/* Scalar Replacement of Aggregates (SRA) converts some structure
references into scalar references, exposing them to the scalar
optimizers.
- Copyright (C) 2003, 2004, 2005 Free Software Foundation, Inc.
+ Copyright (C) 2003, 2004, 2005, 2006 Free Software Foundation, Inc.
Contributed by Diego Novillo <dnovillo@redhat.com>
This file is part of GCC.
*/
+/* The set of todo flags to return from tree_sra. */
+static unsigned int todoflags;
+
/* The set of aggregate variables that are candidates for scalarization. */
static bitmap sra_candidates;
static bitmap sra_type_decomp_cache;
static bitmap sra_type_inst_cache;
-/* One of these structures is created for each candidate aggregate
- and each (accessed) member of such an aggregate. */
+/* One of these structures is created for each candidate aggregate and
+ each (accessed) member or group of members of such an aggregate. */
struct sra_elt
{
/* A tree of the elements. Used when we want to traverse everything. */
struct sra_elt *parent;
+ struct sra_elt *groups;
struct sra_elt *children;
struct sra_elt *sibling;
/* If this element is a root, then this is the VAR_DECL. If this is
a sub-element, this is some token used to identify the reference.
In the case of COMPONENT_REF, this is the FIELD_DECL. In the case
- of an ARRAY_REF, this is the (constant) index. In the case of a
- complex number, this is a zero or one. */
+ of an ARRAY_REF, this is the (constant) index. In the case of an
+ ARRAY_RANGE_REF, this is the (constant) RANGE_EXPR. In the case
+ of a complex number, this is a zero or one. */
tree element;
/* The type of the element. */
/* True if TYPE is scalar. */
bool is_scalar;
+ /* True if this element is a group of members of its parent. */
+ bool is_group;
+
/* True if we saw something about this element that prevents scalarization,
such as non-constant indexing. */
bool cannot_scalarize;
should happen via memcpy and not per-element. */
bool use_block_copy;
+ /* True if everything under this element has been marked TREE_NO_WARNING. */
+ bool all_no_warning;
+
/* A flag for use with/after random access traversals. */
bool visited;
};
+#define IS_ELEMENT_FOR_GROUP(ELEMENT) (TREE_CODE (ELEMENT) == RANGE_EXPR)
+
+#define FOR_EACH_ACTUAL_CHILD(CHILD, ELT) \
+ for ((CHILD) = (ELT)->is_group \
+ ? next_child_for_group (NULL, (ELT)) \
+ : (ELT)->children; \
+ (CHILD); \
+ (CHILD) = (ELT)->is_group \
+ ? next_child_for_group ((CHILD), (ELT)) \
+ : (CHILD)->sibling)
+
+/* Helper function for above macro. Return next child in group. */
+static struct sra_elt *
+next_child_for_group (struct sra_elt *child, struct sra_elt *group)
+{
+ gcc_assert (group->is_group);
+
+ /* Find the next child in the parent. */
+ if (child)
+ child = child->sibling;
+ else
+ child = group->parent->children;
+
+ /* Skip siblings that do not belong to the group. */
+ while (child)
+ {
+ tree g_elt = group->element;
+ if (TREE_CODE (g_elt) == RANGE_EXPR)
+ {
+ if (!tree_int_cst_lt (child->element, TREE_OPERAND (g_elt, 0))
+ && !tree_int_cst_lt (TREE_OPERAND (g_elt, 1), child->element))
+ break;
+ }
+ else
+ gcc_unreachable ();
+
+ child = child->sibling;
+ }
+
+ return child;
+}
+
/* Random access to the child of a parent is performed by hashing.
This prevents quadratic behavior, and allows SRA to function
reasonably on larger records. */
enum tree_code code = TREE_CODE (type);
return (code == INTEGER_TYPE || code == REAL_TYPE || code == VECTOR_TYPE
|| code == ENUMERAL_TYPE || code == BOOLEAN_TYPE
- || code == CHAR_TYPE || code == POINTER_TYPE || code == OFFSET_TYPE
+ || code == POINTER_TYPE || code == OFFSET_TYPE
|| code == REFERENCE_TYPE);
}
if (elt->cannot_scalarize)
return false;
- for (c = elt->children; c ; c = c->sibling)
+ for (c = elt->children; c; c = c->sibling)
+ if (!can_completely_scalarize_p (c))
+ return false;
+
+ for (c = elt->groups; c; c = c->sibling)
if (!can_completely_scalarize_p (c))
return false;
h = TREE_INT_CST_LOW (t) ^ TREE_INT_CST_HIGH (t);
break;
+ case RANGE_EXPR:
+ h = iterative_hash_expr (TREE_OPERAND (t, 0), 0);
+ h = iterative_hash_expr (TREE_OPERAND (t, 1), h);
+ break;
+
case FIELD_DECL:
/* We can have types that are compatible, but have different member
lists, so we can't hash fields by ID. Use offsets instead. */
/* Integers are not pointer unique, so compare their values. */
return tree_int_cst_equal (ae, be);
+ case RANGE_EXPR:
+ return
+ tree_int_cst_equal (TREE_OPERAND (ae, 0), TREE_OPERAND (be, 0))
+ && tree_int_cst_equal (TREE_OPERAND (ae, 1), TREE_OPERAND (be, 1));
+
case FIELD_DECL:
/* Fields are unique within a record, but not between
compatible records. */
struct sra_elt **slot;
struct sra_elt *elt;
- dummy.parent = parent;
+ if (parent)
+ dummy.parent = parent->is_group ? parent->parent : parent;
+ else
+ dummy.parent = NULL;
dummy.element = child;
slot = (struct sra_elt **) htab_find_slot (sra_map, &dummy, insert);
if (parent)
{
- elt->sibling = parent->children;
- parent->children = elt;
+ if (IS_ELEMENT_FOR_GROUP (elt->element))
+ {
+ elt->is_group = true;
+ elt->sibling = parent->groups;
+ parent->groups = elt;
+ }
+ else
+ {
+ elt->sibling = parent->children;
+ parent->children = elt;
+ }
}
/* If this is a parameter, then if we want to scalarize, we have
return elt;
}
-/* Return true if the ARRAY_REF in EXPR is a constant, in bounds access. */
-
-static bool
-is_valid_const_index (tree expr)
-{
- tree dom, t, index = TREE_OPERAND (expr, 1);
-
- if (TREE_CODE (index) != INTEGER_CST)
- return false;
-
- /* Watch out for stupid user tricks, indexing outside the array.
-
- Careful, we're not called only on scalarizable types, so do not
- assume constant array bounds. We needn't do anything with such
- cases, since they'll be referring to objects that we should have
- already rejected for scalarization, so returning false is fine. */
-
- dom = TYPE_DOMAIN (TREE_TYPE (TREE_OPERAND (expr, 0)));
- if (dom == NULL)
- return false;
-
- t = TYPE_MIN_VALUE (dom);
- if (!t || TREE_CODE (t) != INTEGER_CST)
- return false;
- if (tree_int_cst_lt (index, t))
- return false;
-
- t = TYPE_MAX_VALUE (dom);
- if (!t || TREE_CODE (t) != INTEGER_CST)
- return false;
- if (tree_int_cst_lt (t, index))
- return false;
-
- return true;
-}
-
/* Create or return the SRA_ELT structure for EXPR if the expression
refers to a scalarizable variable. */
return NULL;
case ARRAY_REF:
- /* We can't scalarize variable array indicies. */
- if (is_valid_const_index (expr))
+ /* We can't scalarize variable array indices. */
+ if (in_array_bounds_p (expr))
child = TREE_OPERAND (expr, 1);
else
return NULL;
break;
+ case ARRAY_RANGE_REF:
+ /* We can't scalarize variable array indices. */
+ if (range_in_array_bounds_p (expr))
+ {
+ tree domain = TYPE_DOMAIN (TREE_TYPE (expr));
+ child = build2 (RANGE_EXPR, integer_type_node,
+ TYPE_MIN_VALUE (domain), TYPE_MAX_VALUE (domain));
+ }
+ else
+ return NULL;
+ break;
+
case COMPONENT_REF:
/* Don't look through unions. */
if (TREE_CODE (TREE_TYPE (TREE_OPERAND (expr, 0))) != RECORD_TYPE)
/* Invoked when ELT is required as a unit. Note that ELT might refer to
a leaf node, in which case this is a simple scalar reference. *EXPR_P
points to the location of the expression. IS_OUTPUT is true if this
- is a left-hand-side reference. */
+ is a left-hand-side reference. USE_ALL is true if we saw something we
+ couldn't quite identify and had to force the use of the entire object. */
void (*use) (struct sra_elt *elt, tree *expr_p,
- block_stmt_iterator *bsi, bool is_output);
+ block_stmt_iterator *bsi, bool is_output, bool use_all);
/* Invoked when we have a copy between two scalarizable references. */
void (*copy) (struct sra_elt *lhs_elt, struct sra_elt *rhs_elt,
tree expr = *expr_p;
tree inner = expr;
bool disable_scalarization = false;
+ bool use_all_p = false;
/* We're looking to collect a reference expression between EXPR and INNER,
such that INNER is a scalarizable decl and all other nodes through EXPR
if (disable_scalarization)
elt->cannot_scalarize = true;
else
- fns->use (elt, expr_p, bsi, is_output);
+ fns->use (elt, expr_p, bsi, is_output, use_all_p);
}
return;
the effort. */
/* ??? Hack. Figure out how to push this into the scan routines
without duplicating too much code. */
- if (!is_valid_const_index (inner))
+ if (!in_array_bounds_p (inner))
{
disable_scalarization = true;
goto use_all;
inner = TREE_OPERAND (inner, 0);
break;
+ case ARRAY_RANGE_REF:
+ if (!range_in_array_bounds_p (inner))
+ {
+ disable_scalarization = true;
+ goto use_all;
+ }
+ /* ??? See above non-constant bounds and stride . */
+ if (TREE_OPERAND (inner, 2) || TREE_OPERAND (inner, 3))
+ goto use_all;
+ inner = TREE_OPERAND (inner, 0);
+ break;
+
case COMPONENT_REF:
/* A reference to a union member constitutes a reference to the
entire union. */
complete outer element, to which walk_tree will bring us next. */
goto use_all;
- case ARRAY_RANGE_REF:
- /* Similarly, a subrange reference is used to modify indexing. Which
- means that the canonical element names that we have won't work. */
- goto use_all;
-
case VIEW_CONVERT_EXPR:
case NOP_EXPR:
/* Similarly, a view/nop explicitly wants to look at an object in a
use_all:
expr_p = &TREE_OPERAND (inner, 0);
inner = expr = *expr_p;
+ use_all_p = true;
break;
default:
if (!rhs_elt->is_scalar)
fns->ldst (rhs_elt, lhs, bsi, false);
else
- fns->use (rhs_elt, &TREE_OPERAND (expr, 1), bsi, false);
+ fns->use (rhs_elt, &TREE_OPERAND (expr, 1), bsi, false, false);
}
/* If it isn't scalarizable, there may be scalarizable variables within, so
/* Otherwise we're being used in some context that requires the
aggregate to be seen as a whole. Invoke USE. */
else
- fns->use (lhs_elt, &TREE_OPERAND (expr, 0), bsi, true);
+ fns->use (lhs_elt, &TREE_OPERAND (expr, 0), bsi, true, false);
}
/* Similarly to above, LHS_ELT being null only means that the LHS as a
static void
scan_use (struct sra_elt *elt, tree *expr_p ATTRIBUTE_UNUSED,
block_stmt_iterator *bsi ATTRIBUTE_UNUSED,
- bool is_output ATTRIBUTE_UNUSED)
+ bool is_output ATTRIBUTE_UNUSED, bool use_all ATTRIBUTE_UNUSED)
{
elt->n_uses += 1;
}
for (c = elt->children; c ; c = c->sibling)
scan_dump (c);
+
+ for (c = elt->groups; c ; c = c->sibling)
+ scan_dump (c);
}
/* Entry point to phase 2. Scan the entire function, building up
}
else
{
- struct sra_elt *c;
+ struct sra_elt *c, *group;
unsigned int this_uses = elt->n_uses + parent_uses;
unsigned int this_copies = elt->n_copies + parent_copies;
+ /* Consider groups of sub-elements as weighing in favour of
+ instantiation whatever their size. */
+ for (group = elt->groups; group ; group = group->sibling)
+ FOR_EACH_ACTUAL_CHILD (c, group)
+ {
+ c->n_uses += group->n_uses;
+ c->n_copies += group->n_copies;
+ }
+
for (c = elt->children; c ; c = c->sibling)
decide_instantiation_1 (c, this_uses, this_copies);
}
struct sra_elt *c;
bool any_inst;
+ /* We shouldn't be invoked on groups of sub-elements as they must
+ behave like their parent as far as block copy is concerned. */
+ gcc_assert (!elt->is_group);
+
/* If scalarization is disabled, respect it. */
if (elt->cannot_scalarize)
{
c->cannot_scalarize = 1;
decide_block_copy (c);
}
+
+ /* Groups behave like their parent. */
+ for (c = elt->groups; c; c = c->sibling)
+ {
+ c->cannot_scalarize = 1;
+ c->use_block_copy = 1;
+ }
+
return false;
}
else if (host_integerp (size_tree, 1))
{
unsigned HOST_WIDE_INT full_size, inst_size = 0;
- unsigned int max_size;
+ unsigned int max_size, max_count, inst_count, full_count;
/* If the sra-max-structure-size parameter is 0, then the
user has not overridden the parameter and we can choose a
max_size = SRA_MAX_STRUCTURE_SIZE
? SRA_MAX_STRUCTURE_SIZE
: MOVE_RATIO * UNITS_PER_WORD;
+ max_count = SRA_MAX_STRUCTURE_COUNT
+ ? SRA_MAX_STRUCTURE_COUNT
+ : MOVE_RATIO;
full_size = tree_low_cst (size_tree, 1);
+ full_count = count_type_elements (elt->type, false);
+ inst_count = sum_instantiated_sizes (elt, &inst_size);
/* ??? What to do here. If there are two fields, and we've only
instantiated one, then instantiating the other is clearly a win.
/* If the structure is small, and we've made copies, go ahead
and instantiate, hoping that the copies will go away. */
if (full_size <= max_size
+ && (full_count - inst_count) <= max_count
&& elt->n_copies > elt->n_uses)
use_block_copy = false;
- else
- {
- sum_instantiated_sizes (elt, &inst_size);
-
- if (inst_size * 100 >= full_size * SRA_FIELD_STRUCTURE_RATIO)
- use_block_copy = false;
- }
+ else if (inst_count * 100 >= full_count * SRA_FIELD_STRUCTURE_RATIO
+ && inst_size * 100 >= full_size * SRA_FIELD_STRUCTURE_RATIO)
+ use_block_copy = false;
/* In order to avoid block copy, we have to be able to instantiate
all elements of the type. See if this is possible. */
|| !type_can_instantiate_all_elements (elt->type)))
use_block_copy = true;
}
+
elt->use_block_copy = use_block_copy;
+ /* Groups behave like their parent. */
+ for (c = elt->groups; c; c = c->sibling)
+ c->use_block_copy = use_block_copy;
+
if (dump_file)
{
fprintf (dump_file, "Using %s for ",
bitmap_and_compl_into (needs_copy_in, &done_head);
}
bitmap_clear (&done_head);
+
+ if (!bitmap_empty_p (sra_candidates))
+ todoflags |= TODO_update_smt_usage;
mark_set_for_renaming (sra_candidates);
}
}
+/* Mark every replacement under ELT with TREE_NO_WARNING. */
+
+static void
+mark_no_warning (struct sra_elt *elt)
+{
+ if (!elt->all_no_warning)
+ {
+ if (elt->replacement)
+ TREE_NO_WARNING (elt->replacement) = 1;
+ else
+ {
+ struct sra_elt *c;
+ FOR_EACH_ACTUAL_CHILD (c, elt)
+ mark_no_warning (c);
+ }
+ elt->all_no_warning = true;
+ }
+}
/* Build a single level component reference to ELT rooted at BASE. */
if (DECL_FIELD_CONTEXT (field) != TYPE_MAIN_VARIANT (TREE_TYPE (base)))
field = find_compatible_field (TREE_TYPE (base), field);
- return build (COMPONENT_REF, elt->type, base, field, NULL);
+ return build3 (COMPONENT_REF, elt->type, base, field, NULL);
}
case ARRAY_TYPE:
- return build (ARRAY_REF, elt->type, base, elt->element, NULL, NULL);
+ todoflags |= TODO_update_smt_usage;
+ if (TREE_CODE (elt->element) == RANGE_EXPR)
+ return build4 (ARRAY_RANGE_REF, elt->type, base,
+ TREE_OPERAND (elt->element, 0), NULL, NULL);
+ else
+ return build4 (ARRAY_REF, elt->type, base, elt->element, NULL, NULL);
case COMPLEX_TYPE:
if (elt->element == integer_zero_node)
- return build (REALPART_EXPR, elt->type, base);
+ return build1 (REALPART_EXPR, elt->type, base);
else
- return build (IMAGPART_EXPR, elt->type, base);
+ return build1 (IMAGPART_EXPR, elt->type, base);
default:
gcc_unreachable ();
c = lookup_element (elt, integer_one_node, NULL, NO_INSERT);
i = c->replacement;
- t = build (COMPLEX_EXPR, elt->type, r, i);
- t = build (MODIFY_EXPR, void_type_node, expr, t);
+ t = build2 (COMPLEX_EXPR, elt->type, r, i);
+ t = build2 (MODIFY_EXPR, void_type_node, expr, t);
SSA_NAME_DEF_STMT (expr) = t;
append_to_statement_list (t, list_p);
}
else if (elt->replacement)
{
if (copy_out)
- t = build (MODIFY_EXPR, void_type_node, elt->replacement, expr);
+ t = build2 (MODIFY_EXPR, void_type_node, elt->replacement, expr);
else
- t = build (MODIFY_EXPR, void_type_node, expr, elt->replacement);
+ t = build2 (MODIFY_EXPR, void_type_node, expr, elt->replacement);
append_to_statement_list (t, list_p);
}
else
{
- for (c = elt->children; c ; c = c->sibling)
+ FOR_EACH_ACTUAL_CHILD (c, elt)
{
t = generate_one_element_ref (c, unshare_expr (expr));
generate_copy_inout (c, copy_out, t, list_p);
{
struct sra_elt *dc, *sc;
- for (dc = dst->children; dc ; dc = dc->sibling)
+ FOR_EACH_ACTUAL_CHILD (dc, dst)
{
sc = lookup_element (src, dc->element, NULL, NO_INSERT);
gcc_assert (sc);
gcc_assert (src->replacement);
- t = build (MODIFY_EXPR, void_type_node, dst->replacement,
- src->replacement);
+ t = build2 (MODIFY_EXPR, void_type_node, dst->replacement,
+ src->replacement);
append_to_statement_list (t, list_p);
}
}
return;
}
- for (c = elt->children; c ; c = c->sibling)
+ FOR_EACH_ACTUAL_CHILD (c, elt)
generate_element_zero (c, list_p);
if (elt->replacement)
gcc_assert (elt->is_scalar);
t = fold_convert (elt->type, integer_zero_node);
- t = build (MODIFY_EXPR, void_type_node, elt->replacement, t);
+ t = build2 (MODIFY_EXPR, void_type_node, elt->replacement, t);
append_to_statement_list (t, list_p);
}
}
generate_one_element_init (tree var, tree init, tree *list_p)
{
/* The replacement can be almost arbitrarily complex. Gimplify. */
- tree stmt = build (MODIFY_EXPR, void_type_node, var, init);
+ tree stmt = build2 (MODIFY_EXPR, void_type_node, var, init);
gimplify_and_add (stmt, list_p);
}
{
case COMPLEX_CST:
case COMPLEX_EXPR:
- for (sub = elt->children; sub ; sub = sub->sibling)
+ FOR_EACH_ACTUAL_CHILD (sub, elt)
{
if (sub->element == integer_zero_node)
t = (init_code == COMPLEX_EXPR
sra_replace (block_stmt_iterator *bsi, tree list)
{
sra_insert_before (bsi, list);
- bsi_remove (bsi);
+ bsi_remove (bsi, false);
if (bsi_end_p (*bsi))
*bsi = bsi_last (bsi->bb);
else
static void
scalarize_use (struct sra_elt *elt, tree *expr_p, block_stmt_iterator *bsi,
- bool is_output)
+ bool is_output, bool use_all)
{
tree list = NULL, stmt = bsi_stmt (*bsi);
if (is_output)
sra_insert_after (bsi, list);
else
- sra_insert_before (bsi, list);
+ {
+ sra_insert_before (bsi, list);
+ if (use_all)
+ mark_no_warning (elt);
+ }
}
}
{
/* Since ELT is not fully instantiated, we have to leave the
block copy in place. Treat this as a USE. */
- scalarize_use (elt, NULL, bsi, is_output);
+ scalarize_use (elt, NULL, bsi, is_output, false);
}
else
{
fputc ('.', f);
print_generic_expr (f, elt->element, dump_flags);
}
+ else if (TREE_CODE (elt->element) == RANGE_EXPR)
+ fprintf (f, "["HOST_WIDE_INT_PRINT_DEC".."HOST_WIDE_INT_PRINT_DEC"]",
+ TREE_INT_CST_LOW (TREE_OPERAND (elt->element, 0)),
+ TREE_INT_CST_LOW (TREE_OPERAND (elt->element, 1)));
else
fprintf (f, "[" HOST_WIDE_INT_PRINT_DEC "]",
TREE_INT_CST_LOW (elt->element));
/* Main entry point. */
-static void
+static unsigned int
tree_sra (void)
{
/* Initialize local variables. */
+ todoflags = 0;
gcc_obstack_init (&sra_obstack);
sra_candidates = BITMAP_ALLOC (NULL);
needs_copy_in = BITMAP_ALLOC (NULL);
BITMAP_FREE (sra_type_decomp_cache);
BITMAP_FREE (sra_type_inst_cache);
obstack_free (&sra_obstack, NULL);
+ return todoflags;
}
static bool
TV_TREE_SRA, /* tv_id */
PROP_cfg | PROP_ssa | PROP_alias, /* properties_required */
0, /* properties_provided */
- 0, /* properties_destroyed */
+ PROP_smt_usage, /* properties_destroyed */
0, /* todo_flags_start */
- TODO_dump_func | TODO_update_ssa
- | TODO_ggc_collect | TODO_verify_ssa, /* todo_flags_finish */
+ TODO_dump_func /* todo_flags_finish */
+ | TODO_update_ssa
+ | TODO_ggc_collect | TODO_verify_ssa,
0 /* letter */
};