This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
[PATCH] Optimize scalars in shared clauses to firstprivate if proven not written (PR tree-optimization/68128)
- From: Jakub Jelinek <jakub at redhat dot com>
- To: Richard Henderson <rth at redhat dot com>, Richard Biener <rguenther at suse dot de>
- Cc: gcc-patches at gcc dot gnu dot org
- Date: Wed, 25 Nov 2015 21:17:30 +0100
- Subject: [PATCH] Optimize scalars in shared clauses to firstprivate if proven not written (PR tree-optimization/68128)
- Authentication-results: sourceware.org; auth=none
- Reply-to: Jakub Jelinek <jakub at redhat dot com>
Hi!
This patch adds a quick optimization for what I believe quite common case,
people mistakenly (because of laziness, lack of knowledge, ...) not using
firstprivate clauses for scalars not written in the body of omp
parallel/task regions. It handles only the easy cases, i.e. non-addressable
scalars, non-references (because references are firstprivatized
differently). As richi said fallback_t during gimplification is not
something that can be fully relied on, I'm instead walk_gimple_seq
the body of constructs right before finalizing clauses during
gimplification, so the body at that point is fully high GIMPLE already,
but it is done during gimplification, because we have the OpenMP contexts
with data sharing of decls around.
Bootstrapped/regtested on x86_64-linux and i686-linux, I'll apply to trunk
tomorrow if there are no objections.
2015-11-25 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/68128
* tree.h (OMP_CLAUSE_SHARED_READONLY): Define.
* gimplify.c: Include gimple-walk.h.
(enum gimplify_omp_var_data): Add GOVD_WRITTEN.
(omp_notice_variable): Set flags to n->value if n already
exists in target region, but we need to jump to do_outer.
(omp_shared_to_firstprivate_optimizable_decl_p,
omp_mark_stores, omp_find_stores_op, omp_find_stores_stmt): New
functions.
(gimplify_adjust_omp_clauses_1): Set OMP_CLAUSE_SHARED_READONLY
on OMP_CLAUSE_SHARED if it is a scalar non-addressable that is
not modified in the body. Call omp_mark_stores for outer
contexts on OMP_CLAUSE_SHARED clauses if they could be written
in the body or on OMP_CLAUSE_LASTPRIVATE.
(gimplify_adjust_omp_clauses): Add body argument, call
omp_find_stores_{stmt,op} on the body through walk_gimple_seq.
Set OMP_CLAUSE_SHARED_READONLY
on OMP_CLAUSE_SHARED if it is a scalar non-addressable that is
not modified in the body. Call omp_mark_stores for outer
contexts on OMP_CLAUSE_SHARED clauses if they could be written
in the body or on OMP_CLAUSE_LASTPRIVATE or on OMP_CLAUSE_LINEAR
without OMP_CLAUSE_LINEAR_NO_COPYOUT or on OMP_CLAUSE_REDUCTION.
(gimplify_oacc_cache, gimplify_omp_parallel, gimplify_omp_task,
gimplify_omp_for, gimplify_omp_workshare, gimplify_omp_target_update,
gimplify_expr): Adjust gimplify_adjust_omp_clauses callers.
* tree-nested.c (convert_nonlocal_omp_clauses,
convert_local_omp_clauses): Clear OMP_CLAUSE_SHARED_READONLY on
non-local vars or local vars referenced from nested routines.
* omp-low.c (scan_sharing_clauses): For OMP_CLAUSE_SHARED_READONLY
attempt to optimize it into OMP_CLAUSE_FIRSTPRIVATE. Even for
TREE_READONLY, don't call use_pointer_for_field with non-NULL
second argument until we are sure we are keeping OMP_CLAUSE_SHARED.
* gcc.dg/gomp/pr68128-1.c: New test.
* gcc.dg/gomp/pr68128-2.c: New test.
--- gcc/tree.h.jj 2015-11-25 09:49:55.000000000 +0100
+++ gcc/tree.h 2015-11-25 10:42:17.695239250 +0100
@@ -1453,6 +1453,11 @@ extern void protected_set_expr_location
#define OMP_CLAUSE_SHARED_FIRSTPRIVATE(NODE) \
(OMP_CLAUSE_SUBCODE_CHECK (NODE, OMP_CLAUSE_SHARED)->base.public_flag)
+/* True on a SHARED clause if a scalar is not modified in the body and
+ thus could be optimized as firstprivate. */
+#define OMP_CLAUSE_SHARED_READONLY(NODE) \
+ TREE_PRIVATE (OMP_CLAUSE_SUBCODE_CHECK (NODE, OMP_CLAUSE_SHARED))
+
#define OMP_CLAUSE_IF_MODIFIER(NODE) \
(OMP_CLAUSE_SUBCODE_CHECK (NODE, OMP_CLAUSE_IF)->omp_clause.subcode.if_modifier)
--- gcc/gimplify.c.jj 2015-11-23 17:13:32.198486706 +0100
+++ gcc/gimplify.c 2015-11-25 17:38:50.784948410 +0100
@@ -56,7 +56,7 @@ along with GCC; see the file COPYING3.
#include "cilk.h"
#include "gomp-constants.h"
#include "tree-dump.h"
-
+#include "gimple-walk.h"
#include "langhooks-def.h" /* FIXME: for lhd_set_decl_assembler_name */
#include "builtins.h"
@@ -87,6 +87,9 @@ enum gimplify_omp_var_data
/* Flag for GOVD_MAP, if it is always, to or always, tofrom mapping. */
GOVD_MAP_ALWAYS_TO = 65536,
+ /* Flag for shared vars that are or might be stored to in the region. */
+ GOVD_WRITTEN = 131072,
+
GOVD_DATA_SHARE_CLASS = (GOVD_SHARED | GOVD_PRIVATE | GOVD_FIRSTPRIVATE
| GOVD_LASTPRIVATE | GOVD_REDUCTION | GOVD_LINEAR
| GOVD_LOCAL)
@@ -6153,7 +6156,8 @@ omp_notice_variable (struct gimplify_omp
/* If nothing changed, there's nothing left to do. */
if ((n->value & flags) == flags)
return ret;
- n->value |= flags;
+ flags |= n->value;
+ n->value = flags;
}
goto do_outer;
}
@@ -7384,6 +7388,123 @@ gimplify_scan_omp_clauses (tree *list_p,
delete struct_map_to_clause;
}
+/* Return true if DECL is a candidate for shared to firstprivate
+ optimization. We only consider non-addressable scalars, not
+ too big, and not references. */
+
+static bool
+omp_shared_to_firstprivate_optimizable_decl_p (tree decl)
+{
+ if (TREE_ADDRESSABLE (decl))
+ return false;
+ tree type = TREE_TYPE (decl);
+ if (!is_gimple_reg_type (type)
+ || TREE_CODE (type) == REFERENCE_TYPE
+ || TREE_ADDRESSABLE (type))
+ return false;
+ /* Don't optimize too large decls, as each thread/task will have
+ its own. */
+ HOST_WIDE_INT len = int_size_in_bytes (type);
+ if (len == -1 || len > 4 * POINTER_SIZE / BITS_PER_UNIT)
+ return false;
+ if (lang_hooks.decls.omp_privatize_by_reference (decl))
+ return false;
+ return true;
+}
+
+/* Helper function of omp_find_stores_op and gimplify_adjust_omp_clauses*.
+ For omp_shared_to_firstprivate_optimizable_decl_p decl mark it as
+ GOVD_WRITTEN in outer contexts. */
+
+static void
+omp_mark_stores (struct gimplify_omp_ctx *ctx, tree decl)
+{
+ for (; ctx; ctx = ctx->outer_context)
+ {
+ splay_tree_node n = splay_tree_lookup (ctx->variables,
+ (splay_tree_key) decl);
+ if (n == NULL)
+ continue;
+ else if (n->value & GOVD_SHARED)
+ {
+ n->value |= GOVD_WRITTEN;
+ return;
+ }
+ else if (n->value & GOVD_DATA_SHARE_CLASS)
+ return;
+ }
+}
+
+/* Helper callback for walk_gimple_seq to discover possible stores
+ to omp_shared_to_firstprivate_optimizable_decl_p decls and set
+ GOVD_WRITTEN if they are GOVD_SHARED in some outer context
+ for those. */
+
+static tree
+omp_find_stores_op (tree *tp, int *walk_subtrees, void *data)
+{
+ struct walk_stmt_info *wi = (struct walk_stmt_info *) data;
+
+ *walk_subtrees = 0;
+ if (!wi->is_lhs)
+ return NULL_TREE;
+
+ tree op = *tp;
+ do
+ {
+ if (handled_component_p (op))
+ op = TREE_OPERAND (op, 0);
+ else if ((TREE_CODE (op) == MEM_REF || TREE_CODE (op) == TARGET_MEM_REF)
+ && TREE_CODE (TREE_OPERAND (op, 0)) == ADDR_EXPR)
+ op = TREE_OPERAND (TREE_OPERAND (op, 0), 0);
+ else
+ break;
+ }
+ while (1);
+ if (!DECL_P (op) || !omp_shared_to_firstprivate_optimizable_decl_p (op))
+ return NULL_TREE;
+
+ omp_mark_stores (gimplify_omp_ctxp, op);
+ return NULL_TREE;
+}
+
+/* Helper callback for walk_gimple_seq to discover possible stores
+ to omp_shared_to_firstprivate_optimizable_decl_p decls and set
+ GOVD_WRITTEN if they are GOVD_SHARED in some outer context
+ for those. */
+
+static tree
+omp_find_stores_stmt (gimple_stmt_iterator *gsi_p,
+ bool *handled_ops_p,
+ struct walk_stmt_info *wi)
+{
+ gimple *stmt = gsi_stmt (*gsi_p);
+ switch (gimple_code (stmt))
+ {
+ /* Don't recurse on OpenMP constructs for which
+ gimplify_adjust_omp_clauses already handled the bodies,
+ except handle gimple_omp_for_pre_body. */
+ case GIMPLE_OMP_FOR:
+ *handled_ops_p = true;
+ if (gimple_omp_for_pre_body (stmt))
+ walk_gimple_seq (gimple_omp_for_pre_body (stmt),
+ omp_find_stores_stmt, omp_find_stores_op, wi);
+ break;
+ case GIMPLE_OMP_PARALLEL:
+ case GIMPLE_OMP_TASK:
+ case GIMPLE_OMP_SECTIONS:
+ case GIMPLE_OMP_SINGLE:
+ case GIMPLE_OMP_TARGET:
+ case GIMPLE_OMP_TEAMS:
+ case GIMPLE_OMP_CRITICAL:
+ *handled_ops_p = true;
+ break;
+ default:
+ break;
+ }
+ return NULL_TREE;
+}
+
struct gimplify_adjust_omp_clauses_data
{
tree *list_p;
@@ -7455,6 +7576,11 @@ gimplify_adjust_omp_clauses_1 (splay_tre
else
gcc_unreachable ();
+ if (((flags & GOVD_LASTPRIVATE)
+ || (code == OMP_CLAUSE_SHARED && (flags & GOVD_WRITTEN)))
+ && omp_shared_to_firstprivate_optimizable_decl_p (decl))
+ omp_mark_stores (gimplify_omp_ctxp->outer_context, decl);
+
clause = build_omp_clause (input_location, code);
OMP_CLAUSE_DECL (clause) = decl;
OMP_CLAUSE_CHAIN (clause) = *list_p;
@@ -7462,6 +7588,10 @@ gimplify_adjust_omp_clauses_1 (splay_tre
OMP_CLAUSE_PRIVATE_DEBUG (clause) = 1;
else if (code == OMP_CLAUSE_PRIVATE && (flags & GOVD_PRIVATE_OUTER_REF))
OMP_CLAUSE_PRIVATE_OUTER_REF (clause) = 1;
+ else if (code == OMP_CLAUSE_SHARED
+ && (flags & GOVD_WRITTEN) == 0
+ && omp_shared_to_firstprivate_optimizable_decl_p (decl))
+ OMP_CLAUSE_SHARED_READONLY (clause) = 1;
else if (code == OMP_CLAUSE_MAP && (flags & GOVD_MAP_0LEN_ARRAY) != 0)
{
tree nc = build_omp_clause (input_location, OMP_CLAUSE_MAP);
@@ -7562,12 +7692,26 @@ gimplify_adjust_omp_clauses_1 (splay_tre
}
static void
-gimplify_adjust_omp_clauses (gimple_seq *pre_p, tree *list_p,
+gimplify_adjust_omp_clauses (gimple_seq *pre_p, gimple_seq body, tree *list_p,
enum tree_code code)
{
struct gimplify_omp_ctx *ctx = gimplify_omp_ctxp;
tree c, decl;
+ if (body)
+ {
+ struct gimplify_omp_ctx *octx;
+ for (octx = ctx; octx; octx = octx->outer_context)
+ if ((octx->region_type & (ORT_PARALLEL | ORT_TASK | ORT_TEAMS)) != 0)
+ break;
+ if (octx)
+ {
+ struct walk_stmt_info wi;
+ memset (&wi, 0, sizeof (wi));
+ walk_gimple_seq (body, omp_find_stores_stmt,
+ omp_find_stores_op, &wi);
+ }
+ }
while ((c = *list_p) != NULL)
{
splay_tree_node n;
@@ -7594,6 +7738,18 @@ gimplify_adjust_omp_clauses (gimple_seq
OMP_CLAUSE_SET_CODE (c, OMP_CLAUSE_PRIVATE);
OMP_CLAUSE_PRIVATE_DEBUG (c) = 1;
}
+ if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_SHARED
+ && (n->value & GOVD_WRITTEN) == 0
+ && DECL_P (decl)
+ && omp_shared_to_firstprivate_optimizable_decl_p (decl))
+ OMP_CLAUSE_SHARED_READONLY (c) = 1;
+ else if (DECL_P (decl)
+ && ((OMP_CLAUSE_CODE (c) == OMP_CLAUSE_SHARED
+ && (n->value & GOVD_WRITTEN) != 1)
+ || (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LINEAR
+ && !OMP_CLAUSE_LINEAR_NO_COPYOUT (c)))
+ && omp_shared_to_firstprivate_optimizable_decl_p (decl))
+ omp_mark_stores (gimplify_omp_ctxp->outer_context, decl);
}
break;
@@ -7620,6 +7776,11 @@ gimplify_adjust_omp_clauses (gimple_seq
"%<lastprivate%> clauses on %<distribute%> "
"construct");
}
+ if (!remove
+ && OMP_CLAUSE_CODE (c) == OMP_CLAUSE_LASTPRIVATE
+ && DECL_P (decl)
+ && omp_shared_to_firstprivate_optimizable_decl_p (decl))
+ omp_mark_stores (gimplify_omp_ctxp->outer_context, decl);
break;
case OMP_CLAUSE_ALIGNED:
@@ -7800,6 +7961,11 @@ gimplify_adjust_omp_clauses (gimple_seq
break;
case OMP_CLAUSE_REDUCTION:
+ decl = OMP_CLAUSE_DECL (c);
+ if (DECL_P (decl)
+ && omp_shared_to_firstprivate_optimizable_decl_p (decl))
+ omp_mark_stores (gimplify_omp_ctxp->outer_context, decl);
+ break;
case OMP_CLAUSE_COPYIN:
case OMP_CLAUSE_COPYPRIVATE:
case OMP_CLAUSE_IF:
@@ -7876,7 +8042,8 @@ gimplify_oacc_cache (tree *expr_p, gimpl
gimplify_scan_omp_clauses (&OACC_CACHE_CLAUSES (expr), pre_p, ORT_ACC,
OACC_CACHE);
- gimplify_adjust_omp_clauses (pre_p, &OACC_CACHE_CLAUSES (expr), OACC_CACHE);
+ gimplify_adjust_omp_clauses (pre_p, NULL, &OACC_CACHE_CLAUSES (expr),
+ OACC_CACHE);
/* TODO: Do something sensible with this information. */
@@ -8023,7 +8190,7 @@ gimplify_omp_parallel (tree *expr_p, gim
else
pop_gimplify_context (NULL);
- gimplify_adjust_omp_clauses (pre_p, &OMP_PARALLEL_CLAUSES (expr),
+ gimplify_adjust_omp_clauses (pre_p, body, &OMP_PARALLEL_CLAUSES (expr),
OMP_PARALLEL);
g = gimple_build_omp_parallel (body,
@@ -8060,7 +8227,8 @@ gimplify_omp_task (tree *expr_p, gimple_
else
pop_gimplify_context (NULL);
- gimplify_adjust_omp_clauses (pre_p, &OMP_TASK_CLAUSES (expr), OMP_TASK);
+ gimplify_adjust_omp_clauses (pre_p, body, &OMP_TASK_CLAUSES (expr),
+ OMP_TASK);
g = gimple_build_omp_task (body,
OMP_TASK_CLAUSES (expr),
@@ -8782,7 +8950,8 @@ gimplify_omp_for (tree *expr_p, gimple_s
TREE_OPERAND (TREE_OPERAND (t, 1), 0) = var;
}
- gimplify_adjust_omp_clauses (pre_p, &OMP_FOR_CLAUSES (orig_for_stmt),
+ gimplify_adjust_omp_clauses (pre_p, for_body,
+ &OMP_FOR_CLAUSES (orig_for_stmt),
TREE_CODE (orig_for_stmt));
int kind;
@@ -9236,7 +9405,8 @@ gimplify_omp_workshare (tree *expr_p, gi
}
else
gimplify_and_add (OMP_BODY (expr), &body);
- gimplify_adjust_omp_clauses (pre_p, &OMP_CLAUSES (expr), TREE_CODE (expr));
+ gimplify_adjust_omp_clauses (pre_p, body, &OMP_CLAUSES (expr),
+ TREE_CODE (expr));
switch (TREE_CODE (expr))
{
@@ -9313,7 +9483,7 @@ gimplify_omp_target_update (tree *expr_p
}
gimplify_scan_omp_clauses (&OMP_STANDALONE_CLAUSES (expr), pre_p,
ort, TREE_CODE (expr));
- gimplify_adjust_omp_clauses (pre_p, &OMP_STANDALONE_CLAUSES (expr),
+ gimplify_adjust_omp_clauses (pre_p, NULL, &OMP_STANDALONE_CLAUSES (expr),
TREE_CODE (expr));
stmt = gimple_build_omp_target (NULL, kind, OMP_STANDALONE_CLAUSES (expr));
@@ -10426,7 +10596,7 @@ gimplify_expr (tree *expr_p, gimple_seq
case OMP_CRITICAL:
gimplify_scan_omp_clauses (&OMP_CRITICAL_CLAUSES (*expr_p),
pre_p, ORT_WORKSHARE, OMP_CRITICAL);
- gimplify_adjust_omp_clauses (pre_p,
+ gimplify_adjust_omp_clauses (pre_p, body,
&OMP_CRITICAL_CLAUSES (*expr_p),
OMP_CRITICAL);
g = gimple_build_omp_critical (body,
--- gcc/tree-nested.c.jj 2015-11-04 11:12:19.000000000 +0100
+++ gcc/tree-nested.c 2015-11-25 12:08:30.218295840 +0100
@@ -1081,6 +1081,8 @@ convert_nonlocal_omp_clauses (tree *pcla
break;
if (decl_function_context (decl) != info->context)
{
+ if (OMP_CLAUSE_CODE (clause) == OMP_CLAUSE_SHARED)
+ OMP_CLAUSE_SHARED_READONLY (clause) = 0;
bitmap_set_bit (new_suppress, DECL_UID (decl));
OMP_CLAUSE_DECL (clause) = get_nonlocal_debug_decl (info, decl);
if (OMP_CLAUSE_CODE (clause) != OMP_CLAUSE_PRIVATE)
@@ -1732,6 +1734,8 @@ convert_local_omp_clauses (tree *pclause
tree field = lookup_field_for_decl (info, decl, NO_INSERT);
if (field)
{
+ if (OMP_CLAUSE_CODE (clause) == OMP_CLAUSE_SHARED)
+ OMP_CLAUSE_SHARED_READONLY (clause) = 0;
bitmap_set_bit (new_suppress, DECL_UID (decl));
OMP_CLAUSE_DECL (clause)
= get_local_debug_decl (info, decl, field);
--- gcc/omp-low.c.jj 2015-11-23 13:29:49.000000000 +0100
+++ gcc/omp-low.c 2015-11-25 16:54:22.086998009 +0100
@@ -1852,14 +1852,18 @@ scan_sharing_clauses (tree clauses, omp_
the receiver side will use them directly. */
if (is_global_var (maybe_lookup_decl_in_outer_ctx (decl, ctx)))
break;
- by_ref = use_pointer_for_field (decl, ctx);
if (OMP_CLAUSE_SHARED_FIRSTPRIVATE (c))
- break;
- if (! TREE_READONLY (decl)
+ {
+ use_pointer_for_field (decl, ctx);
+ break;
+ }
+ by_ref = use_pointer_for_field (decl, NULL);
+ if ((! TREE_READONLY (decl) && !OMP_CLAUSE_SHARED_READONLY (c))
|| TREE_ADDRESSABLE (decl)
|| by_ref
|| is_reference (decl))
{
+ by_ref = use_pointer_for_field (decl, ctx);
install_var_field (decl, by_ref, 3, ctx);
install_var_local (decl, ctx);
break;
--- gcc/testsuite/gcc.dg/gomp/pr68128-1.c.jj 2015-11-25 15:49:27.101292541 +0100
+++ gcc/testsuite/gcc.dg/gomp/pr68128-1.c 2015-11-25 16:11:15.864730952 +0100
@@ -0,0 +1,32 @@
+/* PR tree-optimization/68128 */
+/* { dg-do compile } */
+/* { dg-options "-Ofast -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target i?86-*-* x86_64-*-* } } */
+
+/* Make sure the following loop is vectorized even when not using
+ firstprivate variables for scalar vars that are not modified
+ in the parallel region. */
+
+void
+foo (float *u, float v, float w, float x, float y, float z, float t)
+{
+ int i, j, k, l;
+ float a, *b, c, s, e;
+#pragma omp parallel for private (i, j, k, l, a, b, c, s, e)
+ for (j = 0; j < 1024; j++)
+ {
+ k = j * 64;
+ l = j * 64 + 63;
+ a = v + j * w;
+ b = u + j * 64;
+ for (i = k; i <= l; i++, b++, a += w)
+ {
+ c = a * a + y;
+ s = (1.f - c * x) * (1.f - c * x);
+ e = t * (1 / __builtin_sqrtf (c)) * s;
+ *b += (c < z ? e : 0);
+ }
+ }
+}
+
+/* { dg-final { scan-tree-dump "note: vectorized 1 loops in function" "vect" { target i?86-*-* x86_64-*-* } } } */
--- gcc/testsuite/gcc.dg/gomp/pr68128-2.c.jj 2015-11-25 16:25:14.789828476 +0100
+++ gcc/testsuite/gcc.dg/gomp/pr68128-2.c 2015-11-25 17:45:12.239524874 +0100
@@ -0,0 +1,198 @@
+/* PR tree-optimization/68128 */
+/* { dg-do compile } */
+/* { dg-options "-O2 -fopenmp -fdump-tree-omplower" } */
+
+extern int omp_get_thread_num (void);
+extern int omp_get_ancestor_thread_num (int);
+
+void b1 (int, int);
+
+int
+f1 (void)
+{
+ int a1 = 1;
+ unsigned char a2 = 2;
+ unsigned long a3 = 3;
+ long long a4 = 4;
+ short a5 = 5;
+ char a6 = 6;
+ #pragma omp parallel shared (a1, a2, a3)
+ {
+ if (omp_get_thread_num () == 0)
+ {
+ a1 = a2;
+ a4 = a5;
+ }
+ b1 (a2, a6);
+ #pragma omp barrier
+ if (omp_get_thread_num () == 1)
+ {
+ a1 += a3;
+ a4 += a6;
+ }
+ }
+ return a1 + a2 + a3 + a4 + a5 + a6;
+}
+
+/* { dg-final { scan-tree-dump "shared\\(a1\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "firstprivate\\(a2\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "firstprivate\\(a3\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a4\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "firstprivate\\(a5\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "firstprivate\\(a6\\)" "omplower" } } */
+
+struct S { int a, b; };
+
+void b2 (int *, int *, int, int, struct S, struct S);
+
+void
+f2 (void)
+{
+ struct S a7 = { 7, 7 }, a8 = { 8, 8 };
+ int a9 = 9, a10 = 10;
+ short a11[2] = { 11, 11 };
+ char a12[1] = { 12 };
+ #pragma omp parallel shared (a7, a9, a11)
+ {
+ b2 (&a9, &a10, a11[1], a12[0], a7, a8);
+ }
+}
+
+/* { dg-final { scan-tree-dump "shared\\(a7\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a8\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a9\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a10\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a11\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a12\\)" "omplower" } } */
+
+void b3 (_Complex float, _Complex float);
+
+_Complex float
+f3 (void)
+{
+ _Complex float a13 = 13.0f, a14 = 14.0f, a15 = 15.0f, a16 = 16.0f;
+ #pragma omp parallel shared (a13, a15)
+ {
+ #pragma omp parallel shared (a14)
+ {
+ if (omp_get_thread_num () == 0 && omp_get_ancestor_thread_num (1) == 1)
+ __imag__ a13 = __real__ a15;
+ else if (omp_get_thread_num () == 1 && omp_get_ancestor_thread_num (1) == 0)
+ __real__ a14 = __imag__ a16;
+ b3 (a15, a16);
+ }
+ }
+ return a13 + a14 + a15 + a16;
+}
+
+/* { dg-final { scan-tree-dump-times "shared\\(a13\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "shared\\(a14\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "firstprivate\\(a15\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "firstprivate\\(a16\\)" 2 "omplower" } } */
+
+int
+f4 (void)
+{
+ int a17 = 17, a18 = 18, a19 = 19, a20 = 20, a21 = 21, a22 = 22, a23 = 23, a24 = 0, a25 = 0, a26 = 0;
+ int i;
+ #pragma omp task shared (a17)
+ b1 (a17, a18);
+ b1 (a17, a18);
+ #pragma omp taskwait
+ #pragma omp parallel shared (a19)
+ {
+ #pragma omp task shared (a19)
+ {
+ a19 = 1;
+ }
+ #pragma omp task shared (a20)
+ a20 = a21;
+ #pragma omp for firstprivate (a25) lastprivate (a22) linear (a23:2) reduction (+:a24) private (a26)
+ for (i = 0; i < 10; i++)
+ {
+ a26 = i;
+ a22 = a26 + 7;
+ a23 += 2;
+ a24 += i;
+ a25++;
+ }
+ }
+ return a22 + a23 + a24 + a25 + a26;
+}
+
+/* { dg-final { scan-tree-dump "firstprivate\\(a17\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "firstprivate\\(a18\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump-times "shared\\(a19\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "shared\\(a20\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "firstprivate\\(a21\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "lastprivate\\(a22\\)" 1 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "shared\\(a22\\)" 1 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "linear\\(a23:2\\)" 1 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "shared\\(a23\\)" 1 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "reduction\\(.:a24\\)" 1 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "shared\\(a24\\)" 1 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "firstprivate\\(a25\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "private\\(a26\\)" 1 "omplower" } } */
+
+void
+f5 (void)
+{
+ int a27 = 27, a28 = 28, a29 = 29, a30 = 30;
+ #pragma omp target data map (tofrom: a27, a28)
+ {
+ #pragma omp target map (tofrom: a27)
+ a27++;
+ #pragma omp parallel shared (a27, a28)
+ {
+ #pragma omp critical
+ {
+ /* This might modify a27 for non-shared memory offloading. */
+ #pragma omp target update to (a27)
+ #pragma omp target map (always, from: a28) private (a30)
+ {
+ a28++;
+ a29++;
+ a30 = a29;
+ }
+ }
+ #pragma omp barrier
+ b1 (a27, a28);
+ }
+ }
+}
+
+/* { dg-final { scan-tree-dump "shared\\(a27\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a28\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump-times "firstprivate\\(a29\\)" 2 "omplower" } } */
+/* { dg-final { scan-tree-dump-times "private\\(a30\\)" 1 "omplower" } } */
+
+int
+f6 (void)
+{
+ int a31 = 31, a32 = 32, a33 = 33, a34 = 34;
+ #pragma omp parallel
+ {
+ #pragma omp sections
+ {
+ #pragma omp section
+ {
+ a31 = 9;
+ }
+ #pragma omp section
+ {
+ int i = 10;
+ __builtin_memcpy (&a32, &i, sizeof (int));
+ }
+ }
+ #pragma omp single
+ a33 = 11;
+ #pragma omp atomic
+ a34++;
+ }
+ return a31 + a32 + a33 + a34;
+}
+
+/* { dg-final { scan-tree-dump "shared\\(a31\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a32\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a33\\)" "omplower" } } */
+/* { dg-final { scan-tree-dump "shared\\(a34\\)" "omplower" } } */
Jakub