenced_var_iterator rvi;
1048 FOR_EACH_REFERENCED_VAR (var, rvi)
1050 if (decl_can_be_decomposed_p (var))
1052 bitmap_set_bit (sra_candidates, DECL_UID (var));
1061 /* Phase Two: Scan all references to scalarizable variables. Count the
1062 number of times they are used or copied respectively. */
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. */
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)
1077 scan_copy (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
1078 block_stmt_iterator *bsi ATTRIBUTE_UNUSED)
1080 lhs_elt->n_copies += 1;
1081 rhs_elt->n_copies += 1;
1085 scan_init (struct sra_elt *lhs_elt, tree rhs ATTRIBUTE_UNUSED,
1086 block_stmt_iterator *bsi ATTRIBUTE_UNUSED)
1088 lhs_elt->n_copies += 1;
1092 scan_ldst (struct sra_elt *elt, tree other ATTRIBUTE_UNUSED,
1093 block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
1094 bool is_output ATTRIBUTE_UNUSED)
1099 /* Dump the values we collected during the scanning phase. */
1102 scan_dump (struct sra_elt *elt)
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);
1109 for (c = elt->children; c ; c = c->sibling)
1112 for (c = elt->groups; c ; c = c->sibling)
1116 /* Entry point to phase 2. Scan the entire function, building up
1117 scalarization data structures, recording copies and uses. */
1120 scan_function (void)
1122 static const struct sra_walk_fns fns = {
1123 scan_use, scan_copy, scan_init, scan_ldst, true
1127 sra_walk_function (&fns);
1129 if (dump_file && (dump_flags & TDF_DETAILS))
1133 fputs ("\nScan results:\n", dump_file);
1134 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1136 tree var = referenced_var (i);
1137 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1141 fputc ('\n', dump_file);
1145 /* Phase Three: Make decisions about which variables to scalarize, if any.
1146 All elements to be scalarized have replacement variables made for them. */
1148 /* A subroutine of build_element_name. Recursively build the element
1149 name on the obstack. */
1152 build_element_name_1 (struct sra_elt *elt)
1159 build_element_name_1 (elt->parent);
1160 obstack_1grow (&sra_obstack, '$');
1162 if (TREE_CODE (elt->parent->type) == COMPLEX_TYPE)
1164 if (elt->element == integer_zero_node)
1165 obstack_grow (&sra_obstack, "real", 4);
1167 obstack_grow (&sra_obstack, "imag", 4);
1173 if (TREE_CODE (t) == INTEGER_CST)
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));
1181 tree name = DECL_NAME (t);
1183 obstack_grow (&sra_obstack, IDENTIFIER_POINTER (name),
1184 IDENTIFIER_LENGTH (name));
1187 sprintf (buffer, "D%u", DECL_UID (t));
1188 obstack_grow (&sra_obstack, buffer, strlen (buffer));
1193 /* Construct a pretty variable name for an element's replacement variable.
1194 The name is built on the obstack. */
1197 build_element_name (struct sra_elt *elt)
1199 build_element_name_1 (elt);
1200 obstack_1grow (&sra_obstack, '\0');
1201 return XOBFINISH (&sra_obstack, char *);
1204 /* Instantiate an element as an independent variable. */
1207 instantiate_element (struct sra_elt *elt)
1209 struct sra_elt *base_elt;
1212 for (base_elt = elt; base_elt->parent; base_elt = base_elt->parent)
1214 base = base_elt->element;
1216 elt->replacement = var = make_rename_temp (elt->type, "SR");
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;
1223 DECL_SOURCE_LOCATION (var) = DECL_SOURCE_LOCATION (base);
1224 DECL_ARTIFICIAL (var) = 1;
1226 if (TREE_THIS_VOLATILE (elt->type))
1228 TREE_THIS_VOLATILE (var) = 1;
1229 TREE_SIDE_EFFECTS (var) = 1;
1232 if (DECL_NAME (base) && !DECL_IGNORED_P (base))
1234 char *pretty_name = build_element_name (elt);
1235 DECL_NAME (var) = get_identifier (pretty_name);
1236 obstack_free (&sra_obstack, pretty_name);
1238 SET_DECL_DEBUG_EXPR (var, generate_element_ref (elt));
1239 DECL_DEBUG_EXPR_IS_FROM (var) = 1;
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;
1248 DECL_IGNORED_P (var) = 1;
1249 /* ??? We can't generate any warning that would be meaningful. */
1250 TREE_NO_WARNING (var) = 1;
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);
1263 /* Make one pass across an element tree deciding whether or not it's
1264 profitable to instantiate individual leaf scalars.
1266 PARENT_USES and PARENT_COPIES are the sum of the N_USES and N_COPIES
1267 fields all the way up the tree. */
1270 decide_instantiation_1 (struct sra_elt *elt, unsigned int parent_uses,
1271 unsigned int parent_copies)
1273 if (dump_file && !elt->parent)
1275 fputs ("Initial instantiation for ", dump_file);
1276 dump_sra_elt_name (dump_file, elt);
1277 fputc ('\n', dump_file);
1280 if (elt->cannot_scalarize)
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);
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;
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)
1301 c->n_uses += group->n_uses;
1302 c->n_copies += group->n_copies;
1305 for (c = elt->children; c ; c = c->sibling)
1306 decide_instantiation_1 (c, this_uses, this_copies);
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. */
1315 sum_instantiated_sizes (struct sra_elt *elt, unsigned HOST_WIDE_INT *sizep)
1317 if (elt->replacement)
1319 *sizep += TREE_INT_CST_LOW (TYPE_SIZE_UNIT (elt->type));
1325 unsigned int count = 0;
1327 for (c = elt->children; c ; c = c->sibling)
1328 count += sum_instantiated_sizes (c, sizep);
1334 /* Instantiate fields in ELT->TYPE that are not currently present as
1337 static void instantiate_missing_elements (struct sra_elt *elt);
1340 instantiate_missing_elements_1 (struct sra_elt *elt, tree child, tree type)
1342 struct sra_elt *sub = lookup_element (elt, child, type, INSERT);
1345 if (sub->replacement == NULL)
1346 instantiate_element (sub);
1349 instantiate_missing_elements (sub);
1353 instantiate_missing_elements (struct sra_elt *elt)
1355 tree type = elt->type;
1357 switch (TREE_CODE (type))
1362 for (f = TYPE_FIELDS (type); f ; f = TREE_CHAIN (f))
1363 if (TREE_CODE (f) == FIELD_DECL)
1365 tree field_type = TREE_TYPE (f);
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
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,
1379 instantiate_missing_elements_1 (elt, f, field_type);
1386 tree i, max, subtype;
1388 i = TYPE_MIN_VALUE (TYPE_DOMAIN (type));
1389 max = TYPE_MAX_VALUE (TYPE_DOMAIN (type));
1390 subtype = TREE_TYPE (type);
1394 instantiate_missing_elements_1 (elt, i, subtype);
1395 if (tree_int_cst_equal (i, max))
1397 i = int_const_binop (PLUS_EXPR, i, integer_one_node, true);
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);
1414 /* Return true if there is only one non aggregate field in the record, TYPE.
1415 Return false otherwise. */
1418 single_scalar_field_in_record_p (tree type)
1422 if (TREE_CODE (type) != RECORD_TYPE)
1425 for (field = TYPE_FIELDS (type); field; field = TREE_CHAIN (field))
1426 if (TREE_CODE (field) == FIELD_DECL)
1430 if (num_fields == 2)
1433 if (AGGREGATE_TYPE_P (TREE_TYPE (field)))
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. */
1445 decide_block_copy (struct sra_elt *elt)
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);
1454 /* If scalarization is disabled, respect it. */
1455 if (elt->cannot_scalarize)
1457 elt->use_block_copy = 1;
1461 fputs ("Scalarization disabled for ", dump_file);
1462 dump_sra_elt_name (dump_file, elt);
1463 fputc ('\n', dump_file);
1466 /* Disable scalarization of sub-elements */
1467 for (c = elt->children; c; c = c->sibling)
1469 c->cannot_scalarize = 1;
1470 decide_block_copy (c);
1473 /* Groups behave like their parent. */
1474 for (c = elt->groups; c; c = c->sibling)
1476 c->cannot_scalarize = 1;
1477 c->use_block_copy = 1;
1483 /* Don't decide if we've no uses. */
1484 if (elt->n_uses == 0 && elt->n_copies == 0)
1487 else if (!elt->is_scalar)
1489 tree size_tree = TYPE_SIZE_UNIT (elt->type);
1490 bool use_block_copy = true;
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;
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))
1502 unsigned HOST_WIDE_INT full_size, inst_size = 0;
1503 unsigned int max_size, max_count, inst_count, full_count;
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
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);
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;
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. */
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;
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. */
1541 && (!can_completely_scalarize_p (elt)
1542 || !type_can_instantiate_all_elements (elt->type)))
1543 use_block_copy = true;
1546 elt->use_block_copy = use_block_copy;
1548 /* Groups behave like their parent. */
1549 for (c = elt->groups; c; c = c->sibling)
1550 c->use_block_copy = use_block_copy;
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);
1560 if (!use_block_copy)
1562 instantiate_missing_elements (elt);
1567 any_inst = elt->replacement != NULL;
1569 for (c = elt->children; c ; c = c->sibling)
1570 any_inst |= decide_block_copy (c);
1575 /* Entry point to phase 3. Instantiate scalar replacement variables. */
1578 decide_instantiations (void)
1582 bitmap_head done_head;
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;
1590 EXECUTE_IF_SET_IN_BITMAP (sra_candidates, 0, i, bi)
1592 tree var = referenced_var (i);
1593 struct sra_elt *elt = lookup_element (NULL, var, NULL, NO_INSERT);
1596 decide_instantiation_1 (elt, 0, 0);
1597 if (!decide_block_copy (elt))
1602 bitmap_set_bit (&done_head, i);
1609 bitmap_and_compl_into (sra_candidates, &done_head);
1610 bitmap_and_compl_into (needs_copy_in, &done_head);
1612 bitmap_clear (&done_head);
1614 mark_set_for_renaming (sra_candidates);
1617 fputc ('\n', dump_file);
1621 /* Phase Four: Update the function to match the replacements created. */
1623 /* Mark all the variables in VDEF/VUSE operators for STMT for
1624 renaming. This becomes necessary when we modify all of a
1628 mark_all_v_defs_1 (tree stmt)
1633 update_stmt_if_modified (stmt);
1635 FOR_EACH_SSA_TREE_OPERAND (sym, stmt, iter, SSA_OP_ALL_VIRTUALS)
1637 if (TREE_CODE (sym) == SSA_NAME)
1638 sym = SSA_NAME_VAR (sym);
1639 mark_sym_for_renaming (sym);
1644 /* Mark all the variables in virtual operands in all the statements in
1645 LIST for renaming. */
1648 mark_all_v_defs (tree list)
1650 if (TREE_CODE (list) != STATEMENT_LIST)
1651 mark_all_v_defs_1 (list);
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));
1661 /* Mark every replacement under ELT with TREE_NO_WARNING. */
1664 mark_no_warning (struct sra_elt *elt)
1666 if (!elt->all_no_warning)
1668 if (elt->replacement)
1669 TREE_NO_WARNING (elt->replacement) = 1;
1673 FOR_EACH_ACTUAL_CHILD (c, elt)
1674 mark_no_warning (c);
1676 elt->all_no_warning = true;
1680 /* Build a single level component reference to ELT rooted at BASE. */
1683 generate_one_element_ref (struct sra_elt *elt, tree base)
1685 switch (TREE_CODE (TREE_TYPE (base)))
1689 tree field = elt->element;
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);
1695 return build3 (COMPONENT_REF, elt->type, base, field, NULL);
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);
1703 return build4 (ARRAY_REF, elt->type, base, elt->element, NULL, NULL);
1706 if (elt->element == integer_zero_node)
1707 return build1 (REALPART_EXPR, elt->type, base);
1709 return build1 (IMAGPART_EXPR, elt->type, base);
1716 /* Build a full component reference to ELT rooted at its native variable. */
1719 generate_element_ref (struct sra_elt *elt)
1722 return generate_one_element_ref (elt, generate_element_ref (elt->parent));
1724 return elt->element;
1727 /* Create an assignment statement from SRC to DST. */
1730 sra_build_assignment (tree dst, tree src)
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);
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. */
1749 generate_copy_inout (struct sra_elt *elt, bool copy_out, tree expr,
1755 if (!copy_out && TREE_CODE (expr) == SSA_NAME
1756 && TREE_CODE (TREE_TYPE (expr)) == COMPLEX_TYPE)
1760 c = lookup_element (elt, integer_zero_node, NULL, NO_INSERT);
1762 c = lookup_element (elt, integer_one_node, NULL, NO_INSERT);
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);
1770 else if (elt->replacement)
1773 t = sra_build_assignment (elt->replacement, expr);
1775 t = sra_build_assignment (expr, elt->replacement);
1776 append_to_statement_list (t, list_p);
1780 FOR_EACH_ACTUAL_CHILD (c, elt)
1782 t = generate_one_element_ref (c, unshare_expr (expr));
1783 generate_copy_inout (c, copy_out, t, list_p);
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. */
1793 generate_element_copy (struct sra_elt *dst, struct sra_elt *src, tree *list_p)
1795 struct sra_elt *dc, *sc;
1797 FOR_EACH_ACTUAL_CHILD (dc, dst)
1799 sc = lookup_element (src, dc->element, NULL, NO_INSERT);
1801 generate_element_copy (dc, sc, list_p);
1804 if (dst->replacement)
1808 gcc_assert (src->replacement);
1810 t = sra_build_assignment (dst->replacement, src->replacement);
1811 append_to_statement_list (t, list_p);
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. */
1821 generate_element_zero (struct sra_elt *elt, tree *list_p)
1827 elt->visited = false;
1831 FOR_EACH_ACTUAL_CHILD (c, elt)
1832 generate_element_zero (c, list_p);
1834 if (elt->replacement)
1838 gcc_assert (elt->is_scalar);
1839 t = fold_convert (elt->type, integer_zero_node);
1841 t = sra_build_assignment (elt->replacement, t);
1842 append_to_statement_list (t, list_p);
1846 /* Generate an assignment VAR = INIT, where INIT may need gimplification.
1847 Add the result to *LIST_P. */
1850 generate_one_element_init (tree var, tree init, tree *list_p)
1852 /* The replacement can be almost arbitrarily complex. Gimplify. */
1853 tree stmt = sra_build_assignment (var, init);
1854 gimplify_and_add (stmt, list_p);
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
1864 generate_element_init_1 (struct sra_elt *elt, tree init, tree *list_p)
1867 enum tree_code init_code;
1868 struct sra_elt *sub;
1870 unsigned HOST_WIDE_INT idx;
1871 tree value, purpose;
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);
1880 if (elt->replacement)
1882 generate_one_element_init (elt->replacement, init, list_p);
1883 elt->visited = true;
1892 FOR_EACH_ACTUAL_CHILD (sub, elt)
1894 if (sub->element == integer_zero_node)
1895 t = (init_code == COMPLEX_EXPR
1896 ? TREE_OPERAND (init, 0) : TREE_REALPART (init));
1898 t = (init_code == COMPLEX_EXPR
1899 ? TREE_OPERAND (init, 1) : TREE_IMAGPART (init));
1900 result &= generate_element_init_1 (sub, t, list_p);
1905 FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (init), idx, purpose, value)
1907 if (TREE_CODE (purpose) == RANGE_EXPR)
1909 tree lower = TREE_OPERAND (purpose, 0);
1910 tree upper = TREE_OPERAND (purpose, 1);
1914 sub = lookup_element (elt, lower, NULL, NO_INSERT);
1916 result &= generate_element_init_1 (sub, value, list_p);
1917 if (tree_int_cst_equal (lower, upper))
1919 lower = int_const_binop (PLUS_EXPR, lower,
1920 integer_one_node, true);
1925 sub = lookup_element (elt, purpose, NULL, NO_INSERT);
1927 result &= generate_element_init_1 (sub, value, list_p);
1933 elt->visited = true;
1940 /* A wrapper function for generate_element_init_1 that handles cleanup after
1944 generate_element_init (struct sra_elt *elt, tree init, tree *list_p)
1948 push_gimplify_context ();
1949 ret = generate_element_init_1 (elt, init, list_p);
1950 pop_gimplify_context (NULL);
1952 /* The replacement can expose previously unreferenced variables. */
1955 tree_stmt_iterator i;
1957 for (i = tsi_start (*list_p); !tsi_end_p (i); tsi_next (&i))
1958 find_new_referenced_vars (tsi_stmt_ptr (i));
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. */
1969 insert_edge_copies (tree stmt, basic_block bb)
1976 FOR_EACH_EDGE (e, ei, bb->succs)
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))
1985 bsi_insert_on_edge (e, stmt);
1989 bsi_insert_on_edge (e, unsave_expr_now (stmt));
1994 /* Helper function to insert LIST before BSI, and set up line number info. */
1997 sra_insert_before (block_stmt_iterator *bsi, tree list)
1999 tree stmt = bsi_stmt (*bsi);
2001 if (EXPR_HAS_LOCATION (stmt))
2002 annotate_all_with_locus (&list, EXPR_LOCATION (stmt));
2003 bsi_insert_before (bsi, list, BSI_SAME_STMT);
2006 /* Similarly, but insert after BSI. Handles insertion onto edges as well. */
2009 sra_insert_after (block_stmt_iterator *bsi, tree list)
2011 tree stmt = bsi_stmt (*bsi);
2013 if (EXPR_HAS_LOCATION (stmt))
2014 annotate_all_with_locus (&list, EXPR_LOCATION (stmt));
2016 if (stmt_ends_bb_p (stmt))
2017 insert_edge_copies (list, bsi->bb);
2019 bsi_insert_after (bsi, list, BSI_SAME_STMT);
2022 /* Similarly, but replace the statement at BSI. */
2025 sra_replace (block_stmt_iterator *bsi, tree list)
2027 sra_insert_before (bsi, list);
2028 bsi_remove (bsi, false);
2029 if (bsi_end_p (*bsi))
2030 *bsi = bsi_last (bsi->bb);
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. */
2040 scalarize_use (struct sra_elt *elt, tree *expr_p, block_stmt_iterator *bsi,
2041 bool is_output, bool use_all)
2043 tree list = NULL, stmt = bsi_stmt (*bsi);
2045 if (elt->replacement)
2047 /* If we have a replacement, then updating the reference is as
2048 simple as modifying the existing statement in place. */
2050 mark_all_v_defs (stmt);
2051 *expr_p = elt->replacement;
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. */
2069 generate_copy_inout (elt, is_output, generate_element_ref (elt), &list);
2072 mark_all_v_defs (list);
2074 sra_insert_after (bsi, list);
2077 sra_insert_before (bsi, list);
2079 mark_no_warning (elt);
2084 /* Scalarize a COPY. To recap, this is an assignment statement between
2085 two scalarizable references, LHS_ELT and RHS_ELT. */
2088 scalarize_copy (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
2089 block_stmt_iterator *bsi)
2093 if (lhs_elt->replacement && rhs_elt->replacement)
2095 /* If we have two scalar operands, modify the existing statement. */
2096 stmt = bsi_stmt (*bsi);
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);
2102 GIMPLE_STMT_OPERAND (stmt, 0) = lhs_elt->replacement;
2103 GIMPLE_STMT_OPERAND (stmt, 1) = rhs_elt->replacement;
2106 else if (lhs_elt->use_block_copy || rhs_elt->use_block_copy)
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. */
2117 generate_copy_inout (rhs_elt, false,
2118 generate_element_ref (rhs_elt), &list);
2121 mark_all_v_defs (list);
2122 sra_insert_before (bsi, list);
2126 generate_copy_inout (lhs_elt, true,
2127 generate_element_ref (lhs_elt), &list);
2130 mark_all_v_defs (list);
2131 sra_insert_after (bsi, list);
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. */
2140 stmt = bsi_stmt (*bsi);
2141 mark_all_v_defs (stmt);
2144 generate_element_copy (lhs_elt, rhs_elt, &list);
2146 mark_all_v_defs (list);
2147 sra_replace (bsi, list);
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
2157 scalarize_init (struct sra_elt *lhs_elt, tree rhs, block_stmt_iterator *bsi)
2162 /* Generate initialization statements for all members extant in the RHS. */
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);
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);
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. */
2183 generate_copy_inout (lhs_elt, true, generate_element_ref (lhs_elt),
2185 append_to_statement_list (list, &list0);
2189 if (lhs_elt->use_block_copy || !result)
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. */
2196 mark_all_v_defs (list);
2197 sra_insert_after (bsi, list);
2202 /* The LHS is fully instantiated. The list of initializations
2203 replaces the original structure assignment. */
2205 mark_all_v_defs (bsi_stmt (*bsi));
2206 mark_all_v_defs (list);
2207 sra_replace (bsi, list);
2211 /* A subroutine of scalarize_ldst called via walk_tree. Set TREE_NO_TRAP
2212 on all INDIRECT_REFs. */
2215 mark_notrap (tree *tp, int *walk_subtrees, void *data ATTRIBUTE_UNUSED)
2219 if (TREE_CODE (t) == INDIRECT_REF)
2221 TREE_THIS_NOTRAP (t) = 1;
2224 else if (IS_TYPE_OR_DECL_P (t))
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. */
2235 scalarize_ldst (struct sra_elt *elt, tree other,
2236 block_stmt_iterator *bsi, bool is_output)
2238 /* Shouldn't have gotten called for a scalar. */
2239 gcc_assert (!elt->replacement);
2241 if (elt->use_block_copy)
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);
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. */
2253 tree list = NULL, stmt = bsi_stmt (*bsi);
2255 mark_all_v_defs (stmt);
2256 generate_copy_inout (elt, is_output, other, &list);
2257 mark_all_v_defs (list);
2260 /* Preserve EH semantics. */
2261 if (stmt_ends_bb_p (stmt))
2263 tree_stmt_iterator tsi;
2266 /* Extract the first statement from LIST. */
2267 tsi = tsi_start (list);
2268 first = tsi_stmt (tsi);
2271 /* Replace the old statement with this new representative. */
2272 bsi_replace (bsi, first, true);
2274 if (!tsi_end_p (tsi))
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. */
2283 walk_tree (tsi_stmt_ptr (tsi), mark_notrap, NULL, NULL);
2286 while (!tsi_end_p (tsi));
2288 insert_edge_copies (list, bsi->bb);
2292 sra_replace (bsi, list);
2296 /* Generate initializations for all scalarizable parameters. */
2299 scalarize_parms (void)
2305 EXECUTE_IF_SET_IN_BITMAP (needs_copy_in, 0, i, bi)
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);
2314 insert_edge_copies (list, ENTRY_BLOCK_PTR);
2315 mark_all_v_defs (list);
2319 /* Entry point to phase 4. Update the function to match replacements. */
2322 scalarize_function (void)
2324 static const struct sra_walk_fns fns = {
2325 scalarize_use, scalarize_copy, scalarize_init, scalarize_ldst, false
2328 sra_walk_function (&fns);
2330 bsi_commit_edge_inserts ();
2334 /* Debug helper function. Print ELT in a nice human-readable format. */
2337 dump_sra_elt_name (FILE *f, struct sra_elt *elt)
2339 if (elt->parent && TREE_CODE (elt->parent->type) == COMPLEX_TYPE)
2341 fputs (elt->element == integer_zero_node ? "__real__ " : "__imag__ ", f);
2342 dump_sra_elt_name (f, elt->parent);
2347 dump_sra_elt_name (f, elt->parent);
2348 if (DECL_P (elt->element))
2350 if (TREE_CODE (elt->element) == FIELD_DECL)
2352 print_generic_expr (f, elt->element, dump_flags);
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)));
2359 fprintf (f, "[" HOST_WIDE_INT_PRINT_DEC "]",
2360 TREE_INT_CST_LOW (elt->element));
2364 /* Likewise, but callable from the debugger. */
2367 debug_sra_elt_name (struct sra_elt *elt)
2369 dump_sra_elt_name (stderr, elt);
2370 fputc ('\n', stderr);
2374 sra_init_cache (void)
2376 if (sra_type_decomp_cache)
2379 sra_type_decomp_cache = BITMAP_ALLOC (NULL);
2380 sra_type_inst_cache = BITMAP_ALLOC (NULL);
2383 /* Main entry point. */
2388 /* Initialize local variables. */
2390 gcc_obstack_init (&sra_obstack);
2391 sra_candidates = BITMAP_ALLOC (NULL);
2392 needs_copy_in = BITMAP_ALLOC (NULL);
2394 sra_map = htab_create (101, sra_elt_hash, sra_elt_eq, NULL);
2396 /* Scan. If we find anything, instantiate and scalarize. */
2397 if (find_candidates_for_sra ())
2400 decide_instantiations ();
2401 scalarize_function ();
2404 /* Free allocated memory. */
2405 htab_delete (sra_map);
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);
2416 tree_sra_early (void)
2430 return flag_tree_sra != 0;
2433 struct tree_opt_pass pass_sra_early =
2436 gate_sra, /* gate */
2437 tree_sra_early, /* execute */
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 */
2449 | TODO_verify_ssa, /* todo_flags_finish */
2453 struct tree_opt_pass pass_sra =
2456 gate_sra, /* gate */
2457 tree_sra, /* execute */
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 */
2469 | TODO_verify_ssa, /* todo_flags_finish */