marshalling to implement data sharing and copying clauses.
Contributed by Diego Novillo <dnovillo@redhat.com>
- Copyright (C) 2005 Free Software Foundation, Inc.
+ Copyright (C) 2005, 2006, 2007, 2008 Free Software Foundation, Inc.
This file is part of GCC.
GCC is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free
-Software Foundation; either version 2, or (at your option) any later
+Software Foundation; either version 3, or (at your option) any later
version.
GCC is distributed in the hope that it will be useful, but WITHOUT ANY
for more details.
You should have received a copy of the GNU General Public License
-along with GCC; see the file COPYING. If not, write to the Free
-Software Foundation, 51 Franklin Street, Fifth Floor, Boston, MA
-02110-1301, USA. */
+along with GCC; see the file COPYING3. If not see
+<http://www.gnu.org/licenses/>. */
#include "config.h"
#include "system.h"
#include "tree-pass.h"
#include "ggc.h"
#include "except.h"
-
+#include "splay-tree.h"
+#include "optabs.h"
+#include "cfgloop.h"
/* Lowering of OpenMP parallel and workshare constructs proceeds in two
phases. The first phase scans the function looking for OMP statements
and then for variables that must be replaced to satisfy data sharing
clauses. The second phase expands code for the constructs, as well as
- re-gimplifing things when variables have been replaced with complex
+ re-gimplifying things when variables have been replaced with complex
expressions.
- Lowering of a parallel statement results in the contents of the
- parallel being moved to a new function, to be invoked by the thread
- library. The variable remapping process is complex enough that only
- one level of parallel statement is handled at one time. If there are
- nested parallel statements, those nested statements are handled when
- the new function is lowered and optimized. The result is not 100%
- optimal, but lexically nested parallels effectively only happens in
- test suites. */
-
-/* Parallel region information. Every parallel and workshare
- directive is enclosed between two markers, the OMP_* directive
- and a corresponding OMP_RETURN_EXPR statement. */
-
-struct omp_region GTY(())
-{
- /* The enclosing region. */
- struct omp_region *outer;
-
- /* First child region. */
- struct omp_region *inner;
-
- /* Next peer region. */
- struct omp_region *next;
-
- /* Entry point to this region. */
- tree entry;
-
- /* Exit label from this region. */
- tree exit;
-
- /* Region number. */
- int num;
-
- /* True if this is a combined parallel+workshare region. */
- bool is_combined_parallel;
-
- /* If this is a combined parallel+workshare region, this is a list
- of additional arguments needed by the combined parallel+workshare
- library call. */
- tree ws_args;
-};
-
+ Final code generation is done by pass_expand_omp. The flowgraph is
+ scanned for parallel regions which are then moved to a new
+ function, to be invoked by the thread library. */
/* Context structure. Used to store information about each parallel
directive in the code. */
static splay_tree all_contexts;
static int parallel_nesting_level;
-static splay_tree omp_regions;
-static struct omp_region *root_omp_region;
+struct omp_region *root_omp_region;
static void scan_omp (tree *, omp_context *);
static void lower_omp (tree *, omp_context *);
+static tree lookup_decl_in_outer_ctx (tree, omp_context *);
+static tree maybe_lookup_decl_in_outer_ctx (tree, omp_context *);
/* Find an OpenMP clause of type KIND within CLAUSES. */
-static tree
+tree
find_omp_clause (tree clauses, enum tree_code kind)
{
for (; clauses ; clauses = OMP_CLAUSE_CHAIN (clauses))
- if (TREE_CODE (clauses) == kind)
+ if (OMP_CLAUSE_CODE (clauses) == kind)
return clauses;
return NULL_TREE;
static void
extract_omp_for_data (tree for_stmt, struct omp_for_data *fd)
{
- tree t;
+ tree t, var;
fd->for_stmt = for_stmt;
fd->pre = NULL;
t = OMP_FOR_INIT (for_stmt);
- gcc_assert (TREE_CODE (t) == MODIFY_EXPR);
- fd->v = TREE_OPERAND (t, 0);
- gcc_assert (DECL_P (fd->v));
+ gcc_assert (TREE_CODE (t) == GIMPLE_MODIFY_STMT);
+ fd->v = GIMPLE_STMT_OPERAND (t, 0);
+ gcc_assert (SSA_VAR_P (fd->v));
gcc_assert (TREE_CODE (TREE_TYPE (fd->v)) == INTEGER_TYPE);
- fd->n1 = TREE_OPERAND (t, 1);
+ var = TREE_CODE (fd->v) == SSA_NAME ? SSA_NAME_VAR (fd->v) : fd->v;
+ fd->n1 = GIMPLE_STMT_OPERAND (t, 1);
t = OMP_FOR_COND (for_stmt);
fd->cond_code = TREE_CODE (t);
- gcc_assert (TREE_OPERAND (t, 0) == fd->v);
+ gcc_assert (TREE_OPERAND (t, 0) == var);
fd->n2 = TREE_OPERAND (t, 1);
switch (fd->cond_code)
{
}
t = OMP_FOR_INCR (fd->for_stmt);
- gcc_assert (TREE_CODE (t) == MODIFY_EXPR);
- gcc_assert (TREE_OPERAND (t, 0) == fd->v);
- t = TREE_OPERAND (t, 1);
- gcc_assert (TREE_OPERAND (t, 0) == fd->v);
+ gcc_assert (TREE_CODE (t) == GIMPLE_MODIFY_STMT);
+ gcc_assert (GIMPLE_STMT_OPERAND (t, 0) == var);
+ t = GIMPLE_STMT_OPERAND (t, 1);
+ gcc_assert (TREE_OPERAND (t, 0) == var);
switch (TREE_CODE (t))
{
case PLUS_EXPR:
fd->chunk_size = NULL_TREE;
for (t = OMP_FOR_CLAUSES (for_stmt); t ; t = OMP_CLAUSE_CHAIN (t))
- switch (TREE_CODE (t))
+ switch (OMP_CLAUSE_CODE (t))
{
case OMP_CLAUSE_NOWAIT:
fd->have_nowait = true;
}
else if (TREE_CODE (ws_stmt) == OMP_SECTIONS)
{
- basic_block bb = bb_for_stmt (ws_stmt);
- t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs));
+ /* Number of sections is equal to the number of edges from the
+ OMP_SECTIONS_SWITCH statement, except for the one to the exit
+ of the sections region. */
+ basic_block bb = single_succ (bb_for_stmt (ws_stmt));
+ t = build_int_cst (unsigned_type_node, EDGE_COUNT (bb->succs) - 1);
t = tree_cons (NULL, t, NULL);
return t;
}
basic_block par_entry_bb, par_exit_bb;
basic_block ws_entry_bb, ws_exit_bb;
- if (region == NULL || region->inner == NULL)
+ if (region == NULL || region->inner == NULL
+ || region->exit == NULL || region->inner->exit == NULL
+ || region->inner->cont == NULL)
return;
/* We only support parallel+for and parallel+sections. */
- if (TREE_CODE (region->entry) != OMP_PARALLEL
- || (TREE_CODE (region->inner->entry) != OMP_FOR
- && TREE_CODE (region->inner->entry) != OMP_SECTIONS))
+ if (region->type != OMP_PARALLEL
+ || (region->inner->type != OMP_FOR
+ && region->inner->type != OMP_SECTIONS))
return;
/* Check for perfect nesting PAR_ENTRY_BB -> WS_ENTRY_BB and
WS_EXIT_BB -> PAR_EXIT_BB. */
- par_entry_bb = bb_for_stmt (region->entry);
- par_exit_bb = bb_for_stmt (region->exit);
-
- ws_entry_bb = bb_for_stmt (region->inner->entry);
- ws_exit_bb = bb_for_stmt (region->inner->exit);
+ par_entry_bb = region->entry;
+ par_exit_bb = region->exit;
+ ws_entry_bb = region->inner->entry;
+ ws_exit_bb = region->inner->exit;
if (single_succ (par_entry_bb) == ws_entry_bb
&& single_succ (ws_exit_bb) == par_exit_bb
- && workshare_safe_to_combine_p (par_entry_bb, ws_entry_bb))
+ && workshare_safe_to_combine_p (par_entry_bb, ws_entry_bb)
+ && (OMP_PARALLEL_COMBINED (last_stmt (par_entry_bb))
+ || (last_and_only_stmt (ws_entry_bb)
+ && last_and_only_stmt (par_exit_bb))))
{
- if (TREE_CODE (region->inner->entry) == OMP_FOR)
+ tree ws_stmt = last_stmt (ws_entry_bb);
+
+ if (region->inner->type == OMP_FOR)
{
/* If this is a combined parallel loop, we need to determine
whether or not to use the combined library calls. There
parallel loop call would still need extra synchronization
to implement ordered semantics, so there would not be any
gain in using the combined call. */
- tree clauses = OMP_FOR_CLAUSES (region->inner->entry);
+ tree clauses = OMP_FOR_CLAUSES (ws_stmt);
tree c = find_omp_clause (clauses, OMP_CLAUSE_SCHEDULE);
if (c == NULL
|| OMP_CLAUSE_SCHEDULE_KIND (c) == OMP_CLAUSE_SCHEDULE_STATIC
region->is_combined_parallel = true;
region->inner->is_combined_parallel = true;
- region->ws_args = get_ws_args_for (region->inner->entry);
+ region->ws_args = get_ws_args_for (ws_stmt);
}
}
/* Return true if EXPR is variable sized. */
static inline bool
-is_variable_sized (tree expr)
+is_variable_sized (const_tree expr)
{
return !TREE_CONSTANT (TYPE_SIZE_UNIT (TREE_TYPE (expr)));
}
static inline tree
lookup_decl (tree var, omp_context *ctx)
{
- splay_tree_node n;
- n = splay_tree_lookup (ctx->cb.decl_map, (splay_tree_key) var);
- return (tree) n->value;
+ tree *n;
+ n = (tree *) pointer_map_contains (ctx->cb.decl_map, var);
+ return *n;
}
static inline tree
maybe_lookup_decl (tree var, omp_context *ctx)
{
- splay_tree_node n;
- n = splay_tree_lookup (ctx->cb.decl_map, (splay_tree_key) var);
- return n ? (tree) n->value : NULL_TREE;
+ tree *n;
+ n = (tree *) pointer_map_contains (ctx->cb.decl_map, var);
+ return n ? *n : NULL_TREE;
}
static inline tree
if DECL is to be shared. */
static bool
-use_pointer_for_field (tree decl, bool shared_p)
+use_pointer_for_field (const_tree decl, bool shared_p)
{
if (AGGREGATE_TYPE_P (TREE_TYPE (decl)))
return true;
- /* We can only use copy-in/copy-out semantics for shared varibles
+ /* We can only use copy-in/copy-out semantics for shared variables
when we know the value is not accessible from an outer scope. */
if (shared_p)
{
without analyzing the expression whether or not its location
is accessible to anyone else. In the case of nested parallel
regions it certainly may be. */
- if (DECL_HAS_VALUE_EXPR_P (decl))
+ if (TREE_CODE (decl) != RESULT_DECL && DECL_HAS_VALUE_EXPR_P (decl))
return true;
/* Do not use copy-in/copy-out for variables that have their
return false;
}
-/* Construct a new automatic decl similar to VAR. */
+/* Create a new VAR_DECL and copy information from VAR to it. */
-static tree
-omp_copy_decl_2 (tree var, tree name, tree type, omp_context *ctx)
+tree
+copy_var_decl (tree var, tree name, tree type)
{
tree copy = build_decl (VAR_DECL, name, type);
TREE_ADDRESSABLE (copy) = TREE_ADDRESSABLE (var);
- DECL_COMPLEX_GIMPLE_REG_P (copy) = DECL_COMPLEX_GIMPLE_REG_P (var);
+ TREE_THIS_VOLATILE (copy) = TREE_THIS_VOLATILE (var);
+ DECL_GIMPLE_REG_P (copy) = DECL_GIMPLE_REG_P (var);
+ DECL_NO_TBAA_P (copy) = DECL_NO_TBAA_P (var);
DECL_ARTIFICIAL (copy) = DECL_ARTIFICIAL (var);
DECL_IGNORED_P (copy) = DECL_IGNORED_P (var);
+ DECL_CONTEXT (copy) = DECL_CONTEXT (var);
+ DECL_SOURCE_LOCATION (copy) = DECL_SOURCE_LOCATION (var);
TREE_USED (copy) = 1;
- DECL_CONTEXT (copy) = current_function_decl;
DECL_SEEN_IN_BIND_EXPR_P (copy) = 1;
+ return copy;
+}
+
+/* Construct a new automatic decl similar to VAR. */
+
+static tree
+omp_copy_decl_2 (tree var, tree name, tree type, omp_context *ctx)
+{
+ tree copy = copy_var_decl (var, name, type);
+
+ DECL_CONTEXT (copy) = current_function_decl;
TREE_CHAIN (copy) = ctx->block_vars;
ctx->block_vars = copy;
{
tree x;
- if (is_global_var (var))
+ if (is_global_var (maybe_lookup_decl_in_outer_ctx (var, ctx)))
x = var;
else if (is_variable_sized (var))
{
}
else if (ctx->outer)
x = lookup_decl (var, ctx->outer);
+ else if (is_reference (var))
+ /* This can happen with orphaned constructs. If var is reference, it is
+ possible it is shared and as such valid. */
+ x = var;
else
gcc_unreachable ();
omp_context *ctx = (omp_context *) cb;
tree new_var;
- if (is_global_var (var) || decl_function_context (var) != ctx->cb.src_fn)
- return var;
-
if (TREE_CODE (var) == LABEL_DECL)
{
new_var = create_artificial_label ();
return new_var;
}
+ if (is_global_var (var) || decl_function_context (var) != ctx->cb.src_fn)
+ return var;
+
return error_mark_node;
}
/* Return the parallel region associated with STMT. */
-static inline struct omp_region *
-lookup_omp_region (tree stmt)
-{
- splay_tree_node n = splay_tree_lookup (omp_regions, (splay_tree_key) stmt);
- return n ? (struct omp_region *) n->value : NULL;
-}
-
-
/* Debugging dumps for parallel regions. */
void dump_omp_region (FILE *, struct omp_region *, int);
void debug_omp_region (struct omp_region *);
void
dump_omp_region (FILE *file, struct omp_region *region, int indent)
{
- fprintf (file, "%*s", indent, "");
- print_generic_stmt (file, region->entry, TDF_SLIM);
+ fprintf (file, "%*sbb %d: %s\n", indent, "", region->entry->index,
+ tree_code_name[region->type]);
if (region->inner)
dump_omp_region (file, region->inner, indent + 4);
- fprintf (file, "%*s", indent, "");
+ if (region->cont)
+ {
+ fprintf (file, "%*sbb %d: OMP_CONTINUE\n", indent, "",
+ region->cont->index);
+ }
+
if (region->exit)
- print_generic_stmt (file, region->exit, 0);
+ fprintf (file, "%*sbb %d: OMP_RETURN\n", indent, "",
+ region->exit->index);
else
- fprintf (file, "[no exit marker]\n");
+ fprintf (file, "%*s[no exit marker]\n", indent, "");
if (region->next)
- {
- fprintf (file, "\n");
- dump_omp_region (file, region->next, indent);
- }
+ dump_omp_region (file, region->next, indent);
}
void
/* Create a new parallel region starting at STMT inside region PARENT. */
-static struct omp_region *
-new_omp_region (tree stmt, struct omp_region *parent)
+struct omp_region *
+new_omp_region (basic_block bb, enum tree_code type, struct omp_region *parent)
{
- struct omp_region *region = ggc_alloc_cleared (sizeof (*region));
- static int num = 0;
+ struct omp_region *region = xcalloc (1, sizeof (*region));
region->outer = parent;
- region->entry = stmt;
- region->num = num++;
+ region->entry = bb;
+ region->type = type;
if (parent)
{
region->next = parent->inner;
parent->inner = region;
}
- else if (omp_regions)
+ else
{
/* This is a toplevel region. Add it to the list of toplevel
regions in ROOT_OMP_REGION. */
region->next = root_omp_region;
root_omp_region = region;
}
- else
+
+ return region;
+}
+
+/* Release the memory associated with the region tree rooted at REGION. */
+
+static void
+free_omp_region_1 (struct omp_region *region)
+{
+ struct omp_region *i, *n;
+
+ for (i = region->inner; i ; i = n)
{
- /* Create a new root region with the first region we find. */
- root_omp_region = region;
- omp_regions = splay_tree_new (splay_tree_compare_pointers, 0, 0);
+ n = i->next;
+ free_omp_region_1 (i);
}
- splay_tree_insert (omp_regions, (splay_tree_key) stmt,
- (splay_tree_value) region);
+ free (region);
+}
- return region;
+/* Release the memory for the entire omp region tree. */
+
+void
+free_omp_regions (void)
+{
+ struct omp_region *r, *n;
+ for (r = root_omp_region; r ; r = n)
+ {
+ n = r->next;
+ free_omp_region_1 (r);
+ }
+ root_omp_region = NULL;
}
ctx->depth = 1;
}
- ctx->cb.decl_map = splay_tree_new (splay_tree_compare_pointers, 0, 0);
+ ctx->cb.decl_map = pointer_map_create ();
return ctx;
}
{
omp_context *ctx = (omp_context *) value;
- splay_tree_delete (ctx->cb.decl_map);
+ pointer_map_destroy (ctx->cb.decl_map);
if (ctx->field_map)
splay_tree_delete (ctx->field_map);
{
bool by_ref;
- switch (TREE_CODE (c))
+ switch (OMP_CLAUSE_CODE (c))
{
case OMP_CLAUSE_PRIVATE:
decl = OMP_CLAUSE_DECL (c);
decl = OMP_CLAUSE_DECL (c);
gcc_assert (!is_variable_sized (decl));
by_ref = use_pointer_for_field (decl, true);
+ /* Global variables don't need to be copied,
+ the receiver side will use them directly. */
+ if (is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)))
+ break;
if (! TREE_READONLY (decl)
|| TREE_ADDRESSABLE (decl)
|| by_ref
break;
}
/* We don't need to copy const scalar vars back. */
- TREE_SET_CODE (c, OMP_CLAUSE_FIRSTPRIVATE);
+ OMP_CLAUSE_SET_CODE (c, OMP_CLAUSE_FIRSTPRIVATE);
goto do_private;
case OMP_CLAUSE_LASTPRIVATE:
do_private:
if (is_variable_sized (decl))
break;
- else if (is_parallel_ctx (ctx))
+ else if (is_parallel_ctx (ctx)
+ && ! is_global_var (maybe_lookup_decl_in_outer_ctx (decl,
+ ctx)))
{
by_ref = use_pointer_for_field (decl, false);
install_var_field (decl, by_ref, ctx);
case OMP_CLAUSE_NUM_THREADS:
case OMP_CLAUSE_SCHEDULE:
if (ctx->outer)
- scan_omp (&TREE_OPERAND (c, 0), ctx->outer);
+ scan_omp (&OMP_CLAUSE_OPERAND (c, 0), ctx->outer);
break;
case OMP_CLAUSE_NOWAIT:
for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
{
- switch (TREE_CODE (c))
+ switch (OMP_CLAUSE_CODE (c))
{
case OMP_CLAUSE_LASTPRIVATE:
/* Let the corresponding firstprivate clause create
if (is_variable_sized (decl))
install_var_local (decl, ctx);
fixup_remapped_decl (decl, ctx,
- TREE_CODE (c) == OMP_CLAUSE_PRIVATE
+ OMP_CLAUSE_CODE (c) == OMP_CLAUSE_PRIVATE
&& OMP_CLAUSE_PRIVATE_DEBUG (c));
- if (TREE_CODE (c) == OMP_CLAUSE_REDUCTION
+ if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
&& OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
scan_array_reductions = true;
break;
case OMP_CLAUSE_SHARED:
decl = OMP_CLAUSE_DECL (c);
- fixup_remapped_decl (decl, ctx, false);
+ if (! is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)))
+ fixup_remapped_decl (decl, ctx, false);
break;
case OMP_CLAUSE_COPYPRIVATE:
if (scan_array_reductions)
for (c = clauses; c; c = OMP_CLAUSE_CHAIN (c))
- if (TREE_CODE (c) == OMP_CLAUSE_REDUCTION
+ if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
&& OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
{
scan_omp (&OMP_CLAUSE_REDUCTION_INIT (c), ctx);
/* Allocate memory for the function structure. The call to
allocate_struct_function clobbers CFUN, so we need to restore
it afterward. */
- allocate_struct_function (decl);
+ push_struct_function (decl);
DECL_SOURCE_LOCATION (decl) = EXPR_LOCATION (ctx->stmt);
cfun->function_end_locus = EXPR_LOCATION (ctx->stmt);
- cfun = ctx->cb.src_cfun;
+ pop_cfun ();
}
}
+/* Check OpenMP nesting restrictions. */
+static void
+check_omp_nesting_restrictions (tree t, omp_context *ctx)
+{
+ switch (TREE_CODE (t))
+ {
+ case OMP_FOR:
+ case OMP_SECTIONS:
+ case OMP_SINGLE:
+ for (; ctx != NULL; ctx = ctx->outer)
+ switch (TREE_CODE (ctx->stmt))
+ {
+ case OMP_FOR:
+ case OMP_SECTIONS:
+ case OMP_SINGLE:
+ case OMP_ORDERED:
+ case OMP_MASTER:
+ warning (0, "work-sharing region may not be closely nested inside "
+ "of work-sharing, critical, ordered or master region");
+ return;
+ case OMP_PARALLEL:
+ return;
+ default:
+ break;
+ }
+ break;
+ case OMP_MASTER:
+ for (; ctx != NULL; ctx = ctx->outer)
+ switch (TREE_CODE (ctx->stmt))
+ {
+ case OMP_FOR:
+ case OMP_SECTIONS:
+ case OMP_SINGLE:
+ warning (0, "master region may not be closely nested inside "
+ "of work-sharing region");
+ return;
+ case OMP_PARALLEL:
+ return;
+ default:
+ break;
+ }
+ break;
+ case OMP_ORDERED:
+ for (; ctx != NULL; ctx = ctx->outer)
+ switch (TREE_CODE (ctx->stmt))
+ {
+ case OMP_CRITICAL:
+ warning (0, "ordered region may not be closely nested inside "
+ "of critical region");
+ return;
+ case OMP_FOR:
+ if (find_omp_clause (OMP_CLAUSES (ctx->stmt),
+ OMP_CLAUSE_ORDERED) == NULL)
+ warning (0, "ordered region must be closely nested inside "
+ "a loop region with an ordered clause");
+ return;
+ case OMP_PARALLEL:
+ return;
+ default:
+ break;
+ }
+ break;
+ case OMP_CRITICAL:
+ for (; ctx != NULL; ctx = ctx->outer)
+ if (TREE_CODE (ctx->stmt) == OMP_CRITICAL
+ && OMP_CRITICAL_NAME (t) == OMP_CRITICAL_NAME (ctx->stmt))
+ {
+ warning (0, "critical region may not be nested inside a critical "
+ "region with the same name");
+ return;
+ }
+ break;
+ default:
+ break;
+ }
+}
+
+
/* Callback for walk_stmts used to scan for OpenMP directives at TP. */
static tree
if (EXPR_HAS_LOCATION (t))
input_location = EXPR_LOCATION (t);
+ /* Check the OpenMP nesting restrictions. */
+ if (OMP_DIRECTIVE_P (t) && ctx != NULL)
+ check_omp_nesting_restrictions (t, ctx);
+
*walk_subtrees = 0;
switch (TREE_CODE (t))
{
case VAR_DECL:
case PARM_DECL:
case LABEL_DECL:
+ case RESULT_DECL:
if (ctx)
*tp = remap_decl (t, &ctx->cb);
break;
/* Build a call to GOMP_barrier. */
-static void
-build_omp_barrier (tree *stmt_list)
+static tree
+build_omp_barrier (void)
{
- tree t;
-
- t = built_in_decls[BUILT_IN_GOMP_BARRIER];
- t = build_function_call_expr (t, NULL);
- gimplify_and_add (t, stmt_list);
+ return build_call_expr (built_in_decls[BUILT_IN_GOMP_BARRIER], 0);
}
/* If a context was created for STMT when it was scanned, return it. */
tree t;
omp_context *up;
- gcc_assert (ctx->is_nested);
-
for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer)
t = maybe_lookup_decl (decl, up);
- gcc_assert (t);
+ gcc_assert (!ctx->is_nested || t || is_global_var (decl));
- return t;
+ return t ? t : decl;
+}
+
+
+/* Similar to lookup_decl_in_outer_ctx, but return DECL if not found
+ in outer contexts. */
+
+static tree
+maybe_lookup_decl_in_outer_ctx (tree decl, omp_context *ctx)
+{
+ tree t = NULL;
+ omp_context *up;
+
+ for (up = ctx->outer, t = NULL; up && t == NULL; up = up->outer)
+ t = maybe_lookup_decl (decl, up);
+
+ return t ? t : decl;
}
static void
lower_rec_input_clauses (tree clauses, tree *ilist, tree *dlist,
- omp_context *ctx)
+ omp_context *ctx)
{
tree_stmt_iterator diter;
- tree c, dtor, copyin_seq, x, args, ptr;
+ tree c, dtor, copyin_seq, x, ptr;
bool copyin_by_ref = false;
+ bool lastprivate_firstprivate = false;
int pass;
*dlist = alloc_stmt_list ();
{
for (c = clauses; c ; c = OMP_CLAUSE_CHAIN (c))
{
- enum tree_code c_kind = TREE_CODE (c);
+ enum omp_clause_code c_kind = OMP_CLAUSE_CODE (c);
tree var, new_var;
bool by_ref;
continue;
break;
case OMP_CLAUSE_SHARED:
+ if (maybe_lookup_decl (OMP_CLAUSE_DECL (c), ctx) == NULL)
+ {
+ gcc_assert (is_global_var (OMP_CLAUSE_DECL (c)));
+ continue;
+ }
case OMP_CLAUSE_FIRSTPRIVATE:
- case OMP_CLAUSE_LASTPRIVATE:
case OMP_CLAUSE_COPYIN:
case OMP_CLAUSE_REDUCTION:
break;
+ case OMP_CLAUSE_LASTPRIVATE:
+ if (OMP_CLAUSE_LASTPRIVATE_FIRSTPRIVATE (c))
+ {
+ lastprivate_firstprivate = true;
+ if (pass != 0)
+ continue;
+ }
+ break;
default:
continue;
}
gcc_assert (DECL_P (ptr));
x = TYPE_SIZE_UNIT (TREE_TYPE (new_var));
- args = tree_cons (NULL, x, NULL);
- x = built_in_decls[BUILT_IN_ALLOCA];
- x = build_function_call_expr (x, args);
+ x = build_call_expr (built_in_decls[BUILT_IN_ALLOCA], 1, x);
x = fold_convert (TREE_TYPE (ptr), x);
- x = build2 (MODIFY_EXPR, void_type_node, ptr, x);
+ x = build_gimple_modify_stmt (ptr, x);
gimplify_and_add (x, ilist);
}
else if (is_reference (var))
code that expects a pointer to something that expects
a direct variable. Note that this doesn't apply to
C++, since reference types are disallowed in data
- sharing clauses there. */
+ sharing clauses there, except for NRV optimized
+ return values. */
if (pass == 0)
continue;
if (DECL_NAME (var))
name = IDENTIFIER_POINTER (DECL_NAME (new_var));
- x = create_tmp_var (TREE_TYPE (TREE_TYPE (new_var)), name);
+ x = create_tmp_var_raw (TREE_TYPE (TREE_TYPE (new_var)),
+ name);
+ gimple_add_tmp_var (x);
x = build_fold_addr_expr_with_type (x, TREE_TYPE (new_var));
}
else
{
- args = tree_cons (NULL, x, NULL);
- x = built_in_decls[BUILT_IN_ALLOCA];
- x = build_function_call_expr (x, args);
+ x = build_call_expr (built_in_decls[BUILT_IN_ALLOCA], 1, x);
x = fold_convert (TREE_TYPE (new_var), x);
}
- x = build2 (MODIFY_EXPR, void_type_node, new_var, x);
+ x = build_gimple_modify_stmt (new_var, x);
gimplify_and_add (x, ilist);
new_var = build_fold_indirect_ref (new_var);
else if (pass != 0)
continue;
- switch (TREE_CODE (c))
+ switch (OMP_CLAUSE_CODE (c))
{
case OMP_CLAUSE_SHARED:
+ /* Shared global vars are just accessed directly. */
+ if (is_global_var (new_var))
+ break;
/* Set up the DECL_VALUE_EXPR for shared variables now. This
needs to be delayed until after fixup_child_record_type so
that we get the correct type during the dereference. */
{
x = omp_reduction_init (c, TREE_TYPE (new_var));
gcc_assert (TREE_CODE (TREE_TYPE (new_var)) != ARRAY_TYPE);
- x = build2 (MODIFY_EXPR, void_type_node, new_var, x);
+ x = build_gimple_modify_stmt (new_var, x);
gimplify_and_add (x, ilist);
}
break;
but it certainly is to C++ operator=. */
if (copyin_seq)
{
- x = built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM];
- x = build_function_call_expr (x, NULL);
+ x = build_call_expr (built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM], 0);
x = build2 (NE_EXPR, boolean_type_node, x,
build_int_cst (TREE_TYPE (x), 0));
x = build3 (COND_EXPR, void_type_node, x, copyin_seq, NULL);
/* If any copyin variable is passed by reference, we must ensure the
master thread doesn't modify it before it is copied over in all
- threads. */
- if (copyin_by_ref)
- build_omp_barrier (ilist);
+ threads. Similarly for variables in both firstprivate and
+ lastprivate clauses we need to ensure the lastprivate copying
+ happens after firstprivate copying in all threads. */
+ if (copyin_by_ref || lastprivate_firstprivate)
+ gimplify_and_add (build_omp_barrier (), ilist);
}
{
tree var, new_var;
- if (TREE_CODE (c) != OMP_CLAUSE_LASTPRIVATE)
+ if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_LASTPRIVATE)
continue;
var = OMP_CLAUSE_DECL (c);
/* First see if there is exactly one reduction clause. Use OMP_ATOMIC
update in that case, otherwise use a lock. */
for (c = clauses; c && count < 2; c = OMP_CLAUSE_CHAIN (c))
- if (TREE_CODE (c) == OMP_CLAUSE_REDUCTION)
+ if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION)
{
if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
{
tree var, ref, new_var;
enum tree_code code;
- if (TREE_CODE (c) != OMP_CLAUSE_REDUCTION)
+ if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_REDUCTION)
continue;
var = OMP_CLAUSE_DECL (c);
{
x = build2 (code, TREE_TYPE (ref), ref, new_var);
ref = build_outer_var_ref (var, ctx);
- x = build2 (MODIFY_EXPR, void_type_node, ref, x);
+ x = build_gimple_modify_stmt (ref, x);
append_to_statement_list (x, &sub_list);
}
}
- x = built_in_decls[BUILT_IN_GOMP_ATOMIC_START];
- x = build_function_call_expr (x, NULL);
+ x = build_call_expr (built_in_decls[BUILT_IN_GOMP_ATOMIC_START], 0);
gimplify_and_add (x, stmt_list);
gimplify_and_add (sub_list, stmt_list);
- x = built_in_decls[BUILT_IN_GOMP_ATOMIC_END];
- x = build_function_call_expr (x, NULL);
+ x = build_call_expr (built_in_decls[BUILT_IN_GOMP_ATOMIC_END], 0);
gimplify_and_add (x, stmt_list);
}
tree var, ref, x;
bool by_ref;
- if (TREE_CODE (c) != OMP_CLAUSE_COPYPRIVATE)
+ if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYPRIVATE)
continue;
var = OMP_CLAUSE_DECL (c);
by_ref = use_pointer_for_field (var, false);
ref = build_sender_ref (var, ctx);
- x = (ctx->is_nested) ? lookup_decl_in_outer_ctx (var, ctx) : var;
+ x = lookup_decl_in_outer_ctx (var, ctx);
x = by_ref ? build_fold_addr_expr (x) : x;
- x = build2 (MODIFY_EXPR, void_type_node, ref, x);
+ x = build_gimple_modify_stmt (ref, x);
gimplify_and_add (x, slist);
ref = build_receiver_ref (var, by_ref, ctx);
tree val, ref, x, var;
bool by_ref, do_in = false, do_out = false;
- switch (TREE_CODE (c))
+ switch (OMP_CLAUSE_CODE (c))
{
case OMP_CLAUSE_FIRSTPRIVATE:
case OMP_CLAUSE_COPYIN:
continue;
}
- var = val = OMP_CLAUSE_DECL (c);
- if (ctx->is_nested)
- var = lookup_decl_in_outer_ctx (val, ctx);
+ val = OMP_CLAUSE_DECL (c);
+ var = lookup_decl_in_outer_ctx (val, ctx);
+ if (OMP_CLAUSE_CODE (c) != OMP_CLAUSE_COPYIN
+ && is_global_var (var))
+ continue;
if (is_variable_sized (val))
continue;
by_ref = use_pointer_for_field (val, false);
- switch (TREE_CODE (c))
+ switch (OMP_CLAUSE_CODE (c))
{
case OMP_CLAUSE_FIRSTPRIVATE:
case OMP_CLAUSE_COPYIN:
{
ref = build_sender_ref (val, ctx);
x = by_ref ? build_fold_addr_expr (var) : var;
- x = build2 (MODIFY_EXPR, void_type_node, ref, x);
+ x = build_gimple_modify_stmt (ref, x);
gimplify_and_add (x, ilist);
}
if (do_out)
{
ref = build_sender_ref (val, ctx);
- x = build2 (MODIFY_EXPR, void_type_node, var, ref);
+ x = build_gimple_modify_stmt (var, ref);
gimplify_and_add (x, olist);
}
}
if (!nvar || !DECL_HAS_VALUE_EXPR_P (nvar))
continue;
- var = ovar;
-
/* If CTX is a nested parallel directive. Find the immediately
enclosing parallel or workshare construct that contains a
mapping for OVAR. */
- if (ctx->is_nested)
- var = lookup_decl_in_outer_ctx (ovar, ctx);
+ var = lookup_decl_in_outer_ctx (ovar, ctx);
if (use_pointer_for_field (ovar, true))
{
x = build_sender_ref (ovar, ctx);
var = build_fold_addr_expr (var);
- x = build2 (MODIFY_EXPR, void_type_node, x, var);
+ x = build_gimple_modify_stmt (x, var);
gimplify_and_add (x, ilist);
}
else
{
x = build_sender_ref (ovar, ctx);
- x = build2 (MODIFY_EXPR, void_type_node, x, var);
+ x = build_gimple_modify_stmt (x, var);
gimplify_and_add (x, ilist);
x = build_sender_ref (ovar, ctx);
- x = build2 (MODIFY_EXPR, void_type_node, var, x);
+ x = build_gimple_modify_stmt (var, x);
gimplify_and_add (x, olist);
}
}
the workshare construct. */
static void
-expand_parallel_call (struct omp_region *region, basic_block bb, tree ws_args)
+expand_parallel_call (struct omp_region *region, basic_block bb,
+ tree entry_stmt, tree ws_args)
{
- tree t, args, val, cond, c, list, clauses;
+ tree t, t1, t2, val, cond, c, clauses;
block_stmt_iterator si;
int start_ix;
- clauses = OMP_PARALLEL_CLAUSES (region->entry);
- push_gimplify_context ();
+ clauses = OMP_PARALLEL_CLAUSES (entry_stmt);
- /* Determine what flavour of GOMP_parallel_start we will be
+ /* Determine what flavor of GOMP_parallel_start we will be
emitting. */
start_ix = BUILT_IN_GOMP_PARALLEL_START;
if (is_combined_parallel (region))
{
- tree stmt = region->inner->entry;
-
- if (TREE_CODE (stmt) == OMP_FOR)
+ switch (region->inner->type)
{
- struct omp_for_data fd;
- extract_omp_for_data (stmt, &fd);
- start_ix = BUILT_IN_GOMP_PARALLEL_LOOP_STATIC_START + fd.sched_kind;
+ case OMP_FOR:
+ start_ix = BUILT_IN_GOMP_PARALLEL_LOOP_STATIC_START
+ + region->inner->sched_kind;
+ break;
+ case OMP_SECTIONS:
+ start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS_START;
+ break;
+ default:
+ gcc_unreachable ();
}
- else if (TREE_CODE (stmt) == OMP_SECTIONS)
- start_ix = BUILT_IN_GOMP_PARALLEL_SECTIONS_START;
- else
- gcc_unreachable ();
}
/* By default, the value of NUM_THREADS is zero (selected at run time)
cond = gimple_boolify (cond);
if (integer_zerop (val))
- val = build2 (EQ_EXPR, unsigned_type_node, cond,
- build_int_cst (TREE_TYPE (cond), 0));
+ val = fold_build2 (EQ_EXPR, unsigned_type_node, cond,
+ build_int_cst (TREE_TYPE (cond), 0));
else
{
basic_block cond_bb, then_bb, else_bb;
- edge e;
- tree t, then_lab, else_lab, tmp;
+ edge e, e_then, e_else;
+ tree t, tmp_then, tmp_else, tmp_join, tmp_var;
+
+ tmp_var = create_tmp_var (TREE_TYPE (val), NULL);
+ if (gimple_in_ssa_p (cfun))
+ {
+ tmp_then = make_ssa_name (tmp_var, NULL_TREE);
+ tmp_else = make_ssa_name (tmp_var, NULL_TREE);
+ tmp_join = make_ssa_name (tmp_var, NULL_TREE);
+ }
+ else
+ {
+ tmp_then = tmp_var;
+ tmp_else = tmp_var;
+ tmp_join = tmp_var;
+ }
- tmp = create_tmp_var (TREE_TYPE (val), NULL);
e = split_block (bb, NULL);
cond_bb = e->src;
bb = e->dest;
then_bb = create_empty_bb (cond_bb);
else_bb = create_empty_bb (then_bb);
- then_lab = create_artificial_label ();
- else_lab = create_artificial_label ();
+ set_immediate_dominator (CDI_DOMINATORS, then_bb, cond_bb);
+ set_immediate_dominator (CDI_DOMINATORS, else_bb, cond_bb);
t = build3 (COND_EXPR, void_type_node,
- cond,
- build_and_jump (&then_lab),
- build_and_jump (&else_lab));
+ cond, NULL_TREE, NULL_TREE);
si = bsi_start (cond_bb);
bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
si = bsi_start (then_bb);
- t = build1 (LABEL_EXPR, void_type_node, then_lab);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
- t = build2 (MODIFY_EXPR, void_type_node, tmp, val);
+ t = build_gimple_modify_stmt (tmp_then, val);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (tmp_then) = t;
bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
si = bsi_start (else_bb);
- t = build1 (LABEL_EXPR, void_type_node, else_lab);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
- t = build2 (MODIFY_EXPR, void_type_node, tmp,
- build_int_cst (unsigned_type_node, 1));
+ t = build_gimple_modify_stmt (tmp_else,
+ build_int_cst (unsigned_type_node, 1));
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (tmp_else) = t;
bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
make_edge (cond_bb, then_bb, EDGE_TRUE_VALUE);
make_edge (cond_bb, else_bb, EDGE_FALSE_VALUE);
- make_edge (then_bb, bb, EDGE_FALLTHRU);
- make_edge (else_bb, bb, EDGE_FALLTHRU);
+ e_then = make_edge (then_bb, bb, EDGE_FALLTHRU);
+ e_else = make_edge (else_bb, bb, EDGE_FALLTHRU);
+
+ if (gimple_in_ssa_p (cfun))
+ {
+ tree phi = create_phi_node (tmp_join, bb);
+ SSA_NAME_DEF_STMT (tmp_join) = phi;
+ add_phi_arg (phi, tmp_then, e_then);
+ add_phi_arg (phi, tmp_else, e_else);
+ }
- val = tmp;
+ val = tmp_join;
}
- list = NULL_TREE;
- val = get_formal_tmp_var (val, &list);
si = bsi_start (bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
+ val = force_gimple_operand_bsi (&si, val, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
}
- list = NULL_TREE;
- args = tree_cons (NULL, val, NULL);
- t = OMP_PARALLEL_DATA_ARG (region->entry);
+ si = bsi_last (bb);
+ t = OMP_PARALLEL_DATA_ARG (entry_stmt);
if (t == NULL)
- t = null_pointer_node;
+ t1 = null_pointer_node;
else
- t = build_fold_addr_expr (t);
- args = tree_cons (NULL, t, args);
- t = build_fold_addr_expr (OMP_PARALLEL_FN (region->entry));
- args = tree_cons (NULL, t, args);
+ t1 = build_fold_addr_expr (t);
+ t2 = build_fold_addr_expr (OMP_PARALLEL_FN (entry_stmt));
if (ws_args)
- args = chainon (args, ws_args);
+ {
+ tree args = tree_cons (NULL, t2,
+ tree_cons (NULL, t1,
+ tree_cons (NULL, val, ws_args)));
+ t = build_function_call_expr (built_in_decls[start_ix], args);
+ }
+ else
+ t = build_call_expr (built_in_decls[start_ix], 3, t2, t1, val);
- t = built_in_decls[start_ix];
- t = build_function_call_expr (t, args);
- gimplify_and_add (t, &list);
+ force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
- t = OMP_PARALLEL_DATA_ARG (region->entry);
+ t = OMP_PARALLEL_DATA_ARG (entry_stmt);
if (t == NULL)
t = null_pointer_node;
else
t = build_fold_addr_expr (t);
- args = tree_cons (NULL, t, NULL);
- t = build_function_call_expr (OMP_PARALLEL_FN (region->entry), args);
- gimplify_and_add (t, &list);
+ t = build_call_expr (OMP_PARALLEL_FN (entry_stmt), 1, t);
+ force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
- t = built_in_decls[BUILT_IN_GOMP_PARALLEL_END];
- t = build_function_call_expr (t, NULL);
- gimplify_and_add (t, &list);
-
- si = bsi_last (bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
-
- pop_gimplify_context (NULL_TREE);
+ t = build_call_expr (built_in_decls[BUILT_IN_GOMP_PARALLEL_END], 0);
+ force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
}
if (lang_protect_cleanup_actions)
t = lang_protect_cleanup_actions ();
else
- {
- t = built_in_decls[BUILT_IN_TRAP];
- t = build_function_call_expr (t, NULL);
- }
+ t = build_call_expr (built_in_decls[BUILT_IN_TRAP], 0);
f = build2 (EH_FILTER_EXPR, void_type_node, NULL, NULL);
EH_FILTER_MUST_NOT_THROW (f) = 1;
gimplify_and_add (t, &EH_FILTER_FAILURE (f));
{
block_stmt_iterator si;
basic_block exit_bb;
+ edge_iterator ei;
+ edge e;
tree t;
- gcc_assert (TREE_CODE (region->entry) == OMP_PARALLEL);
+ exit_bb = region->exit;
- exit_bb = bb_for_stmt (region->exit);
+ /* If the parallel region doesn't return, we don't have REGION->EXIT
+ block at all. */
+ if (! exit_bb)
+ return;
- /* The barrier should be immediately before OMP_RETURN_EXPR.
- Otherwise, we cannot remove it. */
+ /* The last insn in the block will be the parallel's OMP_RETURN. The
+ workshare's OMP_RETURN will be in a preceding block. The kinds of
+ statements that can appear in between are extremely limited -- no
+ memory operations at all. Here, we allow nothing at all, so the
+ only thing we allow to precede this OMP_RETURN is a label. */
si = bsi_last (exit_bb);
- t = bsi_stmt (si);
- gcc_assert (TREE_CODE (t) == OMP_RETURN_EXPR);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_RETURN);
bsi_prev (&si);
- if (bsi_end_p (si))
+ if (!bsi_end_p (si) && TREE_CODE (bsi_stmt (si)) != LABEL_EXPR)
return;
- t = bsi_stmt (si);
- if (TREE_CODE (t) == CALL_EXPR
- && get_callee_fndecl (t) == built_in_decls[BUILT_IN_GOMP_BARRIER])
- bsi_remove (&si, true);
+ FOR_EACH_EDGE (e, ei, exit_bb->preds)
+ {
+ si = bsi_last (e->src);
+ if (bsi_end_p (si))
+ continue;
+ t = bsi_stmt (si);
+ if (TREE_CODE (t) == OMP_RETURN)
+ OMP_RETURN_NOWAIT (t) = 1;
+ }
+}
+
+static void
+remove_exit_barriers (struct omp_region *region)
+{
+ if (region->type == OMP_PARALLEL)
+ remove_exit_barrier (region);
+
+ if (region->inner)
+ {
+ region = region->inner;
+ remove_exit_barriers (region);
+ while (region->next)
+ {
+ region = region->next;
+ remove_exit_barriers (region);
+ }
+ }
}
+/* Optimize omp_get_thread_num () and omp_get_num_threads ()
+ calls. These can't be declared as const functions, but
+ within one parallel body they are constant, so they can be
+ transformed there into __builtin_omp_get_{thread_num,num_threads} ()
+ which are declared const. */
+
+static void
+optimize_omp_library_calls (void)
+{
+ basic_block bb;
+ block_stmt_iterator bsi;
+ tree thr_num_id
+ = DECL_ASSEMBLER_NAME (built_in_decls [BUILT_IN_OMP_GET_THREAD_NUM]);
+ tree num_thr_id
+ = DECL_ASSEMBLER_NAME (built_in_decls [BUILT_IN_OMP_GET_NUM_THREADS]);
+
+ FOR_EACH_BB (bb)
+ for (bsi = bsi_start (bb); !bsi_end_p (bsi); bsi_next (&bsi))
+ {
+ tree stmt = bsi_stmt (bsi);
+ tree call = get_call_expr_in (stmt);
+ tree decl;
+
+ if (call
+ && (decl = get_callee_fndecl (call))
+ && DECL_EXTERNAL (decl)
+ && TREE_PUBLIC (decl)
+ && DECL_INITIAL (decl) == NULL)
+ {
+ tree built_in;
+
+ if (DECL_NAME (decl) == thr_num_id)
+ built_in = built_in_decls [BUILT_IN_OMP_GET_THREAD_NUM];
+ else if (DECL_NAME (decl) == num_thr_id)
+ built_in = built_in_decls [BUILT_IN_OMP_GET_NUM_THREADS];
+ else
+ continue;
+
+ if (DECL_ASSEMBLER_NAME (decl) != DECL_ASSEMBLER_NAME (built_in)
+ || call_expr_nargs (call) != 0)
+ continue;
+
+ if (flag_exceptions && !TREE_NOTHROW (decl))
+ continue;
+
+ if (TREE_CODE (TREE_TYPE (decl)) != FUNCTION_TYPE
+ || TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (decl)))
+ != TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (built_in))))
+ continue;
+
+ CALL_EXPR_FN (call) = build_fold_addr_expr (built_in);
+ }
+ }
+}
/* Expand the OpenMP parallel directive starting at REGION. */
expand_omp_parallel (struct omp_region *region)
{
basic_block entry_bb, exit_bb, new_bb;
- struct function *child_cfun, *saved_cfun;
+ struct function *child_cfun;
tree child_fn, block, t, ws_args;
block_stmt_iterator si;
+ tree entry_stmt;
edge e;
- child_fn = OMP_PARALLEL_FN (region->entry);
+ entry_stmt = last_stmt (region->entry);
+ child_fn = OMP_PARALLEL_FN (entry_stmt);
child_cfun = DECL_STRUCT_FUNCTION (child_fn);
- saved_cfun = cfun;
+ /* If this function has been already instrumented, make sure
+ the child function isn't instrumented again. */
+ child_cfun->after_tree_profile = cfun->after_tree_profile;
- entry_bb = bb_for_stmt (region->entry);
- exit_bb = bb_for_stmt (region->exit);
-
- /* Barriers at the end of the function are not necessary and can be
- removed. Since the caller will have a barrier of its own, this
- one is superfluous. */
- remove_exit_barrier (region);
+ entry_bb = region->entry;
+ exit_bb = region->exit;
if (is_combined_parallel (region))
ws_args = region->ws_args;
else
ws_args = NULL_TREE;
- if (DECL_STRUCT_FUNCTION (OMP_PARALLEL_FN (region->entry))->cfg)
+ if (child_cfun->cfg)
{
/* Due to inlining, it may happen that we have already outlined
the region, in which case all we need to do is make the
block_stmt_iterator si;
entry_succ_e = single_succ_edge (entry_bb);
- exit_succ_e = single_succ_edge (exit_bb);
si = bsi_last (entry_bb);
- gcc_assert (!bsi_end_p (si) && TREE_CODE (bsi_stmt (si)) == OMP_PARALLEL);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_PARALLEL);
bsi_remove (&si, true);
new_bb = entry_bb;
- remove_edge (entry_succ_e);
- make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU);
+ if (exit_bb)
+ {
+ exit_succ_e = single_succ_edge (exit_bb);
+ make_edge (new_bb, exit_succ_e->dest, EDGE_FALLTHRU);
+ }
+ remove_edge_and_dominated_blocks (entry_succ_e);
}
else
{
/* If the parallel region needs data sent from the parent
- function, then the very first statement of the parallel body
+ function, then the very first statement (except possible
+ tree profile counter updates) of the parallel body
is a copy assignment .OMP_DATA_I = &.OMP_DATA_O. Since
&.OMP_DATA_O is passed as an argument to the child function,
we need to replace it with the argument as seen by the child
a function call that has been inlined, the original PARM_DECL
.OMP_DATA_I may have been converted into a different local
variable. In which case, we need to keep the assignment. */
- if (OMP_PARALLEL_DATA_ARG (region->entry))
+ if (OMP_PARALLEL_DATA_ARG (entry_stmt))
{
basic_block entry_succ_bb = single_succ (entry_bb);
- block_stmt_iterator si = bsi_start (entry_succ_bb);
- tree stmt;
+ block_stmt_iterator si;
+ tree parcopy_stmt = NULL_TREE, arg, narg;
+
+ for (si = bsi_start (entry_succ_bb); ; bsi_next (&si))
+ {
+ tree stmt, arg;
+
+ gcc_assert (!bsi_end_p (si));
+ stmt = bsi_stmt (si);
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
+ continue;
- gcc_assert (!bsi_end_p (si));
+ arg = GIMPLE_STMT_OPERAND (stmt, 1);
+ STRIP_NOPS (arg);
+ if (TREE_CODE (arg) == ADDR_EXPR
+ && TREE_OPERAND (arg, 0)
+ == OMP_PARALLEL_DATA_ARG (entry_stmt))
+ {
+ parcopy_stmt = stmt;
+ break;
+ }
+ }
- stmt = bsi_stmt (si);
- gcc_assert (TREE_CODE (stmt) == MODIFY_EXPR
- && TREE_CODE (TREE_OPERAND (stmt, 1)) == ADDR_EXPR
- && TREE_OPERAND (TREE_OPERAND (stmt, 1), 0)
- == OMP_PARALLEL_DATA_ARG (region->entry));
+ gcc_assert (parcopy_stmt != NULL_TREE);
+ arg = DECL_ARGUMENTS (child_fn);
- if (TREE_OPERAND (stmt, 0) == DECL_ARGUMENTS (child_fn))
- bsi_remove (&si, true);
+ if (!gimple_in_ssa_p (cfun))
+ {
+ if (GIMPLE_STMT_OPERAND (parcopy_stmt, 0) == arg)
+ bsi_remove (&si, true);
+ else
+ GIMPLE_STMT_OPERAND (parcopy_stmt, 1) = arg;
+ }
else
- TREE_OPERAND (stmt, 1) = DECL_ARGUMENTS (child_fn);
+ {
+ /* If we are in ssa form, we must load the value from the default
+ definition of the argument. That should not be defined now,
+ since the argument is not used uninitialized. */
+ gcc_assert (gimple_default_def (cfun, arg) == NULL);
+ narg = make_ssa_name (arg, build_empty_stmt ());
+ set_default_def (arg, narg);
+ GIMPLE_STMT_OPERAND (parcopy_stmt, 1) = narg;
+ update_stmt (parcopy_stmt);
+ }
}
/* Declare local variables needed in CHILD_CFUN. */
block = DECL_INITIAL (child_fn);
BLOCK_VARS (block) = list2chain (child_cfun->unexpanded_var_list);
- DECL_SAVED_TREE (child_fn) = single_succ (entry_bb)->stmt_list;
-
- /* Reset DECL_CONTEXT on locals and function arguments. */
- for (t = BLOCK_VARS (block); t; t = TREE_CHAIN (t))
- DECL_CONTEXT (t) = child_fn;
+ DECL_SAVED_TREE (child_fn) = bb_stmt_list (single_succ (entry_bb));
+ /* Reset DECL_CONTEXT on function arguments. */
for (t = DECL_ARGUMENTS (child_fn); t; t = TREE_CHAIN (t))
DECL_CONTEXT (t) = child_fn;
entry_bb = e->dest;
single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU;
- /* Move the parallel region into CHILD_CFUN. We need to reset
- dominance information because the expansion of the inner
- regions has invalidated it. */
- free_dominance_info (CDI_DOMINATORS);
- new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb);
- single_succ_edge (new_bb)->flags = EDGE_FALLTHRU;
- cgraph_add_new_function (child_fn);
-
/* Convert OMP_RETURN into a RETURN_EXPR. */
- si = bsi_last (exit_bb);
- gcc_assert (!bsi_end_p (si)
- && TREE_CODE (bsi_stmt (si)) == OMP_RETURN_EXPR);
- t = build1 (RETURN_EXPR, void_type_node, NULL);
- bsi_insert_after (&si, t, TSI_SAME_STMT);
- bsi_remove (&si, true);
- }
+ if (exit_bb)
+ {
+ si = bsi_last (exit_bb);
+ gcc_assert (!bsi_end_p (si)
+ && TREE_CODE (bsi_stmt (si)) == OMP_RETURN);
+ t = build1 (RETURN_EXPR, void_type_node, NULL);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+ bsi_remove (&si, true);
+ }
+ /* Move the parallel region into CHILD_CFUN. */
+
+ if (gimple_in_ssa_p (cfun))
+ {
+ push_cfun (child_cfun);
+ init_tree_ssa ();
+ init_ssa_operands ();
+ cfun->gimple_df->in_ssa_p = true;
+ pop_cfun ();
+ }
+ new_bb = move_sese_region_to_fn (child_cfun, entry_bb, exit_bb);
+ if (exit_bb)
+ single_succ_edge (new_bb)->flags = EDGE_FALLTHRU;
+
+ /* Inform the callgraph about the new function. */
+ DECL_STRUCT_FUNCTION (child_fn)->curr_properties
+ = cfun->curr_properties;
+ cgraph_add_new_function (child_fn, true);
+
+ /* Fix the callgraph edges for child_cfun. Those for cfun will be
+ fixed in a following pass. */
+ push_cfun (child_cfun);
+ if (optimize)
+ optimize_omp_library_calls ();
+ rebuild_cgraph_edges ();
+
+ /* Some EH regions might become dead, see PR34608. If
+ pass_cleanup_cfg isn't the first pass to happen with the
+ new child, these dead EH edges might cause problems.
+ Clean them up now. */
+ if (flag_exceptions)
+ {
+ basic_block bb;
+ tree save_current = current_function_decl;
+ bool changed = false;
+
+ current_function_decl = child_fn;
+ FOR_EACH_BB (bb)
+ changed |= tree_purge_dead_eh_edges (bb);
+ if (changed)
+ cleanup_tree_cfg ();
+ current_function_decl = save_current;
+ }
+ pop_cfun ();
+ }
+
/* Emit a library call to launch the children threads. */
- expand_parallel_call (region, new_bb, ws_args);
+ expand_parallel_call (region, new_bb, entry_stmt, ws_args);
+ update_ssa (TODO_update_ssa_only_virtuals);
}
L3:
If this is a combined omp parallel loop, instead of the call to
- GOMP_loop_foo_start, we emit 'goto L3'. */
+ GOMP_loop_foo_start, we call GOMP_loop_foo_next. */
-static basic_block
+static void
expand_omp_for_generic (struct omp_region *region,
struct omp_for_data *fd,
enum built_in_function start_fn,
enum built_in_function next_fn)
{
- tree l0, l1, l2, l3;
- tree type, istart0, iend0, iend;
- tree t, args, list;
- basic_block entry_bb, exit_bb, l0_bb, l1_bb, l2_bb;
- edge exit_edge;
+ tree type, istart0, iend0, iend, phi;
+ tree t, vmain, vback;
+ basic_block entry_bb, cont_bb, exit_bb, l0_bb, l1_bb;
+ basic_block l2_bb = NULL, l3_bb = NULL;
block_stmt_iterator si;
bool in_combined_parallel = is_combined_parallel (region);
+ bool broken_loop = region->cont == NULL;
+ edge e, ne;
+
+ gcc_assert (!broken_loop || !in_combined_parallel);
type = TREE_TYPE (fd->v);
istart0 = create_tmp_var (long_integer_type_node, ".istart0");
iend0 = create_tmp_var (long_integer_type_node, ".iend0");
+ TREE_ADDRESSABLE (istart0) = 1;
+ TREE_ADDRESSABLE (iend0) = 1;
+ if (gimple_in_ssa_p (cfun))
+ {
+ add_referenced_var (istart0);
+ add_referenced_var (iend0);
+ }
- l0 = create_artificial_label ();
- l1 = create_artificial_label ();
- l2 = create_artificial_label ();
- l3 = create_artificial_label ();
- iend = create_tmp_var (type, NULL);
-
- entry_bb = bb_for_stmt (region->entry);
- l1_bb = single_succ (entry_bb);
- exit_bb = bb_for_stmt (region->exit);
+ entry_bb = region->entry;
+ cont_bb = region->cont;
+ gcc_assert (EDGE_COUNT (entry_bb->succs) == 2);
+ gcc_assert (broken_loop
+ || BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest);
+ l0_bb = split_edge (FALLTHRU_EDGE (entry_bb));
+ l1_bb = single_succ (l0_bb);
+ if (!broken_loop)
+ {
+ l2_bb = create_empty_bb (cont_bb);
+ gcc_assert (BRANCH_EDGE (cont_bb)->dest == l1_bb);
+ gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
+ }
+ else
+ l2_bb = NULL;
+ l3_bb = BRANCH_EDGE (entry_bb)->dest;
+ exit_bb = region->exit;
si = bsi_last (entry_bb);
- gcc_assert (bsi_stmt (si) && TREE_CODE (bsi_stmt (si)) == OMP_FOR);
- bsi_remove (&si, true);
- list = alloc_stmt_list ();
-
- if (!in_combined_parallel)
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_FOR);
+ if (in_combined_parallel)
+ {
+ /* In a combined parallel loop, emit a call to
+ GOMP_loop_foo_next. */
+ t = build_call_expr (built_in_decls[next_fn], 2,
+ build_fold_addr_expr (istart0),
+ build_fold_addr_expr (iend0));
+ }
+ else
{
+ tree t0, t1, t2, t3, t4;
/* If this is not a combined parallel loop, emit a call to
GOMP_loop_foo_start in ENTRY_BB. */
- t = build_fold_addr_expr (iend0);
- args = tree_cons (NULL, t, NULL);
- t = build_fold_addr_expr (istart0);
- args = tree_cons (NULL, t, args);
+ t4 = build_fold_addr_expr (iend0);
+ t3 = build_fold_addr_expr (istart0);
+ t2 = fold_convert (long_integer_type_node, fd->step);
+ t1 = fold_convert (long_integer_type_node, fd->n2);
+ t0 = fold_convert (long_integer_type_node, fd->n1);
if (fd->chunk_size)
{
t = fold_convert (long_integer_type_node, fd->chunk_size);
- args = tree_cons (NULL, t, args);
+ t = build_call_expr (built_in_decls[start_fn], 6,
+ t0, t1, t2, t, t3, t4);
}
- t = fold_convert (long_integer_type_node, fd->step);
- args = tree_cons (NULL, t, args);
- t = fold_convert (long_integer_type_node, fd->n2);
- args = tree_cons (NULL, t, args);
- t = fold_convert (long_integer_type_node, fd->n1);
- args = tree_cons (NULL, t, args);
- t = build_function_call_expr (built_in_decls[start_fn], args);
- t = get_formal_tmp_var (t, &list);
- t = build3 (COND_EXPR, void_type_node, t, build_and_jump (&l0),
- build_and_jump (&l3));
- append_to_statement_list (t, &list);
- si = bsi_last (entry_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
+ else
+ t = build_call_expr (built_in_decls[start_fn], 5,
+ t0, t1, t2, t3, t4);
+ }
+ t = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+
+ /* V may be used outside of the loop (e.g., to handle lastprivate clause).
+ If this is the case, its value is undefined if the loop is not entered
+ at all. To handle this case, set its initial value to N1. */
+ if (gimple_in_ssa_p (cfun))
+ {
+ e = find_edge (entry_bb, l3_bb);
+ for (phi = phi_nodes (l3_bb); phi; phi = PHI_CHAIN (phi))
+ if (PHI_ARG_DEF_FROM_EDGE (phi, e) == fd->v)
+ SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), fd->n1);
+ }
+ else
+ {
+ t = build_gimple_modify_stmt (fd->v, fd->n1);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
}
- /* Iteration setup for sequential loop goes in L0_BB. */
- list = alloc_stmt_list ();
- t = build1 (LABEL_EXPR, void_type_node, l0);
- gimplify_and_add (t, &list);
+ /* Remove the OMP_FOR statement. */
+ bsi_remove (&si, true);
+ /* Iteration setup for sequential loop goes in L0_BB. */
+ si = bsi_start (l0_bb);
t = fold_convert (type, istart0);
- t = build2 (MODIFY_EXPR, void_type_node, fd->v, t);
- gimplify_and_add (t, &list);
+ t = force_gimple_operand_bsi (&si, t, false, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
+ t = build_gimple_modify_stmt (fd->v, t);
+ bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (fd->v) = t;
t = fold_convert (type, iend0);
- t = build2 (MODIFY_EXPR, void_type_node, iend, t);
- gimplify_and_add (t, &list);
+ iend = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
- l0_bb = create_empty_bb (entry_bb);
- si = bsi_start (l0_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
-
- /* Loop body goes in L1_BB. */
- list = alloc_stmt_list ();
- si = bsi_start (l1_bb);
- bsi_insert_before (&si, build1 (LABEL_EXPR, void_type_node, l1),
- BSI_CONTINUE_LINKING);
-
- /* Code to control the increment and predicate for the sequential
- loop goes in the first half of EXIT_BB (we split EXIT_BB so
- that we can inherit all the edges going out of the loop
- body). */
- list = alloc_stmt_list ();
-
- t = build2 (PLUS_EXPR, type, fd->v, fd->step);
- t = build2 (MODIFY_EXPR, void_type_node, fd->v, t);
- gimplify_and_add (t, &list);
+ if (!broken_loop)
+ {
+ /* Code to control the increment and predicate for the sequential
+ loop goes in the CONT_BB. */
+ si = bsi_last (cont_bb);
+ t = bsi_stmt (si);
+ gcc_assert (TREE_CODE (t) == OMP_CONTINUE);
+ vmain = TREE_OPERAND (t, 1);
+ vback = TREE_OPERAND (t, 0);
+
+ t = fold_build2 (PLUS_EXPR, type, vmain, fd->step);
+ t = force_gimple_operand_bsi (&si, t, false, NULL_TREE,
+ true, BSI_SAME_STMT);
+ t = build_gimple_modify_stmt (vback, t);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (vback) = t;
- t = build2 (fd->cond_code, boolean_type_node, fd->v, iend);
- t = get_formal_tmp_var (t, &list);
- t = build3 (COND_EXPR, void_type_node, t, build_and_jump (&l1),
- build_and_jump (&l2));
- append_to_statement_list (t, &list);
+ t = build2 (fd->cond_code, boolean_type_node, vback, iend);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
- si = bsi_last (exit_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_RETURN_EXPR);
- bsi_remove (&si, true);
- exit_edge = split_block (exit_bb, t);
- exit_edge->flags = EDGE_FALSE_VALUE;
+ /* Remove OMP_CONTINUE. */
+ bsi_remove (&si, true);
- si = bsi_last (exit_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
+ /* Emit code to get the next parallel iteration in L2_BB. */
+ si = bsi_start (l2_bb);
- /* Emit code to get the next parallel iteration in L2_BB. */
- list = alloc_stmt_list ();
- t = build1 (LABEL_EXPR, void_type_node, l2);
- gimplify_and_add (t, &list);
-
- t = build_fold_addr_expr (iend0);
- args = tree_cons (NULL, t, NULL);
- t = build_fold_addr_expr (istart0);
- args = tree_cons (NULL, t, args);
- t = build_function_call_expr (built_in_decls[next_fn], args);
- t = get_formal_tmp_var (t, &list);
- t = build3 (COND_EXPR, void_type_node, t, build_and_jump (&l0),
- build_and_jump (&l3));
- append_to_statement_list (t, &list);
-
- l2_bb = exit_edge->dest;
- si = bsi_start (l2_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
+ t = build_call_expr (built_in_decls[next_fn], 2,
+ build_fold_addr_expr (istart0),
+ build_fold_addr_expr (iend0));
+ t = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ }
- /* Insert exit label on EXIT_EDGE. */
- exit_edge = single_succ_edge (l2_bb);
- t = build1 (LABEL_EXPR, void_type_node, l3);
- bsi_insert_on_edge_immediate (exit_edge, t);
- exit_edge->flags = EDGE_FALSE_VALUE;
+ /* Add the loop cleanup function. */
+ si = bsi_last (exit_bb);
+ if (OMP_RETURN_NOWAIT (bsi_stmt (si)))
+ t = built_in_decls[BUILT_IN_GOMP_LOOP_END_NOWAIT];
+ else
+ t = built_in_decls[BUILT_IN_GOMP_LOOP_END];
+ t = build_call_expr (t, 0);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+ bsi_remove (&si, true);
/* Connect the new blocks. */
- remove_edge (single_succ_edge (entry_bb));
- if (in_combined_parallel)
- make_edge (entry_bb, l2_bb, EDGE_FALLTHRU);
- else
+ find_edge (entry_bb, l0_bb)->flags = EDGE_TRUE_VALUE;
+ find_edge (entry_bb, l3_bb)->flags = EDGE_FALSE_VALUE;
+
+ if (!broken_loop)
{
- make_edge (entry_bb, l0_bb, EDGE_TRUE_VALUE);
- make_edge (entry_bb, exit_edge->dest, EDGE_FALSE_VALUE);
+ e = find_edge (cont_bb, l3_bb);
+ ne = make_edge (l2_bb, l3_bb, EDGE_FALSE_VALUE);
+
+ for (phi = phi_nodes (l3_bb); phi; phi = PHI_CHAIN (phi))
+ SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, ne),
+ PHI_ARG_DEF_FROM_EDGE (phi, e));
+ remove_edge (e);
+
+ find_edge (cont_bb, l1_bb)->flags = EDGE_TRUE_VALUE;
+ make_edge (cont_bb, l2_bb, EDGE_FALSE_VALUE);
+ make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE);
+
+ set_immediate_dominator (CDI_DOMINATORS, l2_bb,
+ recompute_dominator (CDI_DOMINATORS, l2_bb));
+ set_immediate_dominator (CDI_DOMINATORS, l3_bb,
+ recompute_dominator (CDI_DOMINATORS, l3_bb));
+ set_immediate_dominator (CDI_DOMINATORS, l0_bb,
+ recompute_dominator (CDI_DOMINATORS, l0_bb));
+ set_immediate_dominator (CDI_DOMINATORS, l1_bb,
+ recompute_dominator (CDI_DOMINATORS, l1_bb));
}
-
- make_edge (l0_bb, l1_bb, EDGE_FALLTHRU);
- make_edge (exit_bb, l1_bb, EDGE_TRUE_VALUE);
- make_edge (l2_bb, l0_bb, EDGE_TRUE_VALUE);
-
- return exit_edge->dest;
}
q += (q * nthreads != n);
s0 = q * threadid;
e0 = min(s0 + q, n);
+ V = s0 * STEP + N1;
if (s0 >= e0) goto L2; else goto L0;
L0:
- V = s0 * STEP + N1;
e = e0 * STEP + N1;
L1:
BODY;
L2:
*/
-static basic_block
+static void
expand_omp_for_static_nochunk (struct omp_region *region,
struct omp_for_data *fd)
{
- tree l0, l1, l2, n, q, s0, e0, e, t, nthreads, threadid;
- tree type, utype, list;
- basic_block entry_bb, exit_bb, seq_start_bb, body_bb, new_exit_bb;
+ tree n, q, s0, e0, e, t, nthreads, threadid;
+ tree type, vmain, vback;
+ basic_block entry_bb, exit_bb, seq_start_bb, body_bb, cont_bb;
+ basic_block fin_bb;
block_stmt_iterator si;
- edge exit_edge;
- l0 = create_artificial_label ();
- l1 = create_artificial_label ();
- l2 = create_artificial_label ();
-
type = TREE_TYPE (fd->v);
- utype = lang_hooks.types.unsigned_type (type);
- entry_bb = bb_for_stmt (region->entry);
- body_bb = single_succ (entry_bb);
- exit_bb = bb_for_stmt (region->exit);
+ entry_bb = region->entry;
+ cont_bb = region->cont;
+ gcc_assert (EDGE_COUNT (entry_bb->succs) == 2);
+ gcc_assert (BRANCH_EDGE (entry_bb)->dest == FALLTHRU_EDGE (cont_bb)->dest);
+ seq_start_bb = split_edge (FALLTHRU_EDGE (entry_bb));
+ body_bb = single_succ (seq_start_bb);
+ gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb);
+ gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
+ fin_bb = FALLTHRU_EDGE (cont_bb)->dest;
+ exit_bb = region->exit;
/* Iteration space partitioning goes in ENTRY_BB. */
- list = alloc_stmt_list ();
- t = built_in_decls[BUILT_IN_OMP_GET_NUM_THREADS];
- t = build_function_call_expr (t, NULL);
- t = fold_convert (utype, t);
- nthreads = get_formal_tmp_var (t, &list);
-
- t = built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM];
- t = build_function_call_expr (t, NULL);
- t = fold_convert (utype, t);
- threadid = get_formal_tmp_var (t, &list);
-
- fd->n1 = fold_convert (type, fd->n1);
- if (!is_gimple_val (fd->n1))
- fd->n1 = get_formal_tmp_var (fd->n1, &list);
-
- fd->n2 = fold_convert (type, fd->n2);
- if (!is_gimple_val (fd->n2))
- fd->n2 = get_formal_tmp_var (fd->n2, &list);
+ si = bsi_last (entry_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_FOR);
- fd->step = fold_convert (type, fd->step);
- if (!is_gimple_val (fd->step))
- fd->step = get_formal_tmp_var (fd->step, &list);
+ t = build_call_expr (built_in_decls[BUILT_IN_OMP_GET_NUM_THREADS], 0);
+ t = fold_convert (type, t);
+ nthreads = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ t = build_call_expr (built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM], 0);
+ t = fold_convert (type, t);
+ threadid = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ fd->n1 = force_gimple_operand_bsi (&si,
+ fold_convert (type, fd->n1),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ fd->n2 = force_gimple_operand_bsi (&si,
+ fold_convert (type, fd->n2),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ fd->step = force_gimple_operand_bsi (&si,
+ fold_convert (type, fd->step),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
t = build_int_cst (type, (fd->cond_code == LT_EXPR ? -1 : 1));
t = fold_build2 (PLUS_EXPR, type, fd->step, t);
t = fold_build2 (PLUS_EXPR, type, t, fd->n2);
t = fold_build2 (MINUS_EXPR, type, t, fd->n1);
t = fold_build2 (TRUNC_DIV_EXPR, type, t, fd->step);
- t = fold_convert (utype, t);
- if (is_gimple_val (t))
- n = t;
- else
- n = get_formal_tmp_var (t, &list);
+ t = fold_convert (type, t);
+ n = force_gimple_operand_bsi (&si, t, true, NULL_TREE, true, BSI_SAME_STMT);
- t = build2 (TRUNC_DIV_EXPR, utype, n, nthreads);
- q = get_formal_tmp_var (t, &list);
+ t = fold_build2 (TRUNC_DIV_EXPR, type, n, nthreads);
+ q = force_gimple_operand_bsi (&si, t, true, NULL_TREE, true, BSI_SAME_STMT);
- t = build2 (MULT_EXPR, utype, q, nthreads);
- t = build2 (NE_EXPR, utype, t, n);
- t = build2 (PLUS_EXPR, utype, q, t);
- q = get_formal_tmp_var (t, &list);
+ t = fold_build2 (MULT_EXPR, type, q, nthreads);
+ t = fold_build2 (NE_EXPR, type, t, n);
+ t = fold_build2 (PLUS_EXPR, type, q, t);
+ q = force_gimple_operand_bsi (&si, t, true, NULL_TREE, true, BSI_SAME_STMT);
- t = build2 (MULT_EXPR, utype, q, threadid);
- s0 = get_formal_tmp_var (t, &list);
+ t = build2 (MULT_EXPR, type, q, threadid);
+ s0 = force_gimple_operand_bsi (&si, t, true, NULL_TREE, true, BSI_SAME_STMT);
- t = build2 (PLUS_EXPR, utype, s0, q);
- t = build2 (MIN_EXPR, utype, t, n);
- e0 = get_formal_tmp_var (t, &list);
+ t = fold_build2 (PLUS_EXPR, type, s0, q);
+ t = fold_build2 (MIN_EXPR, type, t, n);
+ e0 = force_gimple_operand_bsi (&si, t, true, NULL_TREE, true, BSI_SAME_STMT);
+
+ t = fold_convert (type, s0);
+ t = fold_build2 (MULT_EXPR, type, t, fd->step);
+ t = fold_build2 (PLUS_EXPR, type, t, fd->n1);
+ t = force_gimple_operand_bsi (&si, t, false, NULL_TREE,
+ true, BSI_SAME_STMT);
+ t = build_gimple_modify_stmt (fd->v, t);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (fd->v) = t;
t = build2 (GE_EXPR, boolean_type_node, s0, e0);
- t = build3 (COND_EXPR, void_type_node, t, build_and_jump (&l2),
- build_and_jump (&l0));
- append_to_statement_list (t, &list);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
- si = bsi_last (entry_bb);
- gcc_assert (bsi_stmt (si) && TREE_CODE (bsi_stmt (si)) == OMP_FOR);
+ /* Remove the OMP_FOR statement. */
bsi_remove (&si, true);
- si = bsi_last (entry_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
/* Setup code for sequential iteration goes in SEQ_START_BB. */
- list = alloc_stmt_list ();
-
- t = build1 (LABEL_EXPR, void_type_node, l0);
- gimplify_and_add (t, &list);
-
- t = fold_convert (type, s0);
- t = build2 (MULT_EXPR, type, t, fd->step);
- t = build2 (PLUS_EXPR, type, t, fd->n1);
- t = build2 (MODIFY_EXPR, void_type_node, fd->v, t);
- gimplify_and_add (t, &list);
-
- t = fold_convert (type, e0);
- t = build2 (MULT_EXPR, type, t, fd->step);
- t = build2 (PLUS_EXPR, type, t, fd->n1);
- e = get_formal_tmp_var (t, &list);
-
- seq_start_bb = create_empty_bb (entry_bb);
si = bsi_start (seq_start_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
- /* Original body goes in BODY_BB. */
- si = bsi_start (body_bb);
- t = build1 (LABEL_EXPR, void_type_node, l1);
- bsi_insert_before (&si, t, BSI_CONTINUE_LINKING);
+ t = fold_convert (type, e0);
+ t = fold_build2 (MULT_EXPR, type, t, fd->step);
+ t = fold_build2 (PLUS_EXPR, type, t, fd->n1);
+ e = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
- /* Split EXIT_BB at the OMP_RETURN. The code controlling the
- sequential loop goes in the original EXIT_BB. The exit out of
- the parallel loop goes in the new block (NEW_EXIT_BB). */
- si = bsi_last (exit_bb);
+ /* The code controlling the sequential loop replaces the OMP_CONTINUE. */
+ si = bsi_last (cont_bb);
t = bsi_stmt (si);
+ gcc_assert (TREE_CODE (t) == OMP_CONTINUE);
+ vmain = TREE_OPERAND (t, 1);
+ vback = TREE_OPERAND (t, 0);
+
+ t = fold_build2 (PLUS_EXPR, type, vmain, fd->step);
+ t = force_gimple_operand_bsi (&si, t, false, NULL_TREE,
+ true, BSI_SAME_STMT);
+ t = build_gimple_modify_stmt (vback, t);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (vback) = t;
+
+ t = build2 (fd->cond_code, boolean_type_node, vback, e);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
+
+ /* Remove the OMP_CONTINUE statement. */
bsi_remove (&si, true);
- gcc_assert (t && TREE_CODE (t) == OMP_RETURN_EXPR);
- exit_edge = split_block (exit_bb, t);
- new_exit_bb = exit_edge->dest;
- list = alloc_stmt_list ();
-
- t = build2 (PLUS_EXPR, type, fd->v, fd->step);
- t = build2 (MODIFY_EXPR, void_type_node, fd->v, t);
- gimplify_and_add (t, &list);
-
- t = build2 (fd->cond_code, boolean_type_node, fd->v, e);
- t = get_formal_tmp_var (t, &list);
- t = build3 (COND_EXPR, void_type_node, t, build_and_jump (&l1),
- build_and_jump (&l2));
- append_to_statement_list (t, &list);
+ /* Replace the OMP_RETURN with a barrier, or nothing. */
si = bsi_last (exit_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
-
- /* Add the exit label to NEW_EXIT_BB. */
- si = bsi_start (new_exit_bb);
- t = build1 (LABEL_EXPR, void_type_node, l2);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
- single_succ_edge (new_exit_bb)->flags = EDGE_FALLTHRU;
+ if (!OMP_RETURN_NOWAIT (bsi_stmt (si)))
+ force_gimple_operand_bsi (&si, build_omp_barrier (), false, NULL_TREE,
+ false, BSI_SAME_STMT);
+ bsi_remove (&si, true);
/* Connect all the blocks. */
- make_edge (seq_start_bb, body_bb, EDGE_FALLTHRU);
-
- remove_edge (single_succ_edge (entry_bb));
- make_edge (entry_bb, new_exit_bb, EDGE_TRUE_VALUE);
- make_edge (entry_bb, seq_start_bb, EDGE_FALSE_VALUE);
-
- make_edge (exit_bb, body_bb, EDGE_TRUE_VALUE);
- find_edge (exit_bb, new_exit_bb)->flags = EDGE_FALSE_VALUE;
-
- return new_exit_bb;
+ find_edge (entry_bb, seq_start_bb)->flags = EDGE_FALSE_VALUE;
+ find_edge (entry_bb, fin_bb)->flags = EDGE_TRUE_VALUE;
+
+ find_edge (cont_bb, body_bb)->flags = EDGE_TRUE_VALUE;
+ find_edge (cont_bb, fin_bb)->flags = EDGE_FALSE_VALUE;
+
+ set_immediate_dominator (CDI_DOMINATORS, seq_start_bb, entry_bb);
+ set_immediate_dominator (CDI_DOMINATORS, body_bb,
+ recompute_dominator (CDI_DOMINATORS, body_bb));
+ set_immediate_dominator (CDI_DOMINATORS, fin_bb,
+ recompute_dominator (CDI_DOMINATORS, fin_bb));
}
adj = STEP + 1;
n = (adj + N2 - N1) / STEP;
trip = 0;
+ V = threadid * CHUNK * STEP + N1; -- this extra definition of V is
+ here so that V is defined
+ if the loop is not entered
L0:
s0 = (trip * nthreads + threadid) * CHUNK;
e0 = min(s0 + CHUNK, n);
L4:
*/
-static basic_block
+static void
expand_omp_for_static_chunk (struct omp_region *region, struct omp_for_data *fd)
{
- tree l0, l1, l2, l3, l4, n, s0, e0, e, t;
- tree trip, nthreads, threadid;
- tree type, utype;
+ tree n, s0, e0, e, t, phi, nphi, args;
+ tree trip_var, trip_init, trip_main, trip_back, nthreads, threadid;
+ tree type, cont, v_main, v_back, v_extra;
basic_block entry_bb, exit_bb, body_bb, seq_start_bb, iter_part_bb;
- basic_block trip_update_bb, new_exit_bb;
- edge exit_edge;
- tree list;
+ basic_block trip_update_bb, cont_bb, fin_bb;
block_stmt_iterator si;
+ edge se, re, ene;
- l0 = create_artificial_label ();
- l1 = create_artificial_label ();
- l2 = create_artificial_label ();
- l3 = create_artificial_label ();
- l4 = create_artificial_label ();
-
type = TREE_TYPE (fd->v);
- utype = lang_hooks.types.unsigned_type (type);
- entry_bb = bb_for_stmt (region->entry);
- body_bb = single_succ (entry_bb);
-
- exit_bb = bb_for_stmt (region->exit);
+ entry_bb = region->entry;
+ se = split_block (entry_bb, last_stmt (entry_bb));
+ entry_bb = se->src;
+ iter_part_bb = se->dest;
+ cont_bb = region->cont;
+ gcc_assert (EDGE_COUNT (iter_part_bb->succs) == 2);
+ gcc_assert (BRANCH_EDGE (iter_part_bb)->dest
+ == FALLTHRU_EDGE (cont_bb)->dest);
+ seq_start_bb = split_edge (FALLTHRU_EDGE (iter_part_bb));
+ body_bb = single_succ (seq_start_bb);
+ gcc_assert (BRANCH_EDGE (cont_bb)->dest == body_bb);
+ gcc_assert (EDGE_COUNT (cont_bb->succs) == 2);
+ fin_bb = FALLTHRU_EDGE (cont_bb)->dest;
+ trip_update_bb = split_edge (FALLTHRU_EDGE (cont_bb));
+ exit_bb = region->exit;
/* Trip and adjustment setup goes in ENTRY_BB. */
- list = alloc_stmt_list ();
+ si = bsi_last (entry_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_FOR);
- t = built_in_decls[BUILT_IN_OMP_GET_NUM_THREADS];
- t = build_function_call_expr (t, NULL);
- t = fold_convert (utype, t);
- nthreads = get_formal_tmp_var (t, &list);
+ t = build_call_expr (built_in_decls[BUILT_IN_OMP_GET_NUM_THREADS], 0);
+ t = fold_convert (type, t);
+ nthreads = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
- t = built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM];
- t = build_function_call_expr (t, NULL);
- t = fold_convert (utype, t);
- threadid = get_formal_tmp_var (t, &list);
-
- fd->n1 = fold_convert (type, fd->n1);
- if (!is_gimple_val (fd->n1))
- fd->n1 = get_formal_tmp_var (fd->n1, &list);
-
- fd->n2 = fold_convert (type, fd->n2);
- if (!is_gimple_val (fd->n2))
- fd->n2 = get_formal_tmp_var (fd->n2, &list);
-
- fd->step = fold_convert (type, fd->step);
- if (!is_gimple_val (fd->step))
- fd->step = get_formal_tmp_var (fd->step, &list);
-
- fd->chunk_size = fold_convert (utype, fd->chunk_size);
- if (!is_gimple_val (fd->chunk_size))
- fd->chunk_size = get_formal_tmp_var (fd->chunk_size, &list);
+ t = build_call_expr (built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM], 0);
+ t = fold_convert (type, t);
+ threadid = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ fd->n1 = force_gimple_operand_bsi (&si, fold_convert (type, fd->n1),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
+ fd->n2 = force_gimple_operand_bsi (&si, fold_convert (type, fd->n2),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
+ fd->step = force_gimple_operand_bsi (&si, fold_convert (type, fd->step),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
+ fd->chunk_size
+ = force_gimple_operand_bsi (&si, fold_convert (type,
+ fd->chunk_size),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
t = build_int_cst (type, (fd->cond_code == LT_EXPR ? -1 : 1));
t = fold_build2 (PLUS_EXPR, type, fd->step, t);
t = fold_build2 (PLUS_EXPR, type, t, fd->n2);
t = fold_build2 (MINUS_EXPR, type, t, fd->n1);
t = fold_build2 (TRUNC_DIV_EXPR, type, t, fd->step);
- t = fold_convert (utype, t);
- if (is_gimple_val (t))
- n = t;
+ t = fold_convert (type, t);
+ n = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ trip_var = create_tmp_var (type, ".trip");
+ if (gimple_in_ssa_p (cfun))
+ {
+ add_referenced_var (trip_var);
+ trip_init = make_ssa_name (trip_var, NULL_TREE);
+ trip_main = make_ssa_name (trip_var, NULL_TREE);
+ trip_back = make_ssa_name (trip_var, NULL_TREE);
+ }
else
- n = get_formal_tmp_var (t, &list);
+ {
+ trip_init = trip_var;
+ trip_main = trip_var;
+ trip_back = trip_var;
+ }
- t = build_int_cst (utype, 0);
- trip = get_initialized_tmp_var (t, &list, NULL);
+ t = build_gimple_modify_stmt (trip_init, build_int_cst (type, 0));
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (trip_init) = t;
- si = bsi_last (entry_bb);
- gcc_assert (bsi_stmt (si) && TREE_CODE (bsi_stmt (si)) == OMP_FOR);
+ t = fold_build2 (MULT_EXPR, type, threadid, fd->chunk_size);
+ t = fold_build2 (MULT_EXPR, type, t, fd->step);
+ t = fold_build2 (PLUS_EXPR, type, t, fd->n1);
+ v_extra = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ true, BSI_SAME_STMT);
+
+ /* Remove the OMP_FOR. */
bsi_remove (&si, true);
- si = bsi_last (entry_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
/* Iteration space partitioning goes in ITER_PART_BB. */
- list = alloc_stmt_list ();
-
- t = build1 (LABEL_EXPR, void_type_node, l0);
- gimplify_and_add (t, &list);
+ si = bsi_last (iter_part_bb);
- t = build2 (MULT_EXPR, utype, trip, nthreads);
- t = build2 (PLUS_EXPR, utype, t, threadid);
- t = build2 (MULT_EXPR, utype, t, fd->chunk_size);
- s0 = get_formal_tmp_var (t, &list);
+ t = fold_build2 (MULT_EXPR, type, trip_main, nthreads);
+ t = fold_build2 (PLUS_EXPR, type, t, threadid);
+ t = fold_build2 (MULT_EXPR, type, t, fd->chunk_size);
+ s0 = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
- t = build2 (PLUS_EXPR, utype, s0, fd->chunk_size);
- t = build2 (MIN_EXPR, utype, t, n);
- e0 = get_formal_tmp_var (t, &list);
+ t = fold_build2 (PLUS_EXPR, type, s0, fd->chunk_size);
+ t = fold_build2 (MIN_EXPR, type, t, n);
+ e0 = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
t = build2 (LT_EXPR, boolean_type_node, s0, n);
- t = build3 (COND_EXPR, void_type_node, t,
- build_and_jump (&l1), build_and_jump (&l4));
- append_to_statement_list (t, &list);
-
- iter_part_bb = create_empty_bb (entry_bb);
- si = bsi_start (iter_part_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
/* Setup code for sequential iteration goes in SEQ_START_BB. */
- list = alloc_stmt_list ();
-
- t = build1 (LABEL_EXPR, void_type_node, l1);
- gimplify_and_add (t, &list);
+ si = bsi_start (seq_start_bb);
t = fold_convert (type, s0);
- t = build2 (MULT_EXPR, type, t, fd->step);
- t = build2 (PLUS_EXPR, type, t, fd->n1);
- t = build2 (MODIFY_EXPR, void_type_node, fd->v, t);
- gimplify_and_add (t, &list);
+ t = fold_build2 (MULT_EXPR, type, t, fd->step);
+ t = fold_build2 (PLUS_EXPR, type, t, fd->n1);
+ t = force_gimple_operand_bsi (&si, t, false, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
+ t = build_gimple_modify_stmt (fd->v, t);
+ bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (fd->v) = t;
t = fold_convert (type, e0);
- t = build2 (MULT_EXPR, type, t, fd->step);
- t = build2 (PLUS_EXPR, type, t, fd->n1);
- e = get_formal_tmp_var (t, &list);
-
- seq_start_bb = create_empty_bb (iter_part_bb);
- si = bsi_start (seq_start_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
-
- /* Main loop body goes in BODY_BB. */
- si = bsi_start (body_bb);
- t = build1 (LABEL_EXPR, void_type_node, l2);
- bsi_insert_before (&si, t, BSI_CONTINUE_LINKING);
-
- /* Split EXIT_BB. The code controlling the sequential loop goes in
- the first half. The trip update code goes into the second half
- (TRIP_UPDATE_BB). */
- list = alloc_stmt_list ();
-
- t = build2 (PLUS_EXPR, type, fd->v, fd->step);
- t = build2 (MODIFY_EXPR, void_type_node, fd->v, t);
- gimplify_and_add (t, &list);
-
- t = build2 (fd->cond_code, boolean_type_node, fd->v, e);
- t = get_formal_tmp_var (t, &list);
- t = build3 (COND_EXPR, void_type_node, t,
- build_and_jump (&l2), build_and_jump (&l3));
- append_to_statement_list (t, &list);
+ t = fold_build2 (MULT_EXPR, type, t, fd->step);
+ t = fold_build2 (PLUS_EXPR, type, t, fd->n1);
+ e = force_gimple_operand_bsi (&si, t, true, NULL_TREE,
+ false, BSI_CONTINUE_LINKING);
+
+ /* The code controlling the sequential loop goes in CONT_BB,
+ replacing the OMP_CONTINUE. */
+ si = bsi_last (cont_bb);
+ cont = bsi_stmt (si);
+ gcc_assert (TREE_CODE (cont) == OMP_CONTINUE);
+ v_main = TREE_OPERAND (cont, 1);
+ v_back = TREE_OPERAND (cont, 0);
+
+ t = build2 (PLUS_EXPR, type, v_main, fd->step);
+ t = build_gimple_modify_stmt (v_back, t);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (v_back) = t;
+
+ t = build2 (fd->cond_code, boolean_type_node, v_back, e);
+ t = build3 (COND_EXPR, void_type_node, t, NULL_TREE, NULL_TREE);
+ bsi_insert_before (&si, t, BSI_SAME_STMT);
- si = bsi_last (exit_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_RETURN_EXPR);
+ /* Remove OMP_CONTINUE. */
bsi_remove (&si, true);
- exit_edge = split_block (exit_bb, t);
- si = bsi_last (exit_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
/* Trip update code goes into TRIP_UPDATE_BB. */
- trip_update_bb = exit_edge->dest;
- list = alloc_stmt_list ();
-
- t = build1 (LABEL_EXPR, void_type_node, l3);
- gimplify_and_add (t, &list);
-
- t = build_int_cst (utype, 1);
- t = build2 (PLUS_EXPR, utype, trip, t);
- t = build2 (MODIFY_EXPR, void_type_node, trip, t);
- gimplify_and_add (t, &list);
-
si = bsi_start (trip_update_bb);
- bsi_insert_after (&si, list, BSI_CONTINUE_LINKING);
- exit_edge = single_succ_edge (trip_update_bb);
- exit_edge->flags = EDGE_FALLTHRU;
- new_exit_bb = exit_edge->dest;
- /* Insert exit label on EXIT_EDGE. */
- t = build1 (LABEL_EXPR, void_type_node, l4);
- bsi_insert_on_edge_immediate (exit_edge, t);
+ t = build_int_cst (type, 1);
+ t = build2 (PLUS_EXPR, type, trip_main, t);
+ t = build_gimple_modify_stmt (trip_back, t);
+ bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (trip_back) = t;
- /* Connect the new blocks. */
- remove_edge (single_succ_edge (entry_bb));
- make_edge (entry_bb, iter_part_bb, EDGE_FALLTHRU);
+ /* Replace the OMP_RETURN with a barrier, or nothing. */
+ si = bsi_last (exit_bb);
+ if (!OMP_RETURN_NOWAIT (bsi_stmt (si)))
+ force_gimple_operand_bsi (&si, build_omp_barrier (), false, NULL_TREE,
+ false, BSI_SAME_STMT);
+ bsi_remove (&si, true);
- make_edge (iter_part_bb, seq_start_bb, EDGE_TRUE_VALUE);
- make_edge (iter_part_bb, new_exit_bb, EDGE_FALSE_VALUE);
- remove_edge (exit_edge);
+ /* Connect the new blocks. */
+ find_edge (iter_part_bb, seq_start_bb)->flags = EDGE_TRUE_VALUE;
+ find_edge (iter_part_bb, fin_bb)->flags = EDGE_FALSE_VALUE;
- make_edge (seq_start_bb, body_bb, EDGE_FALLTHRU);
+ find_edge (cont_bb, body_bb)->flags = EDGE_TRUE_VALUE;
+ find_edge (cont_bb, trip_update_bb)->flags = EDGE_FALSE_VALUE;
- make_edge (exit_bb, body_bb, EDGE_TRUE_VALUE);
- find_edge (exit_bb, trip_update_bb)->flags = EDGE_FALSE_VALUE;
+ redirect_edge_and_branch (single_succ_edge (trip_update_bb), iter_part_bb);
- make_edge (trip_update_bb, iter_part_bb, EDGE_FALLTHRU);
+ if (gimple_in_ssa_p (cfun))
+ {
+ /* When we redirect the edge from trip_update_bb to iter_part_bb, we
+ remove arguments of the phi nodes in fin_bb. We need to create
+ appropriate phi nodes in iter_part_bb instead. */
+ se = single_pred_edge (fin_bb);
+ re = single_succ_edge (trip_update_bb);
+ ene = single_succ_edge (entry_bb);
+
+ args = PENDING_STMT (re);
+ PENDING_STMT (re) = NULL_TREE;
+ for (phi = phi_nodes (fin_bb);
+ phi && args;
+ phi = PHI_CHAIN (phi), args = TREE_CHAIN (args))
+ {
+ t = PHI_RESULT (phi);
+ gcc_assert (t == TREE_PURPOSE (args));
+ nphi = create_phi_node (t, iter_part_bb);
+ SSA_NAME_DEF_STMT (t) = nphi;
+
+ t = PHI_ARG_DEF_FROM_EDGE (phi, se);
+ /* A special case -- fd->v is not yet computed in iter_part_bb, we
+ need to use v_extra instead. */
+ if (t == fd->v)
+ t = v_extra;
+ add_phi_arg (nphi, t, ene);
+ add_phi_arg (nphi, TREE_VALUE (args), re);
+ }
+ gcc_assert (!phi && !args);
+ while ((phi = phi_nodes (fin_bb)) != NULL_TREE)
+ remove_phi_node (phi, NULL_TREE, false);
+
+ /* Make phi node for trip. */
+ phi = create_phi_node (trip_main, iter_part_bb);
+ SSA_NAME_DEF_STMT (trip_main) = phi;
+ add_phi_arg (phi, trip_back, single_succ_edge (trip_update_bb));
+ add_phi_arg (phi, trip_init, single_succ_edge (entry_bb));
+ }
- return new_exit_bb;
+ set_immediate_dominator (CDI_DOMINATORS, trip_update_bb, cont_bb);
+ set_immediate_dominator (CDI_DOMINATORS, iter_part_bb,
+ recompute_dominator (CDI_DOMINATORS, iter_part_bb));
+ set_immediate_dominator (CDI_DOMINATORS, fin_bb,
+ recompute_dominator (CDI_DOMINATORS, fin_bb));
+ set_immediate_dominator (CDI_DOMINATORS, seq_start_bb,
+ recompute_dominator (CDI_DOMINATORS, seq_start_bb));
+ set_immediate_dominator (CDI_DOMINATORS, body_bb,
+ recompute_dominator (CDI_DOMINATORS, body_bb));
}
expand_omp_for (struct omp_region *region)
{
struct omp_for_data fd;
- basic_block last_bb = NULL;
- push_gimplify_context ();
+ extract_omp_for_data (last_stmt (region->entry), &fd);
+ region->sched_kind = fd.sched_kind;
- extract_omp_for_data (region->entry, &fd);
+ gcc_assert (EDGE_COUNT (region->entry->succs) == 2);
+ BRANCH_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL;
+ FALLTHRU_EDGE (region->entry)->flags &= ~EDGE_ABNORMAL;
+ if (region->cont)
+ {
+ gcc_assert (EDGE_COUNT (region->cont->succs) == 2);
+ BRANCH_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL;
+ FALLTHRU_EDGE (region->cont)->flags &= ~EDGE_ABNORMAL;
+ }
- if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC && !fd.have_ordered)
+ if (fd.sched_kind == OMP_CLAUSE_SCHEDULE_STATIC
+ && !fd.have_ordered
+ && region->cont != NULL)
{
if (fd.chunk_size == NULL)
- last_bb = expand_omp_for_static_nochunk (region, &fd);
+ expand_omp_for_static_nochunk (region, &fd);
else
- last_bb = expand_omp_for_static_chunk (region, &fd);
+ expand_omp_for_static_chunk (region, &fd);
}
else
{
int fn_index = fd.sched_kind + fd.have_ordered * 4;
int start_ix = BUILT_IN_GOMP_LOOP_STATIC_START + fn_index;
int next_ix = BUILT_IN_GOMP_LOOP_STATIC_NEXT + fn_index;
- last_bb = expand_omp_for_generic (region, &fd, start_ix, next_ix);
+ expand_omp_for_generic (region, &fd, start_ix, next_ix);
}
- pop_gimplify_context (NULL);
+ update_ssa (TODO_update_ssa_only_virtuals);
}
reduction;
If this is a combined parallel sections, replace the call to
- GOMP_sections_start with 'goto L1'. */
+ GOMP_sections_start with call to GOMP_sections_next. */
static void
expand_omp_sections (struct omp_region *region)
{
- tree label_vec, l0, l1, l2, t, u, v;
- unsigned i, len;
- basic_block entry_bb, exit_bb, l0_bb, l1_bb, default_bb;
- edge e, entry_edge, exit_edge;
- edge_iterator ei;
+ tree label_vec, l1, l2, t, u, sections_stmt, vin, vmain, vnext, cont;
+ unsigned i, casei, len;
+ basic_block entry_bb, l0_bb, l1_bb, l2_bb, default_bb;
block_stmt_iterator si;
-
- entry_bb = bb_for_stmt (region->entry);
- exit_bb = bb_for_stmt (region->exit);
-
- l0 = create_artificial_label ();
- l1 = create_artificial_label ();
- l2 = create_artificial_label ();
-
- v = create_tmp_var (unsigned_type_node, ".section");
+ struct omp_region *inner;
+ bool exit_reachable = region->cont != NULL;
+
+ gcc_assert (exit_reachable == (region->exit != NULL));
+ entry_bb = region->entry;
+ l0_bb = single_succ (entry_bb);
+ l1_bb = region->cont;
+ l2_bb = region->exit;
+ if (exit_reachable)
+ {
+ gcc_assert (single_pred (l2_bb) == l0_bb);
+ default_bb = create_empty_bb (l1_bb->prev_bb);
+ l1 = tree_block_label (l1_bb);
+ l2 = tree_block_label (l2_bb);
+ }
+ else
+ {
+ default_bb = create_empty_bb (l0_bb);
+ l1 = NULL_TREE;
+ l2 = tree_block_label (default_bb);
+ }
/* We will build a switch() with enough cases for all the
OMP_SECTION regions, a '0' case to handle the end of more work
and a default case to abort if something goes wrong. */
- len = EDGE_COUNT (entry_bb->succs);
- label_vec = make_tree_vec (len + 2);
+ len = EDGE_COUNT (l0_bb->succs);
+ label_vec = make_tree_vec (len + 1);
- /* Split ENTRY_BB. The call to GOMP_sections_start goes in the
- first half. The second half contains the switch(). */
+ /* The call to GOMP_sections_start goes in ENTRY_BB, replacing the
+ OMP_SECTIONS statement. */
si = bsi_last (entry_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_SECTIONS);
- bsi_remove (&si, true);
- entry_edge = split_block (entry_bb, t);
- l0_bb = entry_edge->dest;
-
+ sections_stmt = bsi_stmt (si);
+ gcc_assert (TREE_CODE (sections_stmt) == OMP_SECTIONS);
+ vin = OMP_SECTIONS_CONTROL (sections_stmt);
if (!is_combined_parallel (region))
{
/* If we are not inside a combined parallel+sections region,
call GOMP_sections_start. */
- t = build_int_cst (unsigned_type_node, len);
- t = tree_cons (NULL, t, NULL);
+ t = build_int_cst (unsigned_type_node,
+ exit_reachable ? len - 1 : len);
u = built_in_decls[BUILT_IN_GOMP_SECTIONS_START];
- t = build_function_call_expr (u, t);
- t = build2 (MODIFY_EXPR, void_type_node, v, t);
- si = bsi_last (entry_bb);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ t = build_call_expr (u, 1, t);
+ }
+ else
+ {
+ /* Otherwise, call GOMP_sections_next. */
+ u = built_in_decls[BUILT_IN_GOMP_SECTIONS_NEXT];
+ t = build_call_expr (u, 0);
}
+ t = build_gimple_modify_stmt (vin, t);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (vin) = t;
+ bsi_remove (&si, true);
- /* The switch() statement replacing OMP_SECTIONS goes in L0_BB. */
+ /* The switch() statement replacing OMP_SECTIONS_SWITCH goes in L0_BB. */
si = bsi_last (l0_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_SECTIONS_SWITCH);
+ if (exit_reachable)
+ {
+ cont = last_stmt (l1_bb);
+ gcc_assert (TREE_CODE (cont) == OMP_CONTINUE);
+ vmain = TREE_OPERAND (cont, 1);
+ vnext = TREE_OPERAND (cont, 0);
+ }
+ else
+ {
+ vmain = vin;
+ vnext = NULL_TREE;
+ }
- t = build1 (LABEL_EXPR, void_type_node, l0);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ t = build3 (SWITCH_EXPR, void_type_node, vmain, NULL, label_vec);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+ bsi_remove (&si, true);
- t = build3 (SWITCH_EXPR, void_type_node, v, NULL, label_vec);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ i = 0;
+ if (exit_reachable)
+ {
+ t = build3 (CASE_LABEL_EXPR, void_type_node,
+ build_int_cst (unsigned_type_node, 0), NULL, l2);
+ TREE_VEC_ELT (label_vec, 0) = t;
+ i++;
+ }
- t = build3 (CASE_LABEL_EXPR, void_type_node,
- build_int_cst (unsigned_type_node, 0), NULL, l2);
- TREE_VEC_ELT (label_vec, 0) = t;
-
/* Convert each OMP_SECTION into a CASE_LABEL_EXPR. */
- i = 1;
- FOR_EACH_EDGE (e, ei, l0_bb->succs)
+ for (inner = region->inner, casei = 1;
+ inner;
+ inner = inner->next, i++, casei++)
{
basic_block s_entry_bb, s_exit_bb;
- e->flags = 0;
- s_entry_bb = e->dest;
- si = bsi_last (s_entry_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_SECTION);
- s_exit_bb = bb_for_stmt (lookup_omp_region (t)->exit);
- bsi_remove (&si, true);
+ s_entry_bb = inner->entry;
+ s_exit_bb = inner->exit;
- t = create_artificial_label ();
- u = build_int_cst (unsigned_type_node, i);
+ t = tree_block_label (s_entry_bb);
+ u = build_int_cst (unsigned_type_node, casei);
u = build3 (CASE_LABEL_EXPR, void_type_node, u, NULL, t);
TREE_VEC_ELT (label_vec, i) = u;
- t = build1 (LABEL_EXPR, void_type_node, t);
+
si = bsi_last (s_entry_bb);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
- i++;
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_SECTION);
+ gcc_assert (i < len || OMP_SECTION_LAST (bsi_stmt (si)));
+ bsi_remove (&si, true);
single_succ_edge (s_entry_bb)->flags = EDGE_FALLTHRU;
+ if (s_exit_bb == NULL)
+ continue;
+
si = bsi_last (s_exit_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_RETURN_EXPR);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_RETURN);
bsi_remove (&si, true);
+
single_succ_edge (s_exit_bb)->flags = EDGE_FALLTHRU;
}
/* Error handling code goes in DEFAULT_BB. */
- default_bb = create_empty_bb (entry_bb);
- si = bsi_start (default_bb);
- t = create_artificial_label ();
+ t = tree_block_label (default_bb);
u = build3 (CASE_LABEL_EXPR, void_type_node, NULL, NULL, t);
- TREE_VEC_ELT (label_vec, len + 1) = u;
- t = build1 (LABEL_EXPR, void_type_node, t);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
-
- t = built_in_decls[BUILT_IN_TRAP];
- t = build_function_call_expr (t, NULL);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
-
+ TREE_VEC_ELT (label_vec, len) = u;
make_edge (l0_bb, default_bb, 0);
- /* Code to get the next section goes in L1_BB. */
- si = bsi_last (exit_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_RETURN_EXPR);
- bsi_remove (&si, true);
- exit_edge = split_block (exit_bb, t);
- l1_bb = exit_edge->src;
- exit_bb = exit_edge->dest;
- si = bsi_start (l1_bb);
- t = build1 (LABEL_EXPR, void_type_node, l1);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
-
- t = built_in_decls[BUILT_IN_GOMP_SECTIONS_NEXT];
- t = build_function_call_expr (t, NULL);
- t = build2 (MODIFY_EXPR, void_type_node, v, t);
+ si = bsi_start (default_bb);
+ t = build_call_expr (built_in_decls[BUILT_IN_TRAP], 0);
bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
- remove_edge (single_succ_edge (l1_bb));
- make_edge (l1_bb, l0_bb, EDGE_FALLTHRU);
-
- /* Exit label in EXIT_BB. */
- si = bsi_last (exit_bb);
- t = build1 (LABEL_EXPR, void_type_node, l2);
- bsi_insert_after (&si, t, BSI_CONTINUE_LINKING);
+ if (exit_reachable)
+ {
+ /* Code to get the next section goes in L1_BB. */
+ si = bsi_last (l1_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_CONTINUE);
+
+ t = build_call_expr (built_in_decls[BUILT_IN_GOMP_SECTIONS_NEXT], 0);
+ t = build_gimple_modify_stmt (vnext, t);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (vnext) = t;
+ bsi_remove (&si, true);
- make_edge (l0_bb, exit_bb, 0);
- single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU;
+ single_succ_edge (l1_bb)->flags = EDGE_FALLTHRU;
- if (is_combined_parallel (region))
- {
- /* If this was a combined parallel+sections region, we did not
- emit a GOMP_sections_start in the entry block, so we just
- need to jump to L1_BB to get the next section. */
- remove_edge (single_succ_edge (entry_bb));
- make_edge (entry_bb, l1_bb, EDGE_FALLTHRU);
+ /* Cleanup function replaces OMP_RETURN in EXIT_BB. */
+ si = bsi_last (l2_bb);
+ if (OMP_RETURN_NOWAIT (bsi_stmt (si)))
+ t = built_in_decls[BUILT_IN_GOMP_SECTIONS_END_NOWAIT];
+ else
+ t = built_in_decls[BUILT_IN_GOMP_SECTIONS_END];
+ t = build_call_expr (t, 0);
+ bsi_insert_after (&si, t, BSI_SAME_STMT);
+ bsi_remove (&si, true);
}
+
+ set_immediate_dominator (CDI_DOMINATORS, default_bb, l0_bb);
}
-/* Generic expansion for OpenMP synchronization directives: single,
- master, ordered and critical. All we need to do here is remove the
- entry and exit markers for REGION. */
+/* Expand code for an OpenMP single directive. We've already expanded
+ much of the code, here we simply place the GOMP_barrier call. */
static void
-expand_omp_synch (struct omp_region *region)
+expand_omp_single (struct omp_region *region)
{
basic_block entry_bb, exit_bb;
block_stmt_iterator si;
- tree t;
+ bool need_barrier = false;
- entry_bb = bb_for_stmt (region->entry);
- exit_bb = bb_for_stmt (region->exit);
+ entry_bb = region->entry;
+ exit_bb = region->exit;
si = bsi_last (entry_bb);
- t = bsi_stmt (si);
- gcc_assert (t
- && (TREE_CODE (t) == OMP_SINGLE
- || TREE_CODE (t) == OMP_MASTER
- || TREE_CODE (t) == OMP_ORDERED
- || TREE_CODE (t) == OMP_CRITICAL));
+ /* The terminal barrier at the end of a GOMP_single_copy sequence cannot
+ be removed. We need to ensure that the thread that entered the single
+ does not exit before the data is copied out by the other threads. */
+ if (find_omp_clause (OMP_SINGLE_CLAUSES (bsi_stmt (si)),
+ OMP_CLAUSE_COPYPRIVATE))
+ need_barrier = true;
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_SINGLE);
bsi_remove (&si, true);
single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU;
si = bsi_last (exit_bb);
- t = bsi_stmt (si);
- gcc_assert (t && TREE_CODE (t) == OMP_RETURN_EXPR);
+ if (!OMP_RETURN_NOWAIT (bsi_stmt (si)) || need_barrier)
+ force_gimple_operand_bsi (&si, build_omp_barrier (), false, NULL_TREE,
+ false, BSI_SAME_STMT);
bsi_remove (&si, true);
single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU;
}
+/* Generic expansion for OpenMP synchronization directives: master,
+ ordered and critical. All we need to do here is remove the entry
+ and exit markers for REGION. */
+
+static void
+expand_omp_synch (struct omp_region *region)
+{
+ basic_block entry_bb, exit_bb;
+ block_stmt_iterator si;
+
+ entry_bb = region->entry;
+ exit_bb = region->exit;
+
+ si = bsi_last (entry_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_SINGLE
+ || TREE_CODE (bsi_stmt (si)) == OMP_MASTER
+ || TREE_CODE (bsi_stmt (si)) == OMP_ORDERED
+ || TREE_CODE (bsi_stmt (si)) == OMP_CRITICAL);
+ bsi_remove (&si, true);
+ single_succ_edge (entry_bb)->flags = EDGE_FALLTHRU;
+
+ if (exit_bb)
+ {
+ si = bsi_last (exit_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (si)) == OMP_RETURN);
+ bsi_remove (&si, true);
+ single_succ_edge (exit_bb)->flags = EDGE_FALLTHRU;
+ }
+}
+
+/* A subroutine of expand_omp_atomic. Attempt to implement the atomic
+ operation as a __sync_fetch_and_op builtin. INDEX is log2 of the
+ size of the data type, and thus usable to find the index of the builtin
+ decl. Returns false if the expression is not of the proper form. */
+
+static bool
+expand_omp_atomic_fetch_op (basic_block load_bb,
+ tree addr, tree loaded_val,
+ tree stored_val, int index)
+{
+ enum built_in_function base;
+ tree decl, itype, call;
+ enum insn_code *optab;
+ tree rhs;
+ basic_block store_bb = single_succ (load_bb);
+ block_stmt_iterator bsi;
+ tree stmt;
+
+ /* We expect to find the following sequences:
+
+ load_bb:
+ OMP_ATOMIC_LOAD (tmp, mem)
+
+ store_bb:
+ val = tmp OP something; (or: something OP tmp)
+ OMP_STORE (val)
+
+ ???FIXME: Allow a more flexible sequence.
+ Perhaps use data flow to pick the statements.
+
+ */
+
+ bsi = bsi_after_labels (store_bb);
+ stmt = bsi_stmt (bsi);
+ if (TREE_CODE (stmt) != GIMPLE_MODIFY_STMT)
+ return false;
+ bsi_next (&bsi);
+ if (TREE_CODE (bsi_stmt (bsi)) != OMP_ATOMIC_STORE)
+ return false;
+
+ if (!operand_equal_p (GIMPLE_STMT_OPERAND (stmt, 0), stored_val, 0))
+ return false;
+
+ rhs = GIMPLE_STMT_OPERAND (stmt, 1);
+
+ /* Check for one of the supported fetch-op operations. */
+ switch (TREE_CODE (rhs))
+ {
+ case PLUS_EXPR:
+ case POINTER_PLUS_EXPR:
+ base = BUILT_IN_FETCH_AND_ADD_N;
+ optab = sync_add_optab;
+ break;
+ case MINUS_EXPR:
+ base = BUILT_IN_FETCH_AND_SUB_N;
+ optab = sync_add_optab;
+ break;
+ case BIT_AND_EXPR:
+ base = BUILT_IN_FETCH_AND_AND_N;
+ optab = sync_and_optab;
+ break;
+ case BIT_IOR_EXPR:
+ base = BUILT_IN_FETCH_AND_OR_N;
+ optab = sync_ior_optab;
+ break;
+ case BIT_XOR_EXPR:
+ base = BUILT_IN_FETCH_AND_XOR_N;
+ optab = sync_xor_optab;
+ break;
+ default:
+ return false;
+ }
+ /* Make sure the expression is of the proper form. */
+ if (operand_equal_p (TREE_OPERAND (rhs, 0), loaded_val, 0))
+ rhs = TREE_OPERAND (rhs, 1);
+ else if (commutative_tree_code (TREE_CODE (rhs))
+ && operand_equal_p (TREE_OPERAND (rhs, 1), loaded_val, 0))
+ rhs = TREE_OPERAND (rhs, 0);
+ else
+ return false;
+
+ decl = built_in_decls[base + index + 1];
+ itype = TREE_TYPE (TREE_TYPE (decl));
+
+ if (optab[TYPE_MODE (itype)] == CODE_FOR_nothing)
+ return false;
+
+ bsi = bsi_last (load_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (bsi)) == OMP_ATOMIC_LOAD);
+ call = build_call_expr (decl, 2, addr, fold_convert (itype, rhs));
+ force_gimple_operand_bsi (&bsi, call, true, NULL_TREE, true, BSI_SAME_STMT);
+ bsi_remove (&bsi, true);
+
+ bsi = bsi_last (store_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (bsi)) == OMP_ATOMIC_STORE);
+ bsi_remove (&bsi, true);
+ bsi = bsi_last (store_bb);
+ bsi_remove (&bsi, true);
+
+ if (gimple_in_ssa_p (cfun))
+ update_ssa (TODO_update_ssa_no_phi);
+
+ return true;
+}
+
+/* A subroutine of expand_omp_atomic. Implement the atomic operation as:
+
+ oldval = *addr;
+ repeat:
+ newval = rhs; // with oldval replacing *addr in rhs
+ oldval = __sync_val_compare_and_swap (addr, oldval, newval);
+ if (oldval != newval)
+ goto repeat;
+
+ INDEX is log2 of the size of the data type, and thus usable to find the
+ index of the builtin decl. */
+
+static bool
+expand_omp_atomic_pipeline (basic_block load_bb, basic_block store_bb,
+ tree addr, tree loaded_val, tree stored_val,
+ int index)
+{
+ tree loadedi, storedi, initial, new_stored, new_storedi, old_vali;
+ tree type, itype, cmpxchg, iaddr;
+ block_stmt_iterator bsi;
+ basic_block loop_header = single_succ (load_bb);
+ tree phi, x;
+ edge e;
+
+ cmpxchg = built_in_decls[BUILT_IN_VAL_COMPARE_AND_SWAP_N + index + 1];
+ type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr)));
+ itype = TREE_TYPE (TREE_TYPE (cmpxchg));
+
+ if (sync_compare_and_swap[TYPE_MODE (itype)] == CODE_FOR_nothing)
+ return false;
+
+ /* Load the initial value, replacing the OMP_ATOMIC_LOAD. */
+ bsi = bsi_last (load_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (bsi)) == OMP_ATOMIC_LOAD);
+ initial = force_gimple_operand_bsi (&bsi, build_fold_indirect_ref (addr),
+ true, NULL_TREE, true, BSI_SAME_STMT);
+ /* Move the value to the LOADED_VAL temporary. */
+ if (gimple_in_ssa_p (cfun))
+ {
+ gcc_assert (phi_nodes (loop_header) == NULL_TREE);
+ phi = create_phi_node (loaded_val, loop_header);
+ SSA_NAME_DEF_STMT (loaded_val) = phi;
+ SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, single_succ_edge (load_bb)),
+ initial);
+ }
+ else
+ bsi_insert_before (&bsi,
+ build_gimple_modify_stmt (loaded_val, initial),
+ BSI_SAME_STMT);
+ bsi_remove (&bsi, true);
+
+ bsi = bsi_last (store_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (bsi)) == OMP_ATOMIC_STORE);
+
+ /* For floating-point values, we'll need to view-convert them to integers
+ so that we can perform the atomic compare and swap. Simplify the
+ following code by always setting up the "i"ntegral variables. */
+ if (INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
+ {
+ loadedi = loaded_val;
+ storedi = stored_val;
+ iaddr = addr;
+ }
+ else
+ {
+ loadedi = force_gimple_operand_bsi (&bsi,
+ build1 (VIEW_CONVERT_EXPR, itype,
+ loaded_val), true,
+ NULL_TREE, true, BSI_SAME_STMT);
+ storedi =
+ force_gimple_operand_bsi (&bsi,
+ build1 (VIEW_CONVERT_EXPR, itype,
+ stored_val), true, NULL_TREE, true,
+ BSI_SAME_STMT);
+ iaddr = fold_convert (build_pointer_type (itype), addr);
+ }
+
+ /* Build the compare&swap statement. */
+ new_storedi = build_call_expr (cmpxchg, 3, iaddr, loadedi, storedi);
+ new_storedi = force_gimple_operand_bsi (&bsi,
+ fold_convert (itype, new_storedi),
+ true, NULL_TREE,
+ true, BSI_SAME_STMT);
+ if (storedi == stored_val)
+ new_stored = new_storedi;
+ else
+ new_stored = force_gimple_operand_bsi (&bsi,
+ build1 (VIEW_CONVERT_EXPR, type,
+ new_storedi), true,
+ NULL_TREE, true, BSI_SAME_STMT);
+
+ if (gimple_in_ssa_p (cfun))
+ old_vali = loadedi;
+ else
+ {
+ old_vali = create_tmp_var (itype, NULL);
+ x = build_gimple_modify_stmt (old_vali, loadedi);
+ bsi_insert_before (&bsi, x, BSI_SAME_STMT);
+
+ x = build_gimple_modify_stmt (loaded_val, new_stored);
+ bsi_insert_before (&bsi, x, BSI_SAME_STMT);
+ }
+
+ /* Note that we always perform the comparison as an integer, even for
+ floating point. This allows the atomic operation to properly
+ succeed even with NaNs and -0.0. */
+ x = build3 (COND_EXPR, void_type_node,
+ build2 (NE_EXPR, boolean_type_node,
+ new_storedi, old_vali), NULL_TREE, NULL_TREE);
+ bsi_insert_before (&bsi, x, BSI_SAME_STMT);
+
+ /* Update cfg. */
+ e = single_succ_edge (store_bb);
+ e->flags &= ~EDGE_FALLTHRU;
+ e->flags |= EDGE_FALSE_VALUE;
+
+ e = make_edge (store_bb, loop_header, EDGE_TRUE_VALUE);
+
+ /* Copy the new value to loaded_val (we already did that before the condition
+ if we are not in SSA). */
+ if (gimple_in_ssa_p (cfun))
+ {
+ phi = phi_nodes (loop_header);
+ SET_USE (PHI_ARG_DEF_PTR_FROM_EDGE (phi, e), new_stored);
+ }
+
+ /* Remove OMP_ATOMIC_STORE. */
+ bsi_remove (&bsi, true);
+
+ if (gimple_in_ssa_p (cfun))
+ update_ssa (TODO_update_ssa_no_phi);
+
+ return true;
+}
+
+/* A subroutine of expand_omp_atomic. Implement the atomic operation as:
+
+ GOMP_atomic_start ();
+ *addr = rhs;
+ GOMP_atomic_end ();
+
+ The result is not globally atomic, but works so long as all parallel
+ references are within #pragma omp atomic directives. According to
+ responses received from omp@openmp.org, appears to be within spec.
+ Which makes sense, since that's how several other compilers handle
+ this situation as well.
+ LOADED_VAL and ADDR are the operands of OMP_ATOMIC_LOAD we're expanding.
+ STORED_VAL is the operand of the matching OMP_ATOMIC_STORE.
+
+ We replace
+ OMP_ATOMIC_LOAD (loaded_val, addr) with
+ loaded_val = *addr;
+
+ and replace
+ OMP_ATOMIC_ATORE (stored_val) with
+ *addr = stored_val;
+*/
+
+static bool
+expand_omp_atomic_mutex (basic_block load_bb, basic_block store_bb,
+ tree addr, tree loaded_val, tree stored_val)
+{
+ block_stmt_iterator bsi;
+ tree t;
+
+ bsi = bsi_last (load_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (bsi)) == OMP_ATOMIC_LOAD);
+
+ t = built_in_decls[BUILT_IN_GOMP_ATOMIC_START];
+ t = build_function_call_expr (t, 0);
+ force_gimple_operand_bsi (&bsi, t, true, NULL_TREE, true, BSI_SAME_STMT);
+
+ t = build_gimple_modify_stmt (loaded_val, build_fold_indirect_ref (addr));
+ if (gimple_in_ssa_p (cfun))
+ SSA_NAME_DEF_STMT (loaded_val) = t;
+ bsi_insert_before (&bsi, t, BSI_SAME_STMT);
+ bsi_remove (&bsi, true);
+
+ bsi = bsi_last (store_bb);
+ gcc_assert (TREE_CODE (bsi_stmt (bsi)) == OMP_ATOMIC_STORE);
+
+ t = build_gimple_modify_stmt (build_fold_indirect_ref (unshare_expr (addr)),
+ stored_val);
+ bsi_insert_before (&bsi, t, BSI_SAME_STMT);
+
+ t = built_in_decls[BUILT_IN_GOMP_ATOMIC_END];
+ t = build_function_call_expr (t, 0);
+ force_gimple_operand_bsi (&bsi, t, true, NULL_TREE, true, BSI_SAME_STMT);
+ bsi_remove (&bsi, true);
+
+ if (gimple_in_ssa_p (cfun))
+ update_ssa (TODO_update_ssa_no_phi);
+ return true;
+}
+
+/* Expand an OMP_ATOMIC statement. We try to expand
+ using expand_omp_atomic_fetch_op. If it failed, we try to
+ call expand_omp_atomic_pipeline, and if it fails too, the
+ ultimate fallback is wrapping the operation in a mutex
+ (expand_omp_atomic_mutex). REGION is the atomic region built
+ by build_omp_regions_1(). */
+
+static void
+expand_omp_atomic (struct omp_region *region)
+{
+ basic_block load_bb = region->entry, store_bb = region->exit;
+ tree load = last_stmt (load_bb), store = last_stmt (store_bb);
+ tree loaded_val = TREE_OPERAND (load, 0);
+ tree addr = TREE_OPERAND (load, 1);
+ tree stored_val = TREE_OPERAND (store, 0);
+ tree type = TYPE_MAIN_VARIANT (TREE_TYPE (TREE_TYPE (addr)));
+ HOST_WIDE_INT index;
+
+ /* Make sure the type is one of the supported sizes. */
+ index = tree_low_cst (TYPE_SIZE_UNIT (type), 1);
+ index = exact_log2 (index);
+ if (index >= 0 && index <= 4)
+ {
+ unsigned int align = TYPE_ALIGN_UNIT (type);
+
+ /* __sync builtins require strict data alignment. */
+ if (exact_log2 (align) >= index)
+ {
+ /* When possible, use specialized atomic update functions. */
+ if ((INTEGRAL_TYPE_P (type) || POINTER_TYPE_P (type))
+ && store_bb == single_succ (load_bb))
+ {
+ if (expand_omp_atomic_fetch_op (load_bb, addr,
+ loaded_val, stored_val, index))
+ return;
+ }
+
+ /* If we don't have specialized __sync builtins, try and implement
+ as a compare and swap loop. */
+ if (expand_omp_atomic_pipeline (load_bb, store_bb, addr,
+ loaded_val, stored_val, index))
+ return;
+ }
+ }
+
+ /* The ultimate fallback is wrapping the operation in a mutex. */
+ expand_omp_atomic_mutex (load_bb, store_bb, addr, loaded_val, stored_val);
+}
+
+
/* Expand the parallel region tree rooted at REGION. Expansion
proceeds in depth-first order. Innermost regions are expanded
first. This way, parallel regions that require a new function to
{
while (region)
{
- enum tree_code code = TREE_CODE (region->entry);
+ /* First, determine whether this is a combined parallel+workshare
+ region. */
+ if (region->type == OMP_PARALLEL)
+ determine_parallel_type (region);
if (region->inner)
expand_omp (region->inner);
- switch (code)
+ switch (region->type)
{
- case OMP_PARALLEL:
- expand_omp_parallel (region);
- break;
+ case OMP_PARALLEL:
+ expand_omp_parallel (region);
+ break;
- case OMP_FOR:
- expand_omp_for (region);
- break;
+ case OMP_FOR:
+ expand_omp_for (region);
+ break;
- case OMP_SECTIONS:
- expand_omp_sections (region);
- break;
+ case OMP_SECTIONS:
+ expand_omp_sections (region);
+ break;
- case OMP_SECTION:
- /* Individual omp sections are handled together with their
- parent OMP_SECTIONS region. */
- break;
+ case OMP_SECTION:
+ /* Individual omp sections are handled together with their
+ parent OMP_SECTIONS region. */
+ break;
- case OMP_SINGLE:
- case OMP_MASTER:
- case OMP_ORDERED:
- case OMP_CRITICAL:
- expand_omp_synch (region);
- break;
+ case OMP_SINGLE:
+ expand_omp_single (region);
+ break;
- default:
- gcc_unreachable ();
+ case OMP_MASTER:
+ case OMP_ORDERED:
+ case OMP_CRITICAL:
+ expand_omp_synch (region);
+ break;
+
+ case OMP_ATOMIC_LOAD:
+ expand_omp_atomic (region);
+ break;
+
+
+ default:
+ gcc_unreachable ();
}
region = region->next;
/* Helper for build_omp_regions. Scan the dominator tree starting at
- block BB. PARENT is the region that contains BB. */
+ block BB. PARENT is the region that contains BB. If SINGLE_TREE is
+ true, the function ends once a single tree is built (otherwise, whole
+ forest of OMP constructs may be built). */
static void
-build_omp_regions_1 (basic_block bb, struct omp_region *parent)
+build_omp_regions_1 (basic_block bb, struct omp_region *parent,
+ bool single_tree)
{
block_stmt_iterator si;
tree stmt;
if (!bsi_end_p (si) && OMP_DIRECTIVE_P (bsi_stmt (si)))
{
struct omp_region *region;
+ enum tree_code code;
stmt = bsi_stmt (si);
-
- if (TREE_CODE (stmt) == OMP_RETURN_EXPR)
+ code = TREE_CODE (stmt);
+ if (code == OMP_RETURN)
{
/* STMT is the return point out of region PARENT. Mark it
as the exit point and make PARENT the immediately
enclosing region. */
gcc_assert (parent);
region = parent;
- region->exit = stmt;
+ region->exit = bb;
+ parent = parent->outer;
+ }
+ else if (code == OMP_ATOMIC_STORE)
+ {
+ /* OMP_ATOMIC_STORE is analoguous to OMP_RETURN, but matches with
+ OMP_ATOMIC_LOAD. */
+ gcc_assert (parent);
+ gcc_assert (parent->type == OMP_ATOMIC_LOAD);
+ region = parent;
+ region->exit = bb;
parent = parent->outer;
+ }
- /* If REGION is a parallel region, determine whether it is
- a combined parallel+workshare region. */
- if (TREE_CODE (region->entry) == OMP_PARALLEL)
- determine_parallel_type (region);
+ else if (code == OMP_CONTINUE)
+ {
+ gcc_assert (parent);
+ parent->cont = bb;
+ }
+ else if (code == OMP_SECTIONS_SWITCH)
+ {
+ /* OMP_SECTIONS_SWITCH is part of OMP_SECTIONS, and we do nothing for
+ it. */ ;
}
else
{
/* Otherwise, this directive becomes the parent for a new
region. */
- region = new_omp_region (stmt, parent);
+ region = new_omp_region (bb, code, parent);
parent = region;
}
-
- gcc_assert (region);
- if (omp_regions == NULL)
- {
- omp_regions = splay_tree_new (splay_tree_compare_pointers, 0, 0);
- root_omp_region = region;
- }
-
- splay_tree_insert (omp_regions, (splay_tree_key) stmt,
- (splay_tree_value) region);
}
+ if (single_tree && !parent)
+ return;
+
for (son = first_dom_son (CDI_DOMINATORS, bb);
son;
son = next_dom_son (CDI_DOMINATORS, son))
- build_omp_regions_1 (son, parent);
+ build_omp_regions_1 (son, parent, single_tree);
+}
+
+/* Builds the tree of OMP regions rooted at ROOT, storing it to
+ root_omp_region. */
+
+static void
+build_omp_regions_root (basic_block root)
+{
+ gcc_assert (root_omp_region == NULL);
+ build_omp_regions_1 (root, NULL, true);
+ gcc_assert (root_omp_region != NULL);
}
+/* Expands omp construct (and its subconstructs) starting in HEAD. */
+
+void
+omp_expand_local (basic_block head)
+{
+ build_omp_regions_root (head);
+ if (dump_file && (dump_flags & TDF_DETAILS))
+ {
+ fprintf (dump_file, "\nOMP region tree\n\n");
+ dump_omp_region (dump_file, root_omp_region, 0);
+ fprintf (dump_file, "\n");
+ }
+
+ remove_exit_barriers (root_omp_region);
+ expand_omp (root_omp_region);
+
+ free_omp_regions ();
+}
/* Scan the CFG and build a tree of OMP regions. Return the root of
the OMP region tree. */
static void
build_omp_regions (void)
{
- gcc_assert (omp_regions == NULL && root_omp_region == NULL);
+ gcc_assert (root_omp_region == NULL);
calculate_dominance_info (CDI_DOMINATORS);
- build_omp_regions_1 (ENTRY_BLOCK_PTR, NULL);
+ build_omp_regions_1 (ENTRY_BLOCK_PTR, NULL, false);
}
/* Main entry point for expanding OMP-GIMPLE into runtime calls. */
-static void
+static unsigned int
execute_expand_omp (void)
{
build_omp_regions ();
- if (root_omp_region)
- {
- if (dump_file)
- {
- fprintf (dump_file, "\nOMP region tree\n\n");
- dump_omp_region (dump_file, root_omp_region, 0);
- fprintf (dump_file, "\n");
- }
+ if (!root_omp_region)
+ return 0;
- expand_omp (root_omp_region);
- splay_tree_delete (omp_regions);
- root_omp_region = NULL;
- omp_regions = NULL;
- free_dominance_info (CDI_DOMINATORS);
- free_dominance_info (CDI_POST_DOMINATORS);
+ if (dump_file)
+ {
+ fprintf (dump_file, "\nOMP region tree\n\n");
+ dump_omp_region (dump_file, root_omp_region, 0);
+ fprintf (dump_file, "\n");
}
- /* Expansion adds basic blocks that may be merged. */
+ remove_exit_barriers (root_omp_region);
+
+ expand_omp (root_omp_region);
+
cleanup_tree_cfg ();
+
+ free_omp_regions ();
+
+ return 0;
+}
+
+/* OMP expansion in SSA form. For testing purposes only. */
+
+static bool
+gate_expand_omp_ssa (void)
+{
+ return flag_openmp_ssa && flag_openmp != 0 && errorcount == 0;
}
+struct tree_opt_pass pass_expand_omp_ssa =
+{
+ "ompexpssa", /* name */
+ gate_expand_omp_ssa, /* gate */
+ execute_expand_omp, /* execute */
+ NULL, /* sub */
+ NULL, /* next */
+ 0, /* static_pass_number */
+ 0, /* tv_id */
+ PROP_gimple_any, /* properties_required */
+ PROP_gimple_lomp, /* properties_provided */
+ 0, /* properties_destroyed */
+ 0, /* todo_flags_start */
+ TODO_dump_func, /* todo_flags_finish */
+ 0 /* letter */
+};
+
+/* OMP expansion -- the default pass, run before creation of SSA form. */
+
static bool
gate_expand_omp (void)
{
- return flag_openmp != 0 && errorcount == 0;
+ return ((!flag_openmp_ssa || !optimize)
+ && flag_openmp != 0 && errorcount == 0);
}
struct tree_opt_pass pass_expand_omp =
static void
lower_omp_sections (tree *stmt_p, omp_context *ctx)
{
- tree new_stmt, stmt, body, bind, block, ilist, olist, new_body;
- tree dlist, region_exit;
+ tree new_stmt, stmt, body, bind, block, ilist, olist, new_body, control;
+ tree t, dlist;
tree_stmt_iterator tsi;
unsigned i, len;
stmt = *stmt_p;
- gcc_assert (OMP_SECTIONS_SECTIONS (stmt) == NULL_TREE);
-
push_gimplify_context ();
dlist = NULL;
for (len = 0; !tsi_end_p (tsi); len++, tsi_next (&tsi))
continue;
- /* There are two markers per section and one end marker for the
- whole construct. */
- OMP_SECTIONS_SECTIONS (stmt) = make_tree_vec (2 * len + 1);
-
tsi = tsi_start (OMP_SECTIONS_BODY (stmt));
body = alloc_stmt_list ();
for (i = 0; i < len; i++, tsi_next (&tsi))
{
omp_context *sctx;
- tree sec_start, sec_end, sec_body;
+ tree sec_start, sec_end;
sec_start = tsi_stmt (tsi);
- sec_body = alloc_stmt_list ();
sctx = maybe_lookup_ctx (sec_start);
gcc_assert (sctx);
+ append_to_statement_list (sec_start, &body);
+
lower_omp (&OMP_SECTION_BODY (sec_start), sctx);
- append_to_statement_list (OMP_SECTION_BODY (sec_start), &sec_body);
+ append_to_statement_list (OMP_SECTION_BODY (sec_start), &body);
+ OMP_SECTION_BODY (sec_start) = NULL;
if (i == len - 1)
{
tree l = alloc_stmt_list ();
lower_lastprivate_clauses (OMP_SECTIONS_CLAUSES (stmt), NULL,
&l, ctx);
- append_to_statement_list (l, &sec_body);
+ append_to_statement_list (l, &body);
+ OMP_SECTION_LAST (sec_start) = 1;
}
- sec_end = make_node (OMP_RETURN_EXPR);
-
- OMP_SECTION_BODY (sec_start) = sec_body;
- append_to_statement_list (sec_start, &body);
+ sec_end = make_node (OMP_RETURN);
append_to_statement_list (sec_end, &body);
-
- TREE_VEC_ELT (OMP_SECTIONS_SECTIONS (stmt), i * 2) = sec_start;
- TREE_VEC_ELT (OMP_SECTIONS_SECTIONS (stmt), i * 2 + 1) = sec_end;
}
block = make_node (BLOCK);
bind = build3 (BIND_EXPR, void_type_node, NULL, body, block);
- maybe_catch_exception (&BIND_EXPR_BODY (bind));
olist = NULL_TREE;
lower_reduction_clauses (OMP_SECTIONS_CLAUSES (stmt), &olist, ctx);
- /* Unless there's a nowait clause, add a barrier afterward. */
- if (!find_omp_clause (OMP_SECTIONS_CLAUSES (stmt), OMP_CLAUSE_NOWAIT))
- build_omp_barrier (&olist);
-
pop_gimplify_context (NULL_TREE);
record_vars_into (ctx->block_vars, ctx->cb.dst_fn);
new_stmt = build3 (BIND_EXPR, void_type_node, NULL, NULL, NULL);
TREE_SIDE_EFFECTS (new_stmt) = 1;
- OMP_SECTIONS_BODY (stmt) = body;
-
- region_exit = make_node (OMP_RETURN_EXPR);
new_body = alloc_stmt_list ();
append_to_statement_list (ilist, &new_body);
append_to_statement_list (stmt, &new_body);
- append_to_statement_list (region_exit, &new_body);
+ append_to_statement_list (make_node (OMP_SECTIONS_SWITCH), &new_body);
+ append_to_statement_list (bind, &new_body);
+
+ control = create_tmp_var (unsigned_type_node, ".section");
+ t = build2 (OMP_CONTINUE, void_type_node, control, control);
+ OMP_SECTIONS_CONTROL (stmt) = control;
+ append_to_statement_list (t, &new_body);
+
append_to_statement_list (olist, &new_body);
append_to_statement_list (dlist, &new_body);
- BIND_EXPR_BODY (new_stmt) = new_body;
- TREE_VEC_ELT (OMP_SECTIONS_SECTIONS (stmt), 2 * len) = region_exit;
+ maybe_catch_exception (&new_body);
+
+ t = make_node (OMP_RETURN);
+ OMP_RETURN_NOWAIT (t) = !!find_omp_clause (OMP_SECTIONS_CLAUSES (stmt),
+ OMP_CLAUSE_NOWAIT);
+ append_to_statement_list (t, &new_body);
+
+ BIND_EXPR_BODY (new_stmt) = new_body;
+ OMP_SECTIONS_BODY (stmt) = NULL;
*stmt_p = new_stmt;
}
{
tree t;
- t = built_in_decls[BUILT_IN_GOMP_SINGLE_START];
- t = build_function_call_expr (t, NULL);
+ t = build_call_expr (built_in_decls[BUILT_IN_GOMP_SINGLE_START], 0);
t = build3 (COND_EXPR, void_type_node, t,
OMP_SINGLE_BODY (single_stmt), NULL);
gimplify_and_add (t, pre_p);
-
- if (!find_omp_clause (OMP_SINGLE_CLAUSES (single_stmt), OMP_CLAUSE_NOWAIT))
- build_omp_barrier (pre_p);
}
static void
lower_omp_single_copy (tree single_stmt, tree *pre_p, omp_context *ctx)
{
- tree ptr_type, t, args, l0, l1, l2, copyin_seq;
+ tree ptr_type, t, l0, l1, l2, copyin_seq;
ctx->sender_decl = create_tmp_var (ctx->record_type, ".omp_copy_o");
l1 = create_artificial_label ();
l2 = create_artificial_label ();
- t = built_in_decls[BUILT_IN_GOMP_SINGLE_COPY_START];
- t = build_function_call_expr (t, NULL);
+ t = build_call_expr (built_in_decls[BUILT_IN_GOMP_SINGLE_COPY_START], 0);
t = fold_convert (ptr_type, t);
- t = build2 (MODIFY_EXPR, void_type_node, ctx->receiver_decl, t);
+ t = build_gimple_modify_stmt (ctx->receiver_decl, t);
gimplify_and_add (t, pre_p);
t = build2 (EQ_EXPR, boolean_type_node, ctx->receiver_decl,
©in_seq, ctx);
t = build_fold_addr_expr (ctx->sender_decl);
- args = tree_cons (NULL, t, NULL);
- t = built_in_decls[BUILT_IN_GOMP_SINGLE_COPY_END];
- t = build_function_call_expr (t, args);
+ t = build_call_expr (built_in_decls[BUILT_IN_GOMP_SINGLE_COPY_END], 1, t);
gimplify_and_add (t, pre_p);
t = build_and_jump (&l2);
t = build1 (LABEL_EXPR, void_type_node, l2);
gimplify_and_add (t, pre_p);
-
- build_omp_barrier (pre_p);
}
push_gimplify_context ();
block = make_node (BLOCK);
- bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
+ *stmt_p = bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
TREE_SIDE_EFFECTS (bind) = 1;
lower_rec_input_clauses (OMP_SINGLE_CLAUSES (single_stmt),
&BIND_EXPR_BODY (bind), &dlist, ctx);
lower_omp (&OMP_SINGLE_BODY (single_stmt), ctx);
+ append_to_statement_list (single_stmt, &BIND_EXPR_BODY (bind));
+
if (ctx->record_type)
lower_omp_single_copy (single_stmt, &BIND_EXPR_BODY (bind), ctx);
else
lower_omp_single_simple (single_stmt, &BIND_EXPR_BODY (bind));
+ OMP_SINGLE_BODY (single_stmt) = NULL;
+
append_to_statement_list (dlist, &BIND_EXPR_BODY (bind));
- t = make_node (OMP_RETURN_EXPR);
- append_to_statement_list (t, &BIND_EXPR_BODY (bind));
+
maybe_catch_exception (&BIND_EXPR_BODY (bind));
+
+ t = make_node (OMP_RETURN);
+ OMP_RETURN_NOWAIT (t) = !!find_omp_clause (OMP_SINGLE_CLAUSES (single_stmt),
+ OMP_CLAUSE_NOWAIT);
+ append_to_statement_list (t, &BIND_EXPR_BODY (bind));
+
pop_gimplify_context (bind);
BIND_EXPR_VARS (bind) = chainon (BIND_EXPR_VARS (bind), ctx->block_vars);
BLOCK_VARS (block) = BIND_EXPR_VARS (bind);
-
- OMP_SINGLE_BODY (single_stmt) = alloc_stmt_list ();
- append_to_statement_list (bind, &OMP_SINGLE_BODY (single_stmt));
}
push_gimplify_context ();
block = make_node (BLOCK);
- bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
+ *stmt_p = bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
TREE_SIDE_EFFECTS (bind) = 1;
- x = built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM];
- x = build_function_call_expr (x, NULL);
+ append_to_statement_list (stmt, &BIND_EXPR_BODY (bind));
+
+ x = build_call_expr (built_in_decls[BUILT_IN_OMP_GET_THREAD_NUM], 0);
x = build2 (EQ_EXPR, boolean_type_node, x, integer_zero_node);
x = build3 (COND_EXPR, void_type_node, x, NULL, build_and_jump (&lab));
gimplify_and_add (x, &BIND_EXPR_BODY (bind));
lower_omp (&OMP_MASTER_BODY (stmt), ctx);
+ maybe_catch_exception (&OMP_MASTER_BODY (stmt));
append_to_statement_list (OMP_MASTER_BODY (stmt), &BIND_EXPR_BODY (bind));
+ OMP_MASTER_BODY (stmt) = NULL;
x = build1 (LABEL_EXPR, void_type_node, lab);
gimplify_and_add (x, &BIND_EXPR_BODY (bind));
- x = make_node (OMP_RETURN_EXPR);
+
+ x = make_node (OMP_RETURN);
+ OMP_RETURN_NOWAIT (x) = 1;
append_to_statement_list (x, &BIND_EXPR_BODY (bind));
- maybe_catch_exception (&BIND_EXPR_BODY (bind));
+
pop_gimplify_context (bind);
BIND_EXPR_VARS (bind) = chainon (BIND_EXPR_VARS (bind), ctx->block_vars);
BLOCK_VARS (block) = BIND_EXPR_VARS (bind);
-
- OMP_MASTER_BODY (stmt) = alloc_stmt_list ();
- append_to_statement_list (bind, &OMP_MASTER_BODY (stmt));
}
push_gimplify_context ();
block = make_node (BLOCK);
- bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
+ *stmt_p = bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
TREE_SIDE_EFFECTS (bind) = 1;
- x = built_in_decls[BUILT_IN_GOMP_ORDERED_START];
- x = build_function_call_expr (x, NULL);
+ append_to_statement_list (stmt, &BIND_EXPR_BODY (bind));
+
+ x = build_call_expr (built_in_decls[BUILT_IN_GOMP_ORDERED_START], 0);
gimplify_and_add (x, &BIND_EXPR_BODY (bind));
lower_omp (&OMP_ORDERED_BODY (stmt), ctx);
+ maybe_catch_exception (&OMP_ORDERED_BODY (stmt));
append_to_statement_list (OMP_ORDERED_BODY (stmt), &BIND_EXPR_BODY (bind));
+ OMP_ORDERED_BODY (stmt) = NULL;
- x = built_in_decls[BUILT_IN_GOMP_ORDERED_END];
- x = build_function_call_expr (x, NULL);
+ x = build_call_expr (built_in_decls[BUILT_IN_GOMP_ORDERED_END], 0);
gimplify_and_add (x, &BIND_EXPR_BODY (bind));
- x = make_node (OMP_RETURN_EXPR);
+
+ x = make_node (OMP_RETURN);
+ OMP_RETURN_NOWAIT (x) = 1;
append_to_statement_list (x, &BIND_EXPR_BODY (bind));
- maybe_catch_exception (&BIND_EXPR_BODY (bind));
+
pop_gimplify_context (bind);
BIND_EXPR_VARS (bind) = chainon (BIND_EXPR_VARS (bind), ctx->block_vars);
BLOCK_VARS (block) = BIND_EXPR_VARS (bind);
-
- OMP_ORDERED_BODY (stmt) = alloc_stmt_list ();
- append_to_statement_list (bind, &OMP_ORDERED_BODY (stmt));
}
name = OMP_CRITICAL_NAME (stmt);
if (name)
{
- tree decl, args;
+ tree decl;
splay_tree_node n;
if (!critical_name_mutexes)
DECL_COMMON (decl) = 1;
DECL_ARTIFICIAL (decl) = 1;
DECL_IGNORED_P (decl) = 1;
- cgraph_varpool_finalize_decl (decl);
+ varpool_finalize_decl (decl);
splay_tree_insert (critical_name_mutexes, (splay_tree_key) name,
(splay_tree_value) decl);
else
decl = (tree) n->value;
- args = tree_cons (NULL, build_fold_addr_expr (decl), NULL);
lock = built_in_decls[BUILT_IN_GOMP_CRITICAL_NAME_START];
- lock = build_function_call_expr (lock, args);
+ lock = build_call_expr (lock, 1, build_fold_addr_expr (decl));
- args = tree_cons (NULL, build_fold_addr_expr (decl), NULL);
unlock = built_in_decls[BUILT_IN_GOMP_CRITICAL_NAME_END];
- unlock = build_function_call_expr (unlock, args);
+ unlock = build_call_expr (unlock, 1, build_fold_addr_expr (decl));
}
else
{
lock = built_in_decls[BUILT_IN_GOMP_CRITICAL_START];
- lock = build_function_call_expr (lock, NULL);
+ lock = build_call_expr (lock, 0);
unlock = built_in_decls[BUILT_IN_GOMP_CRITICAL_END];
- unlock = build_function_call_expr (unlock, NULL);
+ unlock = build_call_expr (unlock, 0);
}
push_gimplify_context ();
block = make_node (BLOCK);
- bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
+ *stmt_p = bind = build3 (BIND_EXPR, void_type_node, NULL, NULL, block);
TREE_SIDE_EFFECTS (bind) = 1;
+ append_to_statement_list (stmt, &BIND_EXPR_BODY (bind));
+
gimplify_and_add (lock, &BIND_EXPR_BODY (bind));
lower_omp (&OMP_CRITICAL_BODY (stmt), ctx);
maybe_catch_exception (&OMP_CRITICAL_BODY (stmt));
append_to_statement_list (OMP_CRITICAL_BODY (stmt), &BIND_EXPR_BODY (bind));
+ OMP_CRITICAL_BODY (stmt) = NULL;
gimplify_and_add (unlock, &BIND_EXPR_BODY (bind));
- t = make_node (OMP_RETURN_EXPR);
+
+ t = make_node (OMP_RETURN);
+ OMP_RETURN_NOWAIT (t) = 1;
append_to_statement_list (t, &BIND_EXPR_BODY (bind));
pop_gimplify_context (bind);
BIND_EXPR_VARS (bind) = chainon (BIND_EXPR_VARS (bind), ctx->block_vars);
BLOCK_VARS (block) = BIND_EXPR_VARS (bind);
-
- OMP_CRITICAL_BODY (stmt) = alloc_stmt_list ();
- append_to_statement_list (bind, &OMP_CRITICAL_BODY (stmt));
}
/* A subroutine of lower_omp_for. Generate code to emit the predicate
for a lastprivate clause. Given a loop control predicate of (V
cond N2), we gate the clause on (!(V cond N2)). The lowered form
- is appended to *BODY_P. */
+ is appended to *DLIST, iterator initialization is appended to
+ *BODY_P. */
static void
lower_omp_for_lastprivate (struct omp_for_data *fd, tree *body_p,
- struct omp_context *ctx)
+ tree *dlist, struct omp_context *ctx)
{
- tree clauses, cond;
+ tree clauses, cond, stmts, vinit, t;
enum tree_code cond_code;
cond_code = fd->cond_code;
cond = build2 (cond_code, boolean_type_node, fd->v, fd->n2);
clauses = OMP_FOR_CLAUSES (fd->for_stmt);
- lower_lastprivate_clauses (clauses, cond, body_p, ctx);
+ stmts = NULL;
+ lower_lastprivate_clauses (clauses, cond, &stmts, ctx);
+ if (stmts != NULL)
+ {
+ append_to_statement_list (stmts, dlist);
+
+ /* Optimize: v = 0; is usually cheaper than v = some_other_constant. */
+ vinit = fd->n1;
+ if (cond_code == EQ_EXPR
+ && host_integerp (fd->n2, 0)
+ && ! integer_zerop (fd->n2))
+ vinit = build_int_cst (TREE_TYPE (fd->v), 0);
+
+ /* Initialize the iterator variable, so that threads that don't execute
+ any iterations don't execute the lastprivate clauses by accident. */
+ t = build_gimple_modify_stmt (fd->v, vinit);
+ gimplify_and_add (t, body_p);
+ }
}
We just need to make sure that VAL1, VAL2 and VAL3 are lowered
using the .omp_data_s mapping, if needed. */
- rhs_p = &TREE_OPERAND (OMP_FOR_INIT (stmt), 1);
+ rhs_p = &GIMPLE_STMT_OPERAND (OMP_FOR_INIT (stmt), 1);
if (!is_gimple_min_invariant (*rhs_p))
*rhs_p = get_formal_tmp_var (*rhs_p, body_p);
if (!is_gimple_min_invariant (*rhs_p))
*rhs_p = get_formal_tmp_var (*rhs_p, body_p);
- rhs_p = &TREE_OPERAND (TREE_OPERAND (OMP_FOR_INCR (stmt), 1), 1);
+ rhs_p = &TREE_OPERAND (GIMPLE_STMT_OPERAND (OMP_FOR_INCR (stmt), 1), 1);
if (!is_gimple_min_invariant (*rhs_p))
*rhs_p = get_formal_tmp_var (*rhs_p, body_p);
/* Once lowered, extract the bounds and clauses. */
extract_omp_for_data (stmt, &fd);
- /* Region exit marker goes at the end of the loop body. */
- t = make_node (OMP_RETURN_EXPR);
- append_to_statement_list (t, &OMP_FOR_BODY (stmt));
- maybe_catch_exception (&OMP_FOR_BODY (stmt));
+ lower_omp_for_lastprivate (&fd, body_p, &dlist, ctx);
+
append_to_statement_list (stmt, body_p);
+ append_to_statement_list (OMP_FOR_BODY (stmt), body_p);
+
+ t = build2 (OMP_CONTINUE, void_type_node, fd.v, fd.v);
+ append_to_statement_list (t, body_p);
+
/* After the loop, add exit clauses. */
- lower_omp_for_lastprivate (&fd, &dlist, ctx);
lower_reduction_clauses (OMP_FOR_CLAUSES (stmt), body_p, ctx);
append_to_statement_list (dlist, body_p);
- /* Add a barrier unless the user specified NOWAIT. Note that if
- this is a combined parallel+loop construct, the barrier will be
- optimized away during expansion (see expand_omp_for). */
- if (!fd.have_nowait)
- {
- tree stmt = alloc_stmt_list ();
- build_omp_barrier (&stmt);
- append_to_statement_list (stmt, body_p);
- }
+ maybe_catch_exception (body_p);
+
+ /* Region exit marker goes at the end of the loop body. */
+ t = make_node (OMP_RETURN);
+ OMP_RETURN_NOWAIT (t) = fd.have_nowait;
+ append_to_statement_list (t, body_p);
pop_gimplify_context (NULL_TREE);
record_vars_into (ctx->block_vars, ctx->cb.dst_fn);
+ OMP_FOR_BODY (stmt) = NULL_TREE;
OMP_FOR_PRE_BODY (stmt) = NULL_TREE;
*stmt_p = new_stmt;
}
+/* Callback for walk_stmts. Check if *TP only contains OMP_FOR
+ or OMP_PARALLEL. */
+
+static tree
+check_combined_parallel (tree *tp, int *walk_subtrees, void *data)
+{
+ struct walk_stmt_info *wi = data;
+ int *info = wi->info;
+
+ *walk_subtrees = 0;
+ switch (TREE_CODE (*tp))
+ {
+ case OMP_FOR:
+ case OMP_SECTIONS:
+ *info = *info == 0 ? 1 : -1;
+ break;
+ default:
+ *info = -1;
+ break;
+ }
+ return NULL;
+}
/* Lower the OpenMP parallel directive in *STMT_P. CTX holds context
information for the directive. */
par_bind = OMP_PARALLEL_BODY (stmt);
par_body = BIND_EXPR_BODY (par_bind);
child_fn = ctx->cb.dst_fn;
+ if (!OMP_PARALLEL_COMBINED (stmt))
+ {
+ struct walk_stmt_info wi;
+ int ws_num = 0;
+
+ memset (&wi, 0, sizeof (wi));
+ wi.callback = check_combined_parallel;
+ wi.info = &ws_num;
+ wi.val_only = true;
+ walk_stmts (&wi, &par_bind);
+ if (ws_num == 1)
+ OMP_PARALLEL_COMBINED (stmt) = 1;
+ }
push_gimplify_context ();
par_ilist = NULL_TREE;
lower_rec_input_clauses (clauses, &par_ilist, &par_olist, ctx);
lower_omp (&par_body, ctx);
- maybe_catch_exception (&par_body);
lower_reduction_clauses (clauses, &par_olist, ctx);
/* Declare all the variables created by mapping and the variables
if (ctx->record_type)
{
t = build_fold_addr_expr (ctx->sender_decl);
- t = build2 (MODIFY_EXPR, void_type_node, ctx->receiver_decl, t);
+ /* fixup_child_record_type might have changed receiver_decl's type. */
+ t = fold_convert (TREE_TYPE (ctx->receiver_decl), t);
+ t = build_gimple_modify_stmt (ctx->receiver_decl, t);
append_to_statement_list (t, &new_body);
}
append_to_statement_list (par_ilist, &new_body);
append_to_statement_list (par_body, &new_body);
append_to_statement_list (par_olist, &new_body);
- t = make_node (OMP_RETURN_EXPR);
+ maybe_catch_exception (&new_body);
+ t = make_node (OMP_RETURN);
append_to_statement_list (t, &new_body);
OMP_PARALLEL_BODY (stmt) = new_body;
tsi_link_before (&wi->tsi, pre, TSI_SAME_STMT);
}
+/* Copy EXP into a temporary. Insert the initialization statement before TSI. */
+
+static tree
+init_tmp_var (tree exp, tree_stmt_iterator *tsi)
+{
+ tree t, stmt;
+
+ t = create_tmp_var (TREE_TYPE (exp), NULL);
+ DECL_GIMPLE_REG_P (t) = 1;
+ stmt = build_gimple_modify_stmt (t, exp);
+ SET_EXPR_LOCUS (stmt, EXPR_LOCUS (tsi_stmt (*tsi)));
+ tsi_link_before (tsi, stmt, TSI_SAME_STMT);
+
+ return t;
+}
+
+/* Similarly, but copy from the temporary and insert the statement
+ after the iterator. */
+
+static tree
+save_tmp_var (tree exp, tree_stmt_iterator *tsi)
+{
+ tree t, stmt;
+
+ t = create_tmp_var (TREE_TYPE (exp), NULL);
+ DECL_GIMPLE_REG_P (t) = 1;
+ stmt = build_gimple_modify_stmt (exp, t);
+ SET_EXPR_LOCUS (stmt, EXPR_LOCUS (tsi_stmt (*tsi)));
+ tsi_link_after (tsi, stmt, TSI_SAME_STMT);
+
+ return t;
+}
/* Callback for walk_stmts. Lower the OpenMP directive pointed by TP. */
case VAR_DECL:
if (ctx && DECL_HAS_VALUE_EXPR_P (t))
- lower_regimplify (tp, wi);
+ {
+ lower_regimplify (&t, wi);
+ if (wi->val_only)
+ {
+ if (wi->is_lhs)
+ t = save_tmp_var (t, &wi->tsi);
+ else
+ t = init_tmp_var (t, &wi->tsi);
+ }
+ *tp = t;
+ }
break;
case ADDR_EXPR:
\f
/* Main entry point. */
-static void
+static unsigned int
execute_lower_omp (void)
{
all_contexts = splay_tree_new (splay_tree_compare_pointers, 0,
splay_tree_delete (all_contexts);
all_contexts = NULL;
}
+ return 0;
}
static bool
};
\f
/* The following is a utility to diagnose OpenMP structured block violations.
- It's part of the "omplower" pass, as that's invoked too late. It should
- be invoked by the respective front ends after gimplification. */
+ It is not part of the "omplower" pass, as that's invoked too late. It
+ should be invoked by the respective front ends after gimplification. */
static splay_tree all_labels;
OSDN Git Service
