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 Free Software Foundation, Inc.
5 Contributed by Diego Novillo <dnovillo@redhat.com>
7 This file is part of GCC.
9 GCC is free software; you can redistribute it and/or modify it
10 under the terms of the GNU General Public License as published by the
11 Free Software Foundation; either version 2, or (at your option) any
14 GCC is distributed in the hope that it will be useful, but WITHOUT
15 ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
19 You should have received a copy of the GNU General Public License
20 along with GCC; see the file COPYING. If not, write to the Free
21 Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
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 /* The set of todo flags to return from tree_sra. */
79 static unsigned int todoflags
;
81 /* The set of aggregate variables that are candidates for scalarization. */
82 static bitmap sra_candidates
;
84 /* Set of scalarizable PARM_DECLs that need copy-in operations at the
85 beginning of the function. */
86 static bitmap needs_copy_in
;
88 /* Sets of bit pairs that cache type decomposition and instantiation. */
89 static bitmap sra_type_decomp_cache
;
90 static bitmap sra_type_inst_cache
;
92 /* One of these structures is created for each candidate aggregate and
93 each (accessed) member or group of members of such an aggregate. */
96 /* A tree of the elements. Used when we want to traverse everything. */
97 struct sra_elt
*parent
;
98 struct sra_elt
*groups
;
99 struct sra_elt
*children
;
100 struct sra_elt
*sibling
;
102 /* If this element is a root, then this is the VAR_DECL. If this is
103 a sub-element, this is some token used to identify the reference.
104 In the case of COMPONENT_REF, this is the FIELD_DECL. In the case
105 of an ARRAY_REF, this is the (constant) index. In the case of an
106 ARRAY_RANGE_REF, this is the (constant) RANGE_EXPR. In the case
107 of a complex number, this is a zero or one. */
110 /* The type of the element. */
113 /* A VAR_DECL, for any sub-element we've decided to replace. */
116 /* The number of times the element is referenced as a whole. I.e.
117 given "a.b.c", this would be incremented for C, but not for A or B. */
120 /* The number of times the element is copied to or from another
121 scalarizable element. */
122 unsigned int n_copies
;
124 /* True if TYPE is scalar. */
127 /* True if this element is a group of members of its parent. */
130 /* True if we saw something about this element that prevents scalarization,
131 such as non-constant indexing. */
132 bool cannot_scalarize
;
134 /* True if we've decided that structure-to-structure assignment
135 should happen via memcpy and not per-element. */
138 /* True if everything under this element has been marked TREE_NO_WARNING. */
141 /* A flag for use with/after random access traversals. */
145 #define IS_ELEMENT_FOR_GROUP(ELEMENT) (TREE_CODE (ELEMENT) == RANGE_EXPR)
147 #define FOR_EACH_ACTUAL_CHILD(CHILD, ELT) \
148 for ((CHILD) = (ELT)->is_group \
149 ? next_child_for_group (NULL, (ELT)) \
152 (CHILD) = (ELT)->is_group \
153 ? next_child_for_group ((CHILD), (ELT)) \
156 /* Helper function for above macro. Return next child in group. */
157 static struct sra_elt
*
158 next_child_for_group (struct sra_elt
*child
, struct sra_elt
*group
)
160 gcc_assert (group
->is_group
);
162 /* Find the next child in the parent. */
164 child
= child
->sibling
;
166 child
= group
->parent
->children
;
168 /* Skip siblings that do not belong to the group. */
171 tree g_elt
= group
->element
;
172 if (TREE_CODE (g_elt
) == RANGE_EXPR
)
174 if (!tree_int_cst_lt (child
->element
, TREE_OPERAND (g_elt
, 0))
175 && !tree_int_cst_lt (TREE_OPERAND (g_elt
, 1), child
->element
))
181 child
= child
->sibling
;
187 /* Random access to the child of a parent is performed by hashing.
188 This prevents quadratic behavior, and allows SRA to function
189 reasonably on larger records. */
190 static htab_t sra_map
;
192 /* All structures are allocated out of the following obstack. */
193 static struct obstack sra_obstack
;
195 /* Debugging functions. */
196 static void dump_sra_elt_name (FILE *, struct sra_elt
*);
197 extern void debug_sra_elt_name (struct sra_elt
*);
199 /* Forward declarations. */
200 static tree
generate_element_ref (struct sra_elt
*);
202 /* Return true if DECL is an SRA candidate. */
205 is_sra_candidate_decl (tree decl
)
207 return DECL_P (decl
) && bitmap_bit_p (sra_candidates
, DECL_UID (decl
));
210 /* Return true if TYPE is a scalar type. */
213 is_sra_scalar_type (tree type
)
215 enum tree_code code
= TREE_CODE (type
);
216 return (code
== INTEGER_TYPE
|| code
== REAL_TYPE
|| code
== VECTOR_TYPE
217 || code
== ENUMERAL_TYPE
|| code
== BOOLEAN_TYPE
218 || code
== POINTER_TYPE
|| code
== OFFSET_TYPE
219 || code
== REFERENCE_TYPE
);
222 /* Return true if TYPE can be decomposed into a set of independent variables.
224 Note that this doesn't imply that all elements of TYPE can be
225 instantiated, just that if we decide to break up the type into
226 separate pieces that it can be done. */
229 sra_type_can_be_decomposed_p (tree type
)
231 unsigned int cache
= TYPE_UID (TYPE_MAIN_VARIANT (type
)) * 2;
234 /* Avoid searching the same type twice. */
235 if (bitmap_bit_p (sra_type_decomp_cache
, cache
+0))
237 if (bitmap_bit_p (sra_type_decomp_cache
, cache
+1))
240 /* The type must have a definite nonzero size. */
241 if (TYPE_SIZE (type
) == NULL
|| TREE_CODE (TYPE_SIZE (type
)) != INTEGER_CST
242 || integer_zerop (TYPE_SIZE (type
)))
245 /* The type must be a non-union aggregate. */
246 switch (TREE_CODE (type
))
250 bool saw_one_field
= false;
252 for (t
= TYPE_FIELDS (type
); t
; t
= TREE_CHAIN (t
))
253 if (TREE_CODE (t
) == FIELD_DECL
)
255 /* Reject incorrectly represented bit fields. */
256 if (DECL_BIT_FIELD (t
)
257 && (tree_low_cst (DECL_SIZE (t
), 1)
258 != TYPE_PRECISION (TREE_TYPE (t
))))
261 saw_one_field
= true;
264 /* Record types must have at least one field. */
271 /* Array types must have a fixed lower and upper bound. */
272 t
= TYPE_DOMAIN (type
);
275 if (TYPE_MIN_VALUE (t
) == NULL
|| !TREE_CONSTANT (TYPE_MIN_VALUE (t
)))
277 if (TYPE_MAX_VALUE (t
) == NULL
|| !TREE_CONSTANT (TYPE_MAX_VALUE (t
)))
288 bitmap_set_bit (sra_type_decomp_cache
, cache
+0);
292 bitmap_set_bit (sra_type_decomp_cache
, cache
+1);
296 /* Return true if DECL can be decomposed into a set of independent
297 (though not necessarily scalar) variables. */
300 decl_can_be_decomposed_p (tree var
)
302 /* Early out for scalars. */
303 if (is_sra_scalar_type (TREE_TYPE (var
)))
306 /* The variable must not be aliased. */
307 if (!is_gimple_non_addressable (var
))
309 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
311 fprintf (dump_file
, "Cannot scalarize variable ");
312 print_generic_expr (dump_file
, var
, dump_flags
);
313 fprintf (dump_file
, " because it must live in memory\n");
318 /* The variable must not be volatile. */
319 if (TREE_THIS_VOLATILE (var
))
321 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
323 fprintf (dump_file
, "Cannot scalarize variable ");
324 print_generic_expr (dump_file
, var
, dump_flags
);
325 fprintf (dump_file
, " because it is declared volatile\n");
330 /* We must be able to decompose the variable's type. */
331 if (!sra_type_can_be_decomposed_p (TREE_TYPE (var
)))
333 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
335 fprintf (dump_file
, "Cannot scalarize variable ");
336 print_generic_expr (dump_file
, var
, dump_flags
);
337 fprintf (dump_file
, " because its type cannot be decomposed\n");
345 /* Return true if TYPE can be *completely* decomposed into scalars. */
348 type_can_instantiate_all_elements (tree type
)
350 if (is_sra_scalar_type (type
))
352 if (!sra_type_can_be_decomposed_p (type
))
355 switch (TREE_CODE (type
))
359 unsigned int cache
= TYPE_UID (TYPE_MAIN_VARIANT (type
)) * 2;
362 if (bitmap_bit_p (sra_type_inst_cache
, cache
+0))
364 if (bitmap_bit_p (sra_type_inst_cache
, cache
+1))
367 for (f
= TYPE_FIELDS (type
); f
; f
= TREE_CHAIN (f
))
368 if (TREE_CODE (f
) == FIELD_DECL
)
370 if (!type_can_instantiate_all_elements (TREE_TYPE (f
)))
372 bitmap_set_bit (sra_type_inst_cache
, cache
+1);
377 bitmap_set_bit (sra_type_inst_cache
, cache
+0);
382 return type_can_instantiate_all_elements (TREE_TYPE (type
));
392 /* Test whether ELT or some sub-element cannot be scalarized. */
395 can_completely_scalarize_p (struct sra_elt
*elt
)
399 if (elt
->cannot_scalarize
)
402 for (c
= elt
->children
; c
; c
= c
->sibling
)
403 if (!can_completely_scalarize_p (c
))
406 for (c
= elt
->groups
; c
; c
= c
->sibling
)
407 if (!can_completely_scalarize_p (c
))
414 /* A simplified tree hashing algorithm that only handles the types of
415 trees we expect to find in sra_elt->element. */
418 sra_hash_tree (tree t
)
422 switch (TREE_CODE (t
))
431 h
= TREE_INT_CST_LOW (t
) ^ TREE_INT_CST_HIGH (t
);
435 h
= iterative_hash_expr (TREE_OPERAND (t
, 0), 0);
436 h
= iterative_hash_expr (TREE_OPERAND (t
, 1), h
);
440 /* We can have types that are compatible, but have different member
441 lists, so we can't hash fields by ID. Use offsets instead. */
442 h
= iterative_hash_expr (DECL_FIELD_OFFSET (t
), 0);
443 h
= iterative_hash_expr (DECL_FIELD_BIT_OFFSET (t
), h
);
453 /* Hash function for type SRA_PAIR. */
456 sra_elt_hash (const void *x
)
458 const struct sra_elt
*e
= x
;
459 const struct sra_elt
*p
;
462 h
= sra_hash_tree (e
->element
);
464 /* Take into account everything back up the chain. Given that chain
465 lengths are rarely very long, this should be acceptable. If we
466 truly identify this as a performance problem, it should work to
467 hash the pointer value "e->parent". */
468 for (p
= e
->parent
; p
; p
= p
->parent
)
469 h
= (h
* 65521) ^ sra_hash_tree (p
->element
);
474 /* Equality function for type SRA_PAIR. */
477 sra_elt_eq (const void *x
, const void *y
)
479 const struct sra_elt
*a
= x
;
480 const struct sra_elt
*b
= y
;
483 if (a
->parent
!= b
->parent
)
491 if (TREE_CODE (ae
) != TREE_CODE (be
))
494 switch (TREE_CODE (ae
))
499 /* These are all pointer unique. */
503 /* Integers are not pointer unique, so compare their values. */
504 return tree_int_cst_equal (ae
, be
);
508 tree_int_cst_equal (TREE_OPERAND (ae
, 0), TREE_OPERAND (be
, 0))
509 && tree_int_cst_equal (TREE_OPERAND (ae
, 1), TREE_OPERAND (be
, 1));
512 /* Fields are unique within a record, but not between
513 compatible records. */
514 if (DECL_FIELD_CONTEXT (ae
) == DECL_FIELD_CONTEXT (be
))
516 return fields_compatible_p (ae
, be
);
523 /* Create or return the SRA_ELT structure for CHILD in PARENT. PARENT
524 may be null, in which case CHILD must be a DECL. */
526 static struct sra_elt
*
527 lookup_element (struct sra_elt
*parent
, tree child
, tree type
,
528 enum insert_option insert
)
530 struct sra_elt dummy
;
531 struct sra_elt
**slot
;
535 dummy
.parent
= parent
->is_group
? parent
->parent
: parent
;
538 dummy
.element
= child
;
540 slot
= (struct sra_elt
**) htab_find_slot (sra_map
, &dummy
, insert
);
541 if (!slot
&& insert
== NO_INSERT
)
545 if (!elt
&& insert
== INSERT
)
547 *slot
= elt
= obstack_alloc (&sra_obstack
, sizeof (*elt
));
548 memset (elt
, 0, sizeof (*elt
));
550 elt
->parent
= parent
;
551 elt
->element
= child
;
553 elt
->is_scalar
= is_sra_scalar_type (type
);
557 if (IS_ELEMENT_FOR_GROUP (elt
->element
))
559 elt
->is_group
= true;
560 elt
->sibling
= parent
->groups
;
561 parent
->groups
= elt
;
565 elt
->sibling
= parent
->children
;
566 parent
->children
= elt
;
570 /* If this is a parameter, then if we want to scalarize, we have
571 one copy from the true function parameter. Count it now. */
572 if (TREE_CODE (child
) == PARM_DECL
)
575 bitmap_set_bit (needs_copy_in
, DECL_UID (child
));
582 /* Create or return the SRA_ELT structure for EXPR if the expression
583 refers to a scalarizable variable. */
585 static struct sra_elt
*
586 maybe_lookup_element_for_expr (tree expr
)
591 switch (TREE_CODE (expr
))
596 if (is_sra_candidate_decl (expr
))
597 return lookup_element (NULL
, expr
, TREE_TYPE (expr
), INSERT
);
601 /* We can't scalarize variable array indices. */
602 if (in_array_bounds_p (expr
))
603 child
= TREE_OPERAND (expr
, 1);
608 case ARRAY_RANGE_REF
:
609 /* We can't scalarize variable array indices. */
610 if (range_in_array_bounds_p (expr
))
612 tree domain
= TYPE_DOMAIN (TREE_TYPE (expr
));
613 child
= build2 (RANGE_EXPR
, integer_type_node
,
614 TYPE_MIN_VALUE (domain
), TYPE_MAX_VALUE (domain
));
621 /* Don't look through unions. */
622 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (expr
, 0))) != RECORD_TYPE
)
624 child
= TREE_OPERAND (expr
, 1);
628 child
= integer_zero_node
;
631 child
= integer_one_node
;
638 elt
= maybe_lookup_element_for_expr (TREE_OPERAND (expr
, 0));
640 return lookup_element (elt
, child
, TREE_TYPE (expr
), INSERT
);
645 /* Functions to walk just enough of the tree to see all scalarizable
646 references, and categorize them. */
648 /* A set of callbacks for phases 2 and 4. They'll be invoked for the
649 various kinds of references seen. In all cases, *BSI is an iterator
650 pointing to the statement being processed. */
653 /* Invoked when ELT is required as a unit. Note that ELT might refer to
654 a leaf node, in which case this is a simple scalar reference. *EXPR_P
655 points to the location of the expression. IS_OUTPUT is true if this
656 is a left-hand-side reference. USE_ALL is true if we saw something we
657 couldn't quite identify and had to force the use of the entire object. */
658 void (*use
) (struct sra_elt
*elt
, tree
*expr_p
,
659 block_stmt_iterator
*bsi
, bool is_output
, bool use_all
);
661 /* Invoked when we have a copy between two scalarizable references. */
662 void (*copy
) (struct sra_elt
*lhs_elt
, struct sra_elt
*rhs_elt
,
663 block_stmt_iterator
*bsi
);
665 /* Invoked when ELT is initialized from a constant. VALUE may be NULL,
666 in which case it should be treated as an empty CONSTRUCTOR. */
667 void (*init
) (struct sra_elt
*elt
, tree value
, block_stmt_iterator
*bsi
);
669 /* Invoked when we have a copy between one scalarizable reference ELT
670 and one non-scalarizable reference OTHER. IS_OUTPUT is true if ELT
671 is on the left-hand side. */
672 void (*ldst
) (struct sra_elt
*elt
, tree other
,
673 block_stmt_iterator
*bsi
, bool is_output
);
675 /* True during phase 2, false during phase 4. */
676 /* ??? This is a hack. */
680 #ifdef ENABLE_CHECKING
681 /* Invoked via walk_tree, if *TP contains a candidate decl, return it. */
684 sra_find_candidate_decl (tree
*tp
, int *walk_subtrees
,
685 void *data ATTRIBUTE_UNUSED
)
688 enum tree_code code
= TREE_CODE (t
);
690 if (code
== VAR_DECL
|| code
== PARM_DECL
|| code
== RESULT_DECL
)
693 if (is_sra_candidate_decl (t
))
703 /* Walk most expressions looking for a scalarizable aggregate.
704 If we find one, invoke FNS->USE. */
707 sra_walk_expr (tree
*expr_p
, block_stmt_iterator
*bsi
, bool is_output
,
708 const struct sra_walk_fns
*fns
)
712 bool disable_scalarization
= false;
713 bool use_all_p
= false;
715 /* We're looking to collect a reference expression between EXPR and INNER,
716 such that INNER is a scalarizable decl and all other nodes through EXPR
717 are references that we can scalarize. If we come across something that
718 we can't scalarize, we reset EXPR. This has the effect of making it
719 appear that we're referring to the larger expression as a whole. */
722 switch (TREE_CODE (inner
))
727 /* If there is a scalarizable decl at the bottom, then process it. */
728 if (is_sra_candidate_decl (inner
))
730 struct sra_elt
*elt
= maybe_lookup_element_for_expr (expr
);
731 if (disable_scalarization
)
732 elt
->cannot_scalarize
= true;
734 fns
->use (elt
, expr_p
, bsi
, is_output
, use_all_p
);
739 /* Non-constant index means any member may be accessed. Prevent the
740 expression from being scalarized. If we were to treat this as a
741 reference to the whole array, we can wind up with a single dynamic
742 index reference inside a loop being overridden by several constant
743 index references during loop setup. It's possible that this could
744 be avoided by using dynamic usage counts based on BB trip counts
745 (based on loop analysis or profiling), but that hardly seems worth
747 /* ??? Hack. Figure out how to push this into the scan routines
748 without duplicating too much code. */
749 if (!in_array_bounds_p (inner
))
751 disable_scalarization
= true;
754 /* ??? Are we assured that non-constant bounds and stride will have
755 the same value everywhere? I don't think Fortran will... */
756 if (TREE_OPERAND (inner
, 2) || TREE_OPERAND (inner
, 3))
758 inner
= TREE_OPERAND (inner
, 0);
761 case ARRAY_RANGE_REF
:
762 if (!range_in_array_bounds_p (inner
))
764 disable_scalarization
= true;
767 /* ??? See above non-constant bounds and stride . */
768 if (TREE_OPERAND (inner
, 2) || TREE_OPERAND (inner
, 3))
770 inner
= TREE_OPERAND (inner
, 0);
774 /* A reference to a union member constitutes a reference to the
776 if (TREE_CODE (TREE_TYPE (TREE_OPERAND (inner
, 0))) != RECORD_TYPE
)
778 /* ??? See above re non-constant stride. */
779 if (TREE_OPERAND (inner
, 2))
781 inner
= TREE_OPERAND (inner
, 0);
786 inner
= TREE_OPERAND (inner
, 0);
790 /* A bit field reference (access to *multiple* fields simultaneously)
791 is not currently scalarized. Consider this an access to the
792 complete outer element, to which walk_tree will bring us next. */
795 case VIEW_CONVERT_EXPR
:
797 /* Similarly, a view/nop explicitly wants to look at an object in a
798 type other than the one we've scalarized. */
802 /* This is a transparent wrapper. The entire inner expression really
807 expr_p
= &TREE_OPERAND (inner
, 0);
808 inner
= expr
= *expr_p
;
813 #ifdef ENABLE_CHECKING
814 /* Validate that we're not missing any references. */
815 gcc_assert (!walk_tree (&inner
, sra_find_candidate_decl
, NULL
, NULL
));
821 /* Walk a TREE_LIST of values looking for scalarizable aggregates.
822 If we find one, invoke FNS->USE. */
825 sra_walk_tree_list (tree list
, block_stmt_iterator
*bsi
, bool is_output
,
826 const struct sra_walk_fns
*fns
)
829 for (op
= list
; op
; op
= TREE_CHAIN (op
))
830 sra_walk_expr (&TREE_VALUE (op
), bsi
, is_output
, fns
);
833 /* Walk the arguments of a CALL_EXPR looking for scalarizable aggregates.
834 If we find one, invoke FNS->USE. */
837 sra_walk_call_expr (tree expr
, block_stmt_iterator
*bsi
,
838 const struct sra_walk_fns
*fns
)
840 sra_walk_tree_list (TREE_OPERAND (expr
, 1), bsi
, false, fns
);
843 /* Walk the inputs and outputs of an ASM_EXPR looking for scalarizable
844 aggregates. If we find one, invoke FNS->USE. */
847 sra_walk_asm_expr (tree expr
, block_stmt_iterator
*bsi
,
848 const struct sra_walk_fns
*fns
)
850 sra_walk_tree_list (ASM_INPUTS (expr
), bsi
, false, fns
);
851 sra_walk_tree_list (ASM_OUTPUTS (expr
), bsi
, true, fns
);
854 /* Walk a MODIFY_EXPR and categorize the assignment appropriately. */
857 sra_walk_modify_expr (tree expr
, block_stmt_iterator
*bsi
,
858 const struct sra_walk_fns
*fns
)
860 struct sra_elt
*lhs_elt
, *rhs_elt
;
863 lhs
= TREE_OPERAND (expr
, 0);
864 rhs
= TREE_OPERAND (expr
, 1);
865 lhs_elt
= maybe_lookup_element_for_expr (lhs
);
866 rhs_elt
= maybe_lookup_element_for_expr (rhs
);
868 /* If both sides are scalarizable, this is a COPY operation. */
869 if (lhs_elt
&& rhs_elt
)
871 fns
->copy (lhs_elt
, rhs_elt
, bsi
);
875 /* If the RHS is scalarizable, handle it. There are only two cases. */
878 if (!rhs_elt
->is_scalar
)
879 fns
->ldst (rhs_elt
, lhs
, bsi
, false);
881 fns
->use (rhs_elt
, &TREE_OPERAND (expr
, 1), bsi
, false, false);
884 /* If it isn't scalarizable, there may be scalarizable variables within, so
885 check for a call or else walk the RHS to see if we need to do any
886 copy-in operations. We need to do it before the LHS is scalarized so
887 that the statements get inserted in the proper place, before any
888 copy-out operations. */
891 tree call
= get_call_expr_in (rhs
);
893 sra_walk_call_expr (call
, bsi
, fns
);
895 sra_walk_expr (&TREE_OPERAND (expr
, 1), bsi
, false, fns
);
898 /* Likewise, handle the LHS being scalarizable. We have cases similar
899 to those above, but also want to handle RHS being constant. */
902 /* If this is an assignment from a constant, or constructor, then
903 we have access to all of the elements individually. Invoke INIT. */
904 if (TREE_CODE (rhs
) == COMPLEX_EXPR
905 || TREE_CODE (rhs
) == COMPLEX_CST
906 || TREE_CODE (rhs
) == CONSTRUCTOR
)
907 fns
->init (lhs_elt
, rhs
, bsi
);
909 /* If this is an assignment from read-only memory, treat this as if
910 we'd been passed the constructor directly. Invoke INIT. */
911 else if (TREE_CODE (rhs
) == VAR_DECL
913 && TREE_READONLY (rhs
)
914 && targetm
.binds_local_p (rhs
))
915 fns
->init (lhs_elt
, DECL_INITIAL (rhs
), bsi
);
917 /* If this is a copy from a non-scalarizable lvalue, invoke LDST.
918 The lvalue requirement prevents us from trying to directly scalarize
919 the result of a function call. Which would result in trying to call
920 the function multiple times, and other evil things. */
921 else if (!lhs_elt
->is_scalar
&& is_gimple_addressable (rhs
))
922 fns
->ldst (lhs_elt
, rhs
, bsi
, true);
924 /* Otherwise we're being used in some context that requires the
925 aggregate to be seen as a whole. Invoke USE. */
927 fns
->use (lhs_elt
, &TREE_OPERAND (expr
, 0), bsi
, true, false);
930 /* Similarly to above, LHS_ELT being null only means that the LHS as a
931 whole is not a scalarizable reference. There may be occurrences of
932 scalarizable variables within, which implies a USE. */
934 sra_walk_expr (&TREE_OPERAND (expr
, 0), bsi
, true, fns
);
937 /* Entry point to the walk functions. Search the entire function,
938 invoking the callbacks in FNS on each of the references to
939 scalarizable variables. */
942 sra_walk_function (const struct sra_walk_fns
*fns
)
945 block_stmt_iterator si
, ni
;
947 /* ??? Phase 4 could derive some benefit to walking the function in
948 dominator tree order. */
951 for (si
= bsi_start (bb
); !bsi_end_p (si
); si
= ni
)
956 stmt
= bsi_stmt (si
);
957 ann
= stmt_ann (stmt
);
962 /* If the statement has no virtual operands, then it doesn't
963 make any structure references that we care about. */
964 if (ZERO_SSA_OPERANDS (stmt
, (SSA_OP_VIRTUAL_DEFS
| SSA_OP_VUSE
)))
967 switch (TREE_CODE (stmt
))
970 /* If we have "return <retval>" then the return value is
971 already exposed for our pleasure. Walk it as a USE to
972 force all the components back in place for the return.
974 If we have an embedded assignment, then <retval> is of
975 a type that gets returned in registers in this ABI, and
976 we do not wish to extend their lifetimes. Treat this
977 as a USE of the variable on the RHS of this assignment. */
979 t
= TREE_OPERAND (stmt
, 0);
980 if (TREE_CODE (t
) == MODIFY_EXPR
)
981 sra_walk_expr (&TREE_OPERAND (t
, 1), &si
, false, fns
);
983 sra_walk_expr (&TREE_OPERAND (stmt
, 0), &si
, false, fns
);
987 sra_walk_modify_expr (stmt
, &si
, fns
);
990 sra_walk_call_expr (stmt
, &si
, fns
);
993 sra_walk_asm_expr (stmt
, &si
, fns
);
1002 /* Phase One: Scan all referenced variables in the program looking for
1003 structures that could be decomposed. */
1006 find_candidates_for_sra (void)
1008 bool any_set
= false;
1010 referenced_var_iterator rvi
;
1012 FOR_EACH_REFERENCED_VAR (var
, rvi
)
1014 if (decl_can_be_decomposed_p (var
))
1016 bitmap_set_bit (sra_candidates
, DECL_UID (var
));
1025 /* Phase Two: Scan all references to scalarizable variables. Count the
1026 number of times they are used or copied respectively. */
1028 /* Callbacks to fill in SRA_WALK_FNS. Everything but USE is
1029 considered a copy, because we can decompose the reference such that
1030 the sub-elements needn't be contiguous. */
1033 scan_use (struct sra_elt
*elt
, tree
*expr_p ATTRIBUTE_UNUSED
,
1034 block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
,
1035 bool is_output ATTRIBUTE_UNUSED
, bool use_all ATTRIBUTE_UNUSED
)
1041 scan_copy (struct sra_elt
*lhs_elt
, struct sra_elt
*rhs_elt
,
1042 block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
)
1044 lhs_elt
->n_copies
+= 1;
1045 rhs_elt
->n_copies
+= 1;
1049 scan_init (struct sra_elt
*lhs_elt
, tree rhs ATTRIBUTE_UNUSED
,
1050 block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
)
1052 lhs_elt
->n_copies
+= 1;
1056 scan_ldst (struct sra_elt
*elt
, tree other ATTRIBUTE_UNUSED
,
1057 block_stmt_iterator
*bsi ATTRIBUTE_UNUSED
,
1058 bool is_output ATTRIBUTE_UNUSED
)
1063 /* Dump the values we collected during the scanning phase. */
1066 scan_dump (struct sra_elt
*elt
)
1070 dump_sra_elt_name (dump_file
, elt
);
1071 fprintf (dump_file
, ": n_uses=%u n_copies=%u\n", elt
->n_uses
, elt
->n_copies
);
1073 for (c
= elt
->children
; c
; c
= c
->sibling
)
1076 for (c
= elt
->groups
; c
; c
= c
->sibling
)
1080 /* Entry point to phase 2. Scan the entire function, building up
1081 scalarization data structures, recording copies and uses. */
1084 scan_function (void)
1086 static const struct sra_walk_fns fns
= {
1087 scan_use
, scan_copy
, scan_init
, scan_ldst
, true
1091 sra_walk_function (&fns
);
1093 if (dump_file
&& (dump_flags
& TDF_DETAILS
))
1097 fputs ("\nScan results:\n", dump_file
);
1098 EXECUTE_IF_SET_IN_BITMAP (sra_candidates
, 0, i
, bi
)
1100 tree var
= referenced_var (i
);
1101 struct sra_elt
*elt
= lookup_element (NULL
, var
, NULL
, NO_INSERT
);
1105 fputc ('\n', dump_file
);
1109 /* Phase Three: Make decisions about which variables to scalarize, if any.
1110 All elements to be scalarized have replacement variables made for them. */
1112 /* A subroutine of build_element_name. Recursively build the element
1113 name on the obstack. */
1116 build_element_name_1 (struct sra_elt
*elt
)
1123 build_element_name_1 (elt
->parent
);
1124 obstack_1grow (&sra_obstack
, '$');
1126 if (TREE_CODE (elt
->parent
->type
) == COMPLEX_TYPE
)
1128 if (elt
->element
== integer_zero_node
)
1129 obstack_grow (&sra_obstack
, "real", 4);
1131 obstack_grow (&sra_obstack
, "imag", 4);
1137 if (TREE_CODE (t
) == INTEGER_CST
)
1139 /* ??? Eh. Don't bother doing double-wide printing. */
1140 sprintf (buffer
, HOST_WIDE_INT_PRINT_DEC
, TREE_INT_CST_LOW (t
));
1141 obstack_grow (&sra_obstack
, buffer
, strlen (buffer
));
1145 tree name
= DECL_NAME (t
);
1147 obstack_grow (&sra_obstack
, IDENTIFIER_POINTER (name
),
1148 IDENTIFIER_LENGTH (name
));
1151 sprintf (buffer
, "D%u", DECL_UID (t
));
1152 obstack_grow (&sra_obstack
, buffer
, strlen (buffer
));
1157 /* Construct a pretty variable name for an element's replacement variable.
1158 The name is built on the obstack. */
1161 build_element_name (struct sra_elt
*elt
)
1163 build_element_name_1 (elt
);
1164 obstack_1grow (&sra_obstack
, '\0');
1165 return XOBFINISH (&sra_obstack
, char *);
1168 /* Instantiate an element as an independent variable. */
1171 instantiate_element (struct sra_elt
*elt
)
1173 struct sra_elt
*base_elt
;
1176 for (base_elt
= elt
; base_elt
->parent
; base_elt
= base_elt
->parent
)
1178 base
= base_elt
->element
;
1180 elt
->replacement
= var
= make_rename_temp (elt
->type
, "SR");
1181 DECL_SOURCE_LOCATION (var
) = DECL_SOURCE_LOCATION (base
);
1182 DECL_ARTIFICIAL (var
) = 1;
1184 if (TREE_THIS_VOLATILE (elt
->type
))
1186 TREE_THIS_VOLATILE (var
) = 1;
1187 TREE_SIDE_EFFECTS (var
) = 1;
1190 if (DECL_NAME (base
) && !DECL_IGNORED_P (base
))
1192 char *pretty_name
= build_element_name (elt
);
1193 DECL_NAME (var
) = get_identifier (pretty_name
);
1194 obstack_free (&sra_obstack
, pretty_name
);
1196 SET_DECL_DEBUG_EXPR (var
, generate_element_ref (elt
));
1197 DECL_DEBUG_EXPR_IS_FROM (var
) = 1;
1199 DECL_IGNORED_P (var
) = 0;
1200 TREE_NO_WARNING (var
) = TREE_NO_WARNING (base
);
1204 DECL_IGNORED_P (var
) = 1;
1205 /* ??? We can't generate any warning that would be meaningful. */
1206 TREE_NO_WARNING (var
) = 1;
1211 fputs (" ", dump_file
);
1212 dump_sra_elt_name (dump_file
, elt
);
1213 fputs (" -> ", dump_file
);
1214 print_generic_expr (dump_file
, var
, dump_flags
);
1215 fputc ('\n', dump_file
);
1219 /* Make one pass across an element tree deciding whether or not it's
1220 profitable to instantiate individual leaf scalars.
1222 PARENT_USES and PARENT_COPIES are the sum of the N_USES and N_COPIES
1223 fields all the way up the tree. */
1226 decide_instantiation_1 (struct sra_elt
*elt
, unsigned int parent_uses
,
1227 unsigned int parent_copies
)
1229 if (dump_file
&& !elt
->parent
)
1231 fputs ("Initial instantiation for ", dump_file
);
1232 dump_sra_elt_name (dump_file
, elt
);
1233 fputc ('\n', dump_file
);
1236 if (elt
->cannot_scalarize
)
1241 /* The decision is simple: instantiate if we're used more frequently
1242 than the parent needs to be seen as a complete unit. */
1243 if (elt
->n_uses
+ elt
->n_copies
+ parent_copies
> parent_uses
)
1244 instantiate_element (elt
);
1248 struct sra_elt
*c
, *group
;
1249 unsigned int this_uses
= elt
->n_uses
+ parent_uses
;
1250 unsigned int this_copies
= elt
->n_copies
+ parent_copies
;
1252 /* Consider groups of sub-elements as weighing in favour of
1253 instantiation whatever their size. */
1254 for (group
= elt
->groups
; group
; group
= group
->sibling
)
1255 FOR_EACH_ACTUAL_CHILD (c
, group
)
1257 c
->n_uses
+= group
->n_uses
;
1258 c
->n_copies
+= group
->n_copies
;
1261 for (c
= elt
->children
; c
; c
= c
->sibling
)
1262 decide_instantiation_1 (c
, this_uses
, this_copies
);
1266 /* Compute the size and number of all instantiated elements below ELT.
1267 We will only care about this if the size of the complete structure
1268 fits in a HOST_WIDE_INT, so we don't have to worry about overflow. */
1271 sum_instantiated_sizes (struct sra_elt
*elt
, unsigned HOST_WIDE_INT
*sizep
)
1273 if (elt
->replacement
)
1275 *sizep
+= TREE_INT_CST_LOW (TYPE_SIZE_UNIT (elt
->type
));
1281 unsigned int count
= 0;
1283 for (c
= elt
->children
; c
; c
= c
->sibling
)
1284 count
+= sum_instantiated_sizes (c
, sizep
);
1290 /* Instantiate fields in ELT->TYPE that are not currently present as
1293 static void instantiate_missing_elements (struct sra_elt
*elt
);
1296 instantiate_missing_elements_1 (struct sra_elt
*elt
, tree child
, tree type
)
1298 struct sra_elt
*sub
= lookup_element (elt
, child
, type
, INSERT
);
1301 if (sub
->replacement
== NULL
)
1302 instantiate_element (sub
);
1305 instantiate_missing_elements (sub
);
1309 instantiate_missing_elements (struct sra_elt
*elt
)
1311 tree type
= elt
->type
;
1313 switch (TREE_CODE (type
))
1318 for (f
= TYPE_FIELDS (type
); f
; f
= TREE_CHAIN (f
))
1319 if (TREE_CODE (f
) == FIELD_DECL
)
1320 instantiate_missing_elements_1 (elt
, f
, TREE_TYPE (f
));
1326 tree i
, max
, subtype
;
1328 i
= TYPE_MIN_VALUE (TYPE_DOMAIN (type
));
1329 max
= TYPE_MAX_VALUE (TYPE_DOMAIN (type
));
1330 subtype
= TREE_TYPE (type
);
1334 instantiate_missing_elements_1 (elt
, i
, subtype
);
1335 if (tree_int_cst_equal (i
, max
))
1337 i
= int_const_binop (PLUS_EXPR
, i
, integer_one_node
, true);
1344 type
= TREE_TYPE (type
);
1345 instantiate_missing_elements_1 (elt
, integer_zero_node
, type
);
1346 instantiate_missing_elements_1 (elt
, integer_one_node
, type
);
1354 /* Return true if there is only one non aggregate field in the record, TYPE.
1355 Return false otherwise. */
1358 single_scalar_field_in_record_p (tree type
)
1362 if (TREE_CODE (type
) != RECORD_TYPE
)
1365 for (field
= TYPE_FIELDS (type
); field
; field
= TREE_CHAIN (field
))
1366 if (TREE_CODE (field
) == FIELD_DECL
)
1370 if (num_fields
== 2)
1373 if (AGGREGATE_TYPE_P (TREE_TYPE (field
)))
1380 /* Make one pass across an element tree deciding whether to perform block
1381 or element copies. If we decide on element copies, instantiate all
1382 elements. Return true if there are any instantiated sub-elements. */
1385 decide_block_copy (struct sra_elt
*elt
)
1390 /* We shouldn't be invoked on groups of sub-elements as they must
1391 behave like their parent as far as block copy is concerned. */
1392 gcc_assert (!elt
->is_group
);
1394 /* If scalarization is disabled, respect it. */
1395 if (elt
->cannot_scalarize
)
1397 elt
->use_block_copy
= 1;
1401 fputs ("Scalarization disabled for ", dump_file
);
1402 dump_sra_elt_name (dump_file
, elt
);
1403 fputc ('\n', dump_file
);
1406 /* Disable scalarization of sub-elements */
1407 for (c
= elt
->children
; c
; c
= c
->sibling
)
1409 c
->cannot_scalarize
= 1;
1410 decide_block_copy (c
);
1413 /* Groups behave like their parent. */
1414 for (c
= elt
->groups
; c
; c
= c
->sibling
)
1416 c
->cannot_scalarize
= 1;
1417 c
->use_block_copy
= 1;
1423 /* Don't decide if we've no uses. */
1424 if (elt
->n_uses
== 0 && elt
->n_copies
== 0)
1427 else if (!elt
->is_scalar
)
1429 tree size_tree
= TYPE_SIZE_UNIT (elt
->type
);
1430 bool use_block_copy
= true;
1432 /* Tradeoffs for COMPLEX types pretty much always make it better
1433 to go ahead and split the components. */
1434 if (TREE_CODE (elt
->type
) == COMPLEX_TYPE
)
1435 use_block_copy
= false;
1437 /* Don't bother trying to figure out the rest if the structure is
1438 so large we can't do easy arithmetic. This also forces block
1439 copies for variable sized structures. */
1440 else if (host_integerp (size_tree
, 1))
1442 unsigned HOST_WIDE_INT full_size
, inst_size
= 0;
1443 unsigned int max_size
, max_count
, inst_count
, full_count
;
1445 /* If the sra-max-structure-size parameter is 0, then the
1446 user has not overridden the parameter and we can choose a
1447 sensible default. */
1448 max_size
= SRA_MAX_STRUCTURE_SIZE
1449 ? SRA_MAX_STRUCTURE_SIZE
1450 : MOVE_RATIO
* UNITS_PER_WORD
;
1451 max_count
= SRA_MAX_STRUCTURE_COUNT
1452 ? SRA_MAX_STRUCTURE_COUNT
1455 full_size
= tree_low_cst (size_tree
, 1);
1456 full_count
= count_type_elements (elt
->type
, false);
1457 inst_count
= sum_instantiated_sizes (elt
, &inst_size
);
1459 /* If there is only one scalar field in the record, don't block copy. */
1460 if (single_scalar_field_in_record_p (elt
->type
))
1461 use_block_copy
= false;
1463 /* ??? What to do here. If there are two fields, and we've only
1464 instantiated one, then instantiating the other is clearly a win.
1465 If there are a large number of fields then the size of the copy
1466 is much more of a factor. */
1468 /* If the structure is small, and we've made copies, go ahead
1469 and instantiate, hoping that the copies will go away. */
1470 if (full_size
<= max_size
1471 && (full_count
- inst_count
) <= max_count
1472 && elt
->n_copies
> elt
->n_uses
)
1473 use_block_copy
= false;
1474 else if (inst_count
* 100 >= full_count
* SRA_FIELD_STRUCTURE_RATIO
1475 && inst_size
* 100 >= full_size
* SRA_FIELD_STRUCTURE_RATIO
)
1476 use_block_copy
= false;
1478 /* In order to avoid block copy, we have to be able to instantiate
1479 all elements of the type. See if this is possible. */
1481 && (!can_completely_scalarize_p (elt
)
1482 || !type_can_instantiate_all_elements (elt
->type
)))
1483 use_block_copy
= true;
1486 elt
->use_block_copy
= use_block_copy
;
1488 /* Groups behave like their parent. */
1489 for (c
= elt
->groups
; c
; c
= c
->sibling
)
1490 c
->use_block_copy
= use_block_copy
;
1494 fprintf (dump_file
, "Using %s for ",
1495 use_block_copy
? "block-copy" : "element-copy");
1496 dump_sra_elt_name (dump_file
, elt
);
1497 fputc ('\n', dump_file
);
1500 if (!use_block_copy
)
1502 instantiate_missing_elements (elt
);
1507 any_inst
= elt
->replacement
!= NULL
;
1509 for (c
= elt
->children
; c
; c
= c
->sibling
)
1510 any_inst
|= decide_block_copy (c
);
1515 /* Entry point to phase 3. Instantiate scalar replacement variables. */
1518 decide_instantiations (void)
1522 bitmap_head done_head
;
1525 /* We cannot clear bits from a bitmap we're iterating over,
1526 so save up all the bits to clear until the end. */
1527 bitmap_initialize (&done_head
, &bitmap_default_obstack
);
1528 cleared_any
= false;
1530 EXECUTE_IF_SET_IN_BITMAP (sra_candidates
, 0, i
, bi
)
1532 tree var
= referenced_var (i
);
1533 struct sra_elt
*elt
= lookup_element (NULL
, var
, NULL
, NO_INSERT
);
1536 decide_instantiation_1 (elt
, 0, 0);
1537 if (!decide_block_copy (elt
))
1542 bitmap_set_bit (&done_head
, i
);
1549 bitmap_and_compl_into (sra_candidates
, &done_head
);
1550 bitmap_and_compl_into (needs_copy_in
, &done_head
);
1552 bitmap_clear (&done_head
);
1554 if (!bitmap_empty_p (sra_candidates
))
1555 todoflags
|= TODO_update_smt_usage
;
1557 mark_set_for_renaming (sra_candidates
);
1560 fputc ('\n', dump_file
);
1564 /* Phase Four: Update the function to match the replacements created. */
1566 /* Mark all the variables in V_MAY_DEF or V_MUST_DEF operands for STMT for
1567 renaming. This becomes necessary when we modify all of a non-scalar. */
1570 mark_all_v_defs_1 (tree stmt
)
1575 update_stmt_if_modified (stmt
);
1577 FOR_EACH_SSA_TREE_OPERAND (sym
, stmt
, iter
, SSA_OP_ALL_VIRTUALS
)
1579 if (TREE_CODE (sym
) == SSA_NAME
)
1580 sym
= SSA_NAME_VAR (sym
);
1581 mark_sym_for_renaming (sym
);
1586 /* Mark all the variables in virtual operands in all the statements in
1587 LIST for renaming. */
1590 mark_all_v_defs (tree list
)
1592 if (TREE_CODE (list
) != STATEMENT_LIST
)
1593 mark_all_v_defs_1 (list
);
1596 tree_stmt_iterator i
;
1597 for (i
= tsi_start (list
); !tsi_end_p (i
); tsi_next (&i
))
1598 mark_all_v_defs_1 (tsi_stmt (i
));
1602 /* Mark every replacement under ELT with TREE_NO_WARNING. */
1605 mark_no_warning (struct sra_elt
*elt
)
1607 if (!elt
->all_no_warning
)
1609 if (elt
->replacement
)
1610 TREE_NO_WARNING (elt
->replacement
) = 1;
1614 FOR_EACH_ACTUAL_CHILD (c
, elt
)
1615 mark_no_warning (c
);
1617 elt
->all_no_warning
= true;
1621 /* Build a single level component reference to ELT rooted at BASE. */
1624 generate_one_element_ref (struct sra_elt
*elt
, tree base
)
1626 switch (TREE_CODE (TREE_TYPE (base
)))
1630 tree field
= elt
->element
;
1632 /* Watch out for compatible records with differing field lists. */
1633 if (DECL_FIELD_CONTEXT (field
) != TYPE_MAIN_VARIANT (TREE_TYPE (base
)))
1634 field
= find_compatible_field (TREE_TYPE (base
), field
);
1636 return build3 (COMPONENT_REF
, elt
->type
, base
, field
, NULL
);
1640 todoflags
|= TODO_update_smt_usage
;
1641 if (TREE_CODE (elt
->element
) == RANGE_EXPR
)
1642 return build4 (ARRAY_RANGE_REF
, elt
->type
, base
,
1643 TREE_OPERAND (elt
->element
, 0), NULL
, NULL
);
1645 return build4 (ARRAY_REF
, elt
->type
, base
, elt
->element
, NULL
, NULL
);
1648 if (elt
->element
== integer_zero_node
)
1649 return build1 (REALPART_EXPR
, elt
->type
, base
);
1651 return build1 (IMAGPART_EXPR
, elt
->type
, base
);
1658 /* Build a full component reference to ELT rooted at its native variable. */
1661 generate_element_ref (struct sra_elt
*elt
)
1664 return generate_one_element_ref (elt
, generate_element_ref (elt
->parent
));
1666 return elt
->element
;
1669 /* Generate a set of assignment statements in *LIST_P to copy all
1670 instantiated elements under ELT to or from the equivalent structure
1671 rooted at EXPR. COPY_OUT controls the direction of the copy, with
1672 true meaning to copy out of EXPR into ELT. */
1675 generate_copy_inout (struct sra_elt
*elt
, bool copy_out
, tree expr
,
1681 if (!copy_out
&& TREE_CODE (expr
) == SSA_NAME
1682 && TREE_CODE (TREE_TYPE (expr
)) == COMPLEX_TYPE
)
1686 c
= lookup_element (elt
, integer_zero_node
, NULL
, NO_INSERT
);
1688 c
= lookup_element (elt
, integer_one_node
, NULL
, NO_INSERT
);
1691 t
= build2 (COMPLEX_EXPR
, elt
->type
, r
, i
);
1692 t
= build2 (MODIFY_EXPR
, void_type_node
, expr
, t
);
1693 SSA_NAME_DEF_STMT (expr
) = t
;
1694 append_to_statement_list (t
, list_p
);
1696 else if (elt
->replacement
)
1699 t
= build2 (MODIFY_EXPR
, void_type_node
, elt
->replacement
, expr
);
1701 t
= build2 (MODIFY_EXPR
, void_type_node
, expr
, elt
->replacement
);
1702 append_to_statement_list (t
, list_p
);
1706 FOR_EACH_ACTUAL_CHILD (c
, elt
)
1708 t
= generate_one_element_ref (c
, unshare_expr (expr
));
1709 generate_copy_inout (c
, copy_out
, t
, list_p
);
1714 /* Generate a set of assignment statements in *LIST_P to copy all instantiated
1715 elements under SRC to their counterparts under DST. There must be a 1-1
1716 correspondence of instantiated elements. */
1719 generate_element_copy (struct sra_elt
*dst
, struct sra_elt
*src
, tree
*list_p
)
1721 struct sra_elt
*dc
, *sc
;
1723 FOR_EACH_ACTUAL_CHILD (dc
, dst
)
1725 sc
= lookup_element (src
, dc
->element
, NULL
, NO_INSERT
);
1727 generate_element_copy (dc
, sc
, list_p
);
1730 if (dst
->replacement
)
1734 gcc_assert (src
->replacement
);
1736 t
= build2 (MODIFY_EXPR
, void_type_node
, dst
->replacement
,
1738 append_to_statement_list (t
, list_p
);
1742 /* Generate a set of assignment statements in *LIST_P to zero all instantiated
1743 elements under ELT. In addition, do not assign to elements that have been
1744 marked VISITED but do reset the visited flag; this allows easy coordination
1745 with generate_element_init. */
1748 generate_element_zero (struct sra_elt
*elt
, tree
*list_p
)
1754 elt
->visited
= false;
1758 FOR_EACH_ACTUAL_CHILD (c
, elt
)
1759 generate_element_zero (c
, list_p
);
1761 if (elt
->replacement
)
1765 gcc_assert (elt
->is_scalar
);
1766 t
= fold_convert (elt
->type
, integer_zero_node
);
1768 t
= build2 (MODIFY_EXPR
, void_type_node
, elt
->replacement
, t
);
1769 append_to_statement_list (t
, list_p
);
1773 /* Generate an assignment VAR = INIT, where INIT may need gimplification.
1774 Add the result to *LIST_P. */
1777 generate_one_element_init (tree var
, tree init
, tree
*list_p
)
1779 /* The replacement can be almost arbitrarily complex. Gimplify. */
1780 tree stmt
= build2 (MODIFY_EXPR
, void_type_node
, var
, init
);
1781 gimplify_and_add (stmt
, list_p
);
1784 /* Generate a set of assignment statements in *LIST_P to set all instantiated
1785 elements under ELT with the contents of the initializer INIT. In addition,
1786 mark all assigned elements VISITED; this allows easy coordination with
1787 generate_element_zero. Return false if we found a case we couldn't
1791 generate_element_init_1 (struct sra_elt
*elt
, tree init
, tree
*list_p
)
1794 enum tree_code init_code
;
1795 struct sra_elt
*sub
;
1797 unsigned HOST_WIDE_INT idx
;
1798 tree value
, purpose
;
1800 /* We can be passed DECL_INITIAL of a static variable. It might have a
1801 conversion, which we strip off here. */
1802 STRIP_USELESS_TYPE_CONVERSION (init
);
1803 init_code
= TREE_CODE (init
);
1807 if (elt
->replacement
)
1809 generate_one_element_init (elt
->replacement
, init
, list_p
);
1810 elt
->visited
= true;
1819 FOR_EACH_ACTUAL_CHILD (sub
, elt
)
1821 if (sub
->element
== integer_zero_node
)
1822 t
= (init_code
== COMPLEX_EXPR
1823 ? TREE_OPERAND (init
, 0) : TREE_REALPART (init
));
1825 t
= (init_code
== COMPLEX_EXPR
1826 ? TREE_OPERAND (init
, 1) : TREE_IMAGPART (init
));
1827 result
&= generate_element_init_1 (sub
, t
, list_p
);
1832 FOR_EACH_CONSTRUCTOR_ELT (CONSTRUCTOR_ELTS (init
), idx
, purpose
, value
)
1834 if (TREE_CODE (purpose
) == RANGE_EXPR
)
1836 tree lower
= TREE_OPERAND (purpose
, 0);
1837 tree upper
= TREE_OPERAND (purpose
, 1);
1841 sub
= lookup_element (elt
, lower
, NULL
, NO_INSERT
);
1843 result
&= generate_element_init_1 (sub
, value
, list_p
);
1844 if (tree_int_cst_equal (lower
, upper
))
1846 lower
= int_const_binop (PLUS_EXPR
, lower
,
1847 integer_one_node
, true);
1852 sub
= lookup_element (elt
, purpose
, NULL
, NO_INSERT
);
1854 result
&= generate_element_init_1 (sub
, value
, list_p
);
1860 elt
->visited
= true;
1867 /* A wrapper function for generate_element_init_1 that handles cleanup after
1871 generate_element_init (struct sra_elt
*elt
, tree init
, tree
*list_p
)
1875 push_gimplify_context ();
1876 ret
= generate_element_init_1 (elt
, init
, list_p
);
1877 pop_gimplify_context (NULL
);
1879 /* The replacement can expose previously unreferenced variables. */
1882 tree_stmt_iterator i
;
1884 for (i
= tsi_start (*list_p
); !tsi_end_p (i
); tsi_next (&i
))
1885 find_new_referenced_vars (tsi_stmt_ptr (i
));
1891 /* Insert STMT on all the outgoing edges out of BB. Note that if BB
1892 has more than one edge, STMT will be replicated for each edge. Also,
1893 abnormal edges will be ignored. */
1896 insert_edge_copies (tree stmt
, basic_block bb
)
1903 FOR_EACH_EDGE (e
, ei
, bb
->succs
)
1905 /* We don't need to insert copies on abnormal edges. The
1906 value of the scalar replacement is not guaranteed to
1907 be valid through an abnormal edge. */
1908 if (!(e
->flags
& EDGE_ABNORMAL
))
1912 bsi_insert_on_edge (e
, stmt
);
1916 bsi_insert_on_edge (e
, unsave_expr_now (stmt
));
1921 /* Helper function to insert LIST before BSI, and set up line number info. */
1924 sra_insert_before (block_stmt_iterator
*bsi
, tree list
)
1926 tree stmt
= bsi_stmt (*bsi
);
1928 if (EXPR_HAS_LOCATION (stmt
))
1929 annotate_all_with_locus (&list
, EXPR_LOCATION (stmt
));
1930 bsi_insert_before (bsi
, list
, BSI_SAME_STMT
);
1933 /* Similarly, but insert after BSI. Handles insertion onto edges as well. */
1936 sra_insert_after (block_stmt_iterator
*bsi
, tree list
)
1938 tree stmt
= bsi_stmt (*bsi
);
1940 if (EXPR_HAS_LOCATION (stmt
))
1941 annotate_all_with_locus (&list
, EXPR_LOCATION (stmt
));
1943 if (stmt_ends_bb_p (stmt
))
1944 insert_edge_copies (list
, bsi
->bb
);
1946 bsi_insert_after (bsi
, list
, BSI_SAME_STMT
);
1949 /* Similarly, but replace the statement at BSI. */
1952 sra_replace (block_stmt_iterator
*bsi
, tree list
)
1954 sra_insert_before (bsi
, list
);
1955 bsi_remove (bsi
, false);
1956 if (bsi_end_p (*bsi
))
1957 *bsi
= bsi_last (bsi
->bb
);
1962 /* Scalarize a USE. To recap, this is either a simple reference to ELT,
1963 if elt is scalar, or some occurrence of ELT that requires a complete
1964 aggregate. IS_OUTPUT is true if ELT is being modified. */
1967 scalarize_use (struct sra_elt
*elt
, tree
*expr_p
, block_stmt_iterator
*bsi
,
1968 bool is_output
, bool use_all
)
1970 tree list
= NULL
, stmt
= bsi_stmt (*bsi
);
1972 if (elt
->replacement
)
1974 /* If we have a replacement, then updating the reference is as
1975 simple as modifying the existing statement in place. */
1977 mark_all_v_defs (stmt
);
1978 *expr_p
= elt
->replacement
;
1983 /* Otherwise we need some copies. If ELT is being read, then we want
1984 to store all (modified) sub-elements back into the structure before
1985 the reference takes place. If ELT is being written, then we want to
1986 load the changed values back into our shadow variables. */
1987 /* ??? We don't check modified for reads, we just always write all of
1988 the values. We should be able to record the SSA number of the VOP
1989 for which the values were last read. If that number matches the
1990 SSA number of the VOP in the current statement, then we needn't
1991 emit an assignment. This would also eliminate double writes when
1992 a structure is passed as more than one argument to a function call.
1993 This optimization would be most effective if sra_walk_function
1994 processed the blocks in dominator order. */
1996 generate_copy_inout (elt
, is_output
, generate_element_ref (elt
), &list
);
1999 mark_all_v_defs (list
);
2001 sra_insert_after (bsi
, list
);
2004 sra_insert_before (bsi
, list
);
2006 mark_no_warning (elt
);
2011 /* Scalarize a COPY. To recap, this is an assignment statement between
2012 two scalarizable references, LHS_ELT and RHS_ELT. */
2015 scalarize_copy (struct sra_elt
*lhs_elt
, struct sra_elt
*rhs_elt
,
2016 block_stmt_iterator
*bsi
)
2020 if (lhs_elt
->replacement
&& rhs_elt
->replacement
)
2022 /* If we have two scalar operands, modify the existing statement. */
2023 stmt
= bsi_stmt (*bsi
);
2025 /* See the commentary in sra_walk_function concerning
2026 RETURN_EXPR, and why we should never see one here. */
2027 gcc_assert (TREE_CODE (stmt
) == MODIFY_EXPR
);
2029 TREE_OPERAND (stmt
, 0) = lhs_elt
->replacement
;
2030 TREE_OPERAND (stmt
, 1) = rhs_elt
->replacement
;
2033 else if (lhs_elt
->use_block_copy
|| rhs_elt
->use_block_copy
)
2035 /* If either side requires a block copy, then sync the RHS back
2036 to the original structure, leave the original assignment
2037 statement (which will perform the block copy), then load the
2038 LHS values out of its now-updated original structure. */
2039 /* ??? Could perform a modified pair-wise element copy. That
2040 would at least allow those elements that are instantiated in
2041 both structures to be optimized well. */
2044 generate_copy_inout (rhs_elt
, false,
2045 generate_element_ref (rhs_elt
), &list
);
2048 mark_all_v_defs (list
);
2049 sra_insert_before (bsi
, list
);
2053 generate_copy_inout (lhs_elt
, true,
2054 generate_element_ref (lhs_elt
), &list
);
2057 mark_all_v_defs (list
);
2058 sra_insert_after (bsi
, list
);
2063 /* Otherwise both sides must be fully instantiated. In which
2064 case perform pair-wise element assignments and replace the
2065 original block copy statement. */
2067 stmt
= bsi_stmt (*bsi
);
2068 mark_all_v_defs (stmt
);
2071 generate_element_copy (lhs_elt
, rhs_elt
, &list
);
2073 mark_all_v_defs (list
);
2074 sra_replace (bsi
, list
);
2078 /* Scalarize an INIT. To recap, this is an assignment to a scalarizable
2079 reference from some form of constructor: CONSTRUCTOR, COMPLEX_CST or
2080 COMPLEX_EXPR. If RHS is NULL, it should be treated as an empty
2084 scalarize_init (struct sra_elt
*lhs_elt
, tree rhs
, block_stmt_iterator
*bsi
)
2089 /* Generate initialization statements for all members extant in the RHS. */
2092 /* Unshare the expression just in case this is from a decl's initial. */
2093 rhs
= unshare_expr (rhs
);
2094 result
= generate_element_init (lhs_elt
, rhs
, &list
);
2097 /* CONSTRUCTOR is defined such that any member not mentioned is assigned
2098 a zero value. Initialize the rest of the instantiated elements. */
2099 generate_element_zero (lhs_elt
, &list
);
2103 /* If we failed to convert the entire initializer, then we must
2104 leave the structure assignment in place and must load values
2105 from the structure into the slots for which we did not find
2106 constants. The easiest way to do this is to generate a complete
2107 copy-out, and then follow that with the constant assignments
2108 that we were able to build. DCE will clean things up. */
2110 generate_copy_inout (lhs_elt
, true, generate_element_ref (lhs_elt
),
2112 append_to_statement_list (list
, &list0
);
2116 if (lhs_elt
->use_block_copy
|| !result
)
2118 /* Since LHS is not fully instantiated, we must leave the structure
2119 assignment in place. Treating this case differently from a USE
2120 exposes constants to later optimizations. */
2123 mark_all_v_defs (list
);
2124 sra_insert_after (bsi
, list
);
2129 /* The LHS is fully instantiated. The list of initializations
2130 replaces the original structure assignment. */
2132 mark_all_v_defs (bsi_stmt (*bsi
));
2133 mark_all_v_defs (list
);
2134 sra_replace (bsi
, list
);
2138 /* A subroutine of scalarize_ldst called via walk_tree. Set TREE_NO_TRAP
2139 on all INDIRECT_REFs. */
2142 mark_notrap (tree
*tp
, int *walk_subtrees
, void *data ATTRIBUTE_UNUSED
)
2146 if (TREE_CODE (t
) == INDIRECT_REF
)
2148 TREE_THIS_NOTRAP (t
) = 1;
2151 else if (IS_TYPE_OR_DECL_P (t
))
2157 /* Scalarize a LDST. To recap, this is an assignment between one scalarizable
2158 reference ELT and one non-scalarizable reference OTHER. IS_OUTPUT is true
2159 if ELT is on the left-hand side. */
2162 scalarize_ldst (struct sra_elt
*elt
, tree other
,
2163 block_stmt_iterator
*bsi
, bool is_output
)
2165 /* Shouldn't have gotten called for a scalar. */
2166 gcc_assert (!elt
->replacement
);
2168 if (elt
->use_block_copy
)
2170 /* Since ELT is not fully instantiated, we have to leave the
2171 block copy in place. Treat this as a USE. */
2172 scalarize_use (elt
, NULL
, bsi
, is_output
, false);
2176 /* The interesting case is when ELT is fully instantiated. In this
2177 case we can have each element stored/loaded directly to/from the
2178 corresponding slot in OTHER. This avoids a block copy. */
2180 tree list
= NULL
, stmt
= bsi_stmt (*bsi
);
2182 mark_all_v_defs (stmt
);
2183 generate_copy_inout (elt
, is_output
, other
, &list
);
2184 mark_all_v_defs (list
);
2187 /* Preserve EH semantics. */
2188 if (stmt_ends_bb_p (stmt
))
2190 tree_stmt_iterator tsi
;
2193 /* Extract the first statement from LIST. */
2194 tsi
= tsi_start (list
);
2195 first
= tsi_stmt (tsi
);
2198 /* Replace the old statement with this new representative. */
2199 bsi_replace (bsi
, first
, true);
2201 if (!tsi_end_p (tsi
))
2203 /* If any reference would trap, then they all would. And more
2204 to the point, the first would. Therefore none of the rest
2205 will trap since the first didn't. Indicate this by
2206 iterating over the remaining statements and set
2207 TREE_THIS_NOTRAP in all INDIRECT_REFs. */
2210 walk_tree (tsi_stmt_ptr (tsi
), mark_notrap
, NULL
, NULL
);
2213 while (!tsi_end_p (tsi
));
2215 insert_edge_copies (list
, bsi
->bb
);
2219 sra_replace (bsi
, list
);
2223 /* Generate initializations for all scalarizable parameters. */
2226 scalarize_parms (void)
2232 EXECUTE_IF_SET_IN_BITMAP (needs_copy_in
, 0, i
, bi
)
2234 tree var
= referenced_var (i
);
2235 struct sra_elt
*elt
= lookup_element (NULL
, var
, NULL
, NO_INSERT
);
2236 generate_copy_inout (elt
, true, var
, &list
);
2241 insert_edge_copies (list
, ENTRY_BLOCK_PTR
);
2242 mark_all_v_defs (list
);
2246 /* Entry point to phase 4. Update the function to match replacements. */
2249 scalarize_function (void)
2251 static const struct sra_walk_fns fns
= {
2252 scalarize_use
, scalarize_copy
, scalarize_init
, scalarize_ldst
, false
2255 sra_walk_function (&fns
);
2257 bsi_commit_edge_inserts ();
2261 /* Debug helper function. Print ELT in a nice human-readable format. */
2264 dump_sra_elt_name (FILE *f
, struct sra_elt
*elt
)
2266 if (elt
->parent
&& TREE_CODE (elt
->parent
->type
) == COMPLEX_TYPE
)
2268 fputs (elt
->element
== integer_zero_node
? "__real__ " : "__imag__ ", f
);
2269 dump_sra_elt_name (f
, elt
->parent
);
2274 dump_sra_elt_name (f
, elt
->parent
);
2275 if (DECL_P (elt
->element
))
2277 if (TREE_CODE (elt
->element
) == FIELD_DECL
)
2279 print_generic_expr (f
, elt
->element
, dump_flags
);
2281 else if (TREE_CODE (elt
->element
) == RANGE_EXPR
)
2282 fprintf (f
, "["HOST_WIDE_INT_PRINT_DEC
".."HOST_WIDE_INT_PRINT_DEC
"]",
2283 TREE_INT_CST_LOW (TREE_OPERAND (elt
->element
, 0)),
2284 TREE_INT_CST_LOW (TREE_OPERAND (elt
->element
, 1)));
2286 fprintf (f
, "[" HOST_WIDE_INT_PRINT_DEC
"]",
2287 TREE_INT_CST_LOW (elt
->element
));
2291 /* Likewise, but callable from the debugger. */
2294 debug_sra_elt_name (struct sra_elt
*elt
)
2296 dump_sra_elt_name (stderr
, elt
);
2297 fputc ('\n', stderr
);
2301 sra_init_cache (void)
2303 if (sra_type_decomp_cache
)
2306 sra_type_decomp_cache
= BITMAP_ALLOC (NULL
);
2307 sra_type_inst_cache
= BITMAP_ALLOC (NULL
);
2310 /* Main entry point. */
2315 /* Initialize local variables. */
2317 gcc_obstack_init (&sra_obstack
);
2318 sra_candidates
= BITMAP_ALLOC (NULL
);
2319 needs_copy_in
= BITMAP_ALLOC (NULL
);
2321 sra_map
= htab_create (101, sra_elt_hash
, sra_elt_eq
, NULL
);
2323 /* Scan. If we find anything, instantiate and scalarize. */
2324 if (find_candidates_for_sra ())
2327 decide_instantiations ();
2328 scalarize_function ();
2331 /* Free allocated memory. */
2332 htab_delete (sra_map
);
2334 BITMAP_FREE (sra_candidates
);
2335 BITMAP_FREE (needs_copy_in
);
2336 BITMAP_FREE (sra_type_decomp_cache
);
2337 BITMAP_FREE (sra_type_inst_cache
);
2338 obstack_free (&sra_obstack
, NULL
);
2345 return flag_tree_sra
!= 0;
2348 struct tree_opt_pass pass_sra
=
2351 gate_sra
, /* gate */
2352 tree_sra
, /* execute */
2355 0, /* static_pass_number */
2356 TV_TREE_SRA
, /* tv_id */
2357 PROP_cfg
| PROP_ssa
| PROP_alias
, /* properties_required */
2358 0, /* properties_provided */
2359 PROP_smt_usage
, /* properties_destroyed */
2360 0, /* todo_flags_start */
2361 TODO_dump_func
/* todo_flags_finish */
2363 | TODO_ggc_collect
| TODO_verify_ssa
,