@@ -147,6 +147,9 @@ struct omp_context
/* True if there is nested scan context with exclusive clause. */
bool scan_exclusive;
+
+ /* True in the second simd loop of for simd with inscan reductions. */
+ bool for_simd_scan_phase;
};
static splay_tree all_contexts;
@@ -2421,6 +2424,85 @@ scan_omp_simd (gimple_stmt_iterator *gsi
scan_omp_for (stmt, outer_ctx)->simt_stmt = new_stmt;
}
+static tree omp_find_scan (gimple_stmt_iterator *, bool *,
+ struct walk_stmt_info *);
+static omp_context *maybe_lookup_ctx (gimple *);
+
+/* Duplicate #pragma omp simd, one for the scan input phase loop and one
+ for scan phase loop. */
+
+static void
+scan_omp_simd_scan (gimple_stmt_iterator *gsi, gomp_for *stmt,
+ omp_context *outer_ctx)
+{
+ /* The only change between inclusive and exclusive scan will be
+ within the first simd loop, so just use inclusive in the
+ worksharing loop. */
+ outer_ctx->scan_inclusive = true;
+ tree c = build_omp_clause (UNKNOWN_LOCATION, OMP_CLAUSE_INCLUSIVE);
+ OMP_CLAUSE_DECL (c) = integer_zero_node;
+
+ gomp_scan *input_stmt = gimple_build_omp_scan (NULL, NULL_TREE);
+ gomp_scan *scan_stmt = gimple_build_omp_scan (NULL, c);
+ gsi_replace (gsi, input_stmt, false);
+ gimple_seq input_body = NULL;
+ gimple_seq_add_stmt (&input_body, stmt);
+ gsi_insert_after (gsi, scan_stmt, GSI_NEW_STMT);
+
+ gimple_stmt_iterator input1_gsi = gsi_none ();
+ struct walk_stmt_info wi;
+ memset (&wi, 0, sizeof (wi));
+ wi.val_only = true;
+ wi.info = (void *) &input1_gsi;
+ walk_gimple_seq_mod (gimple_omp_body_ptr (stmt), omp_find_scan, NULL, &wi);
+ gcc_assert (!gsi_end_p (input1_gsi));
+
+ gimple *input_stmt1 = gsi_stmt (input1_gsi);
+ gsi_next (&input1_gsi);
+ gimple *scan_stmt1 = gsi_stmt (input1_gsi);
+ gcc_assert (scan_stmt1 && gimple_code (scan_stmt1) == GIMPLE_OMP_SCAN);
+ c = gimple_omp_scan_clauses (as_a <gomp_scan *> (scan_stmt1));
+ if (c && OMP_CLAUSE_CODE (c) == OMP_CLAUSE_EXCLUSIVE)
+ std::swap (input_stmt1, scan_stmt1);
+
+ gimple_seq input_body1 = gimple_omp_body (input_stmt1);
+ gimple_omp_set_body (input_stmt1, NULL);
+
+ gimple_seq scan_body = copy_gimple_seq_and_replace_locals (stmt);
+ gomp_for *new_stmt = as_a <gomp_for *> (scan_body);
+
+ gimple_omp_set_body (input_stmt1, input_body1);
+ gimple_omp_set_body (scan_stmt1, NULL);
+
+ gimple_stmt_iterator input2_gsi = gsi_none ();
+ memset (&wi, 0, sizeof (wi));
+ wi.val_only = true;
+ wi.info = (void *) &input2_gsi;
+ walk_gimple_seq_mod (gimple_omp_body_ptr (new_stmt), omp_find_scan,
+ NULL, &wi);
+ gcc_assert (!gsi_end_p (input2_gsi));
+
+ gimple *input_stmt2 = gsi_stmt (input2_gsi);
+ gsi_next (&input2_gsi);
+ gimple *scan_stmt2 = gsi_stmt (input2_gsi);
+ gcc_assert (scan_stmt2 && gimple_code (scan_stmt2) == GIMPLE_OMP_SCAN);
+ if (c && OMP_CLAUSE_CODE (c) == OMP_CLAUSE_EXCLUSIVE)
+ std::swap (input_stmt2, scan_stmt2);
+
+ gimple_omp_set_body (input_stmt2, NULL);
+
+ gimple_omp_set_body (input_stmt, input_body);
+ gimple_omp_set_body (scan_stmt, scan_body);
+
+ omp_context *ctx = new_omp_context (input_stmt, outer_ctx);
+ scan_omp (gimple_omp_body_ptr (input_stmt), ctx);
+
+ ctx = new_omp_context (scan_stmt, outer_ctx);
+ scan_omp (gimple_omp_body_ptr (scan_stmt), ctx);
+
+ maybe_lookup_ctx (new_stmt)->for_simd_scan_phase = true;
+}
+
/* Scan an OpenMP sections directive. */
static void
@@ -3321,6 +3403,19 @@ scan_omp_1_stmt (gimple_stmt_iterator *g
case GIMPLE_OMP_FOR:
if ((gimple_omp_for_kind (as_a <gomp_for *> (stmt))
== GF_OMP_FOR_KIND_SIMD)
+ && gimple_omp_for_combined_into_p (stmt)
+ && gimple_code (ctx->stmt) != GIMPLE_OMP_SCAN)
+ {
+ tree clauses = gimple_omp_for_clauses (as_a <gomp_for *> (stmt));
+ tree c = omp_find_clause (clauses, OMP_CLAUSE_REDUCTION);
+ if (c && OMP_CLAUSE_REDUCTION_INSCAN (c) && !seen_error ())
+ {
+ scan_omp_simd_scan (gsi, as_a <gomp_for *> (stmt), ctx);
+ break;
+ }
+ }
+ if ((gimple_omp_for_kind (as_a <gomp_for *> (stmt))
+ == GF_OMP_FOR_KIND_SIMD)
&& omp_maybe_offloaded_ctx (ctx)
&& omp_max_simt_vf ())
scan_omp_simd (gsi, as_a <gomp_for *> (stmt), ctx);
@@ -3757,7 +3852,7 @@ lower_rec_simd_input_clauses (tree new_v
DECL_ATTRIBUTES (avar));
gimple_add_tmp_var (avar);
tree iavar = avar;
- if (rvar)
+ if (rvar && !ctx->for_simd_scan_phase)
{
/* For inscan reductions, create another array temporary,
which will hold the reduced value. */
@@ -5213,7 +5308,16 @@ lower_rec_input_clauses (tree clauses, g
x = lang_hooks.decls.omp_clause_default_ctor
(c, unshare_expr (ivar),
build_outer_var_ref (var, ctx));
- if (rvarp)
+ if (rvarp && ctx->for_simd_scan_phase)
+ {
+ if (x)
+ gimplify_and_add (x, &llist[0]);
+ x = lang_hooks.decls.omp_clause_dtor (c, ivar);
+ if (x)
+ gimplify_and_add (x, &llist[1]);
+ break;
+ }
+ else if (rvarp)
{
if (x)
{
@@ -5371,6 +5475,8 @@ lower_rec_input_clauses (tree clauses, g
if (OMP_CLAUSE_CODE (c) == OMP_CLAUSE_REDUCTION
&& OMP_CLAUSE_REDUCTION_INSCAN (c))
{
+ if (ctx->for_simd_scan_phase)
+ goto do_dtor;
if (x || (!is_simd
&& OMP_CLAUSE_REDUCTION_OMP_ORIG_REF (c)))
{
@@ -5532,6 +5638,8 @@ lower_rec_input_clauses (tree clauses, g
if (rvarp)
{
+ if (ctx->for_simd_scan_phase)
+ break;
gimplify_assign (ivar, ref, &llist[0]);
ref = build_outer_var_ref (var, ctx);
gimplify_assign (ref, rvar, &llist[3]);
@@ -8629,11 +8737,13 @@ lower_omp_scan (gimple_stmt_iterator *gs
bool input_phase = has_clauses ^ octx->scan_inclusive;
bool is_simd = (gimple_code (octx->stmt) == GIMPLE_OMP_FOR
- && (gimple_omp_for_kind (octx->stmt) & GF_OMP_FOR_SIMD)
- && !gimple_omp_for_combined_into_p (octx->stmt));
+ && (gimple_omp_for_kind (octx->stmt) & GF_OMP_FOR_SIMD));
bool is_for = (gimple_code (octx->stmt) == GIMPLE_OMP_FOR
&& gimple_omp_for_kind (octx->stmt) == GF_OMP_FOR_KIND_FOR
&& !gimple_omp_for_combined_p (octx->stmt));
+ bool is_for_simd = is_simd && gimple_omp_for_combined_into_p (octx->stmt);
+ if (is_for_simd && octx->for_simd_scan_phase)
+ is_simd = false;
if (is_simd)
if (tree c = omp_find_clause (gimple_omp_for_clauses (octx->stmt),
OMP_CLAUSE__SIMDUID_))
@@ -8866,10 +8976,7 @@ lower_omp_scan (gimple_stmt_iterator *gs
}
}
}
- else if (has_clauses)
- sorry_at (gimple_location (stmt),
- "%<#pragma omp scan%> not supported yet");
- if (!is_for)
+ if (is_simd && !is_for_simd)
{
gsi_insert_seq_after (gsi_p, gimple_omp_body (stmt), GSI_SAME_STMT);
gsi_insert_seq_after (gsi_p, before, GSI_SAME_STMT);
@@ -9115,6 +9222,12 @@ omp_find_scan (gimple_stmt_iterator *gsi
{
WALK_SUBSTMTS;
+ case GIMPLE_OMP_FOR:
+ if ((gimple_omp_for_kind (stmt) & GF_OMP_FOR_SIMD)
+ && gimple_omp_for_combined_into_p (stmt))
+ *handled_ops_p = false;
+ break;
+
case GIMPLE_OMP_SCAN:
*(gimple_stmt_iterator *) (wi->info) = *gsi_p;
return integer_zero_node;
@@ -9255,6 +9368,7 @@ static void
lower_omp_for_scan (gimple_seq *body_p, gimple_seq *dlist, gomp_for *stmt,
struct omp_for_data *fd, omp_context *ctx)
{
+ bool is_for_simd = gimple_omp_for_combined_p (stmt);
gcc_assert (ctx->scan_inclusive || ctx->scan_exclusive);
gimple_seq body = gimple_omp_body (stmt);
@@ -9299,6 +9413,45 @@ lower_omp_for_scan (gimple_seq *body_p,
gcc_assert (scan_stmt2 && gimple_code (scan_stmt2) == GIMPLE_OMP_SCAN);
gimple_omp_set_body (scan_stmt2, scan_body);
+ gimple_stmt_iterator input3_gsi = gsi_none ();
+ gimple_stmt_iterator scan3_gsi = gsi_none ();
+ gimple_stmt_iterator input4_gsi = gsi_none ();
+ gimple_stmt_iterator scan4_gsi = gsi_none ();
+ gimple *input_stmt3 = NULL, *scan_stmt3 = NULL;
+ gimple *input_stmt4 = NULL, *scan_stmt4 = NULL;
+ omp_context *input_simd_ctx = NULL, *scan_simd_ctx = NULL;
+ if (is_for_simd)
+ {
+ memset (&wi, 0, sizeof (wi));
+ wi.val_only = true;
+ wi.info = (void *) &input3_gsi;
+ walk_gimple_seq_mod (&input_body, omp_find_scan, NULL, &wi);
+ gcc_assert (!gsi_end_p (input3_gsi));
+
+ input_stmt3 = gsi_stmt (input3_gsi);
+ gsi = input3_gsi;
+ gsi_next (&gsi);
+ scan3_gsi = gsi;
+ scan_stmt3 = gsi_stmt (gsi);
+ gcc_assert (scan_stmt3 && gimple_code (scan_stmt3) == GIMPLE_OMP_SCAN);
+
+ memset (&wi, 0, sizeof (wi));
+ wi.val_only = true;
+ wi.info = (void *) &input4_gsi;
+ walk_gimple_seq_mod (&scan_body, omp_find_scan, NULL, &wi);
+ gcc_assert (!gsi_end_p (input4_gsi));
+
+ input_stmt4 = gsi_stmt (input4_gsi);
+ gsi = input4_gsi;
+ gsi_next (&gsi);
+ scan4_gsi = gsi;
+ scan_stmt4 = gsi_stmt (gsi);
+ gcc_assert (scan_stmt4 && gimple_code (scan_stmt4) == GIMPLE_OMP_SCAN);
+
+ input_simd_ctx = maybe_lookup_ctx (input_stmt3)->outer;
+ scan_simd_ctx = maybe_lookup_ctx (input_stmt4)->outer;
+ }
+
tree num_threads = create_tmp_var (integer_type_node);
tree thread_num = create_tmp_var (integer_type_node);
tree nthreads_decl = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS);
@@ -9390,6 +9543,18 @@ lower_omp_for_scan (gimple_seq *body_p,
x = fold_build2 (POINTER_PLUS_EXPR, TREE_TYPE (rprivb), rprivb, x);
tree rprivb_ref = build_simple_mem_ref_loc (clause_loc, x);
+ tree var4 = is_for_simd ? new_var : var2;
+ tree var5 = NULL_TREE, var6 = NULL_TREE;
+ if (is_for_simd)
+ {
+ var5 = lookup_decl (var, input_simd_ctx);
+ var6 = lookup_decl (var, scan_simd_ctx);
+ if (new_vard != new_var)
+ {
+ var5 = build_simple_mem_ref_loc (clause_loc, var5);
+ var6 = build_simple_mem_ref_loc (clause_loc, var6);
+ }
+ }
if (OMP_CLAUSE_REDUCTION_PLACEHOLDER (c))
{
tree placeholder = OMP_CLAUSE_REDUCTION_PLACEHOLDER (c);
@@ -9401,16 +9566,19 @@ lower_omp_for_scan (gimple_seq *body_p,
gimplify_and_add (x, &clist);
x = build_outer_var_ref (var, ctx);
- x = lang_hooks.decls.omp_clause_assign_op (c, var2, x);
+ x = lang_hooks.decls.omp_clause_assign_op (c, unshare_expr (var4),
+ x);
gimplify_and_add (x, &thr01_list);
tree y = (DECL_HAS_VALUE_EXPR_P (new_vard)
? DECL_VALUE_EXPR (new_vard) : NULL_TREE);
if (var3)
{
- x = lang_hooks.decls.omp_clause_assign_op (c, var2, var3);
+ x = unshare_expr (var4);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, var3);
gimplify_and_add (x, &thrn1_list);
- x = lang_hooks.decls.omp_clause_assign_op (c, var2, var3);
+ x = unshare_expr (var4);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, var3);
gimplify_and_add (x, &thr02_list);
}
else if (OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c))
@@ -9418,11 +9586,13 @@ lower_omp_for_scan (gimple_seq *body_p,
/* Otherwise, assign to it the identity element. */
gimple_seq tseq = OMP_CLAUSE_REDUCTION_GIMPLE_INIT (c);
tseq = copy_gimple_seq_and_replace_locals (tseq);
-
- if (new_vard != new_var)
- val = build_fold_addr_expr_loc (clause_loc, val);
- SET_DECL_VALUE_EXPR (new_vard, val);
- DECL_HAS_VALUE_EXPR_P (new_vard) = 1;
+ if (!is_for_simd)
+ {
+ if (new_vard != new_var)
+ val = build_fold_addr_expr_loc (clause_loc, val);
+ SET_DECL_VALUE_EXPR (new_vard, val);
+ DECL_HAS_VALUE_EXPR_P (new_vard) = 1;
+ }
SET_DECL_VALUE_EXPR (placeholder, error_mark_node);
DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
lower_omp (&tseq, ctx);
@@ -9442,47 +9612,64 @@ lower_omp_for_scan (gimple_seq *body_p,
}
}
- x = lang_hooks.decls.omp_clause_assign_op (c, var2, rprivam1_ref);
+ x = unshare_expr (var4);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, rprivam1_ref);
gimplify_and_add (x, &thrn2_list);
- if (ctx->scan_exclusive)
+ if (is_for_simd)
{
x = unshare_expr (rprivb_ref);
- x = lang_hooks.decls.omp_clause_assign_op (c, x, var2);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, var5);
gimplify_and_add (x, &scan1_list);
}
+ else
+ {
+ if (ctx->scan_exclusive)
+ {
+ x = unshare_expr (rprivb_ref);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, var2);
+ gimplify_and_add (x, &scan1_list);
+ }
- gimple_seq tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c);
- tseq = copy_gimple_seq_and_replace_locals (tseq);
- SET_DECL_VALUE_EXPR (placeholder, var2);
- DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
- lower_omp (&tseq, ctx);
- gimple_seq_add_seq (&scan1_list, tseq);
+ gimple_seq tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c);
+ tseq = copy_gimple_seq_and_replace_locals (tseq);
+ SET_DECL_VALUE_EXPR (placeholder, var2);
+ DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
+ lower_omp (&tseq, ctx);
+ gimple_seq_add_seq (&scan1_list, tseq);
- if (ctx->scan_inclusive)
- {
- x = unshare_expr (rprivb_ref);
- x = lang_hooks.decls.omp_clause_assign_op (c, x, var2);
- gimplify_and_add (x, &scan1_list);
+ if (ctx->scan_inclusive)
+ {
+ x = unshare_expr (rprivb_ref);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, var2);
+ gimplify_and_add (x, &scan1_list);
+ }
}
x = unshare_expr (rpriva_ref);
- x = lang_hooks.decls.omp_clause_assign_op (c, x, var2);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x,
+ unshare_expr (var4));
gimplify_and_add (x, &mdlist);
- x = unshare_expr (new_var);
- x = lang_hooks.decls.omp_clause_assign_op (c, x, var2);
+ x = unshare_expr (is_for_simd ? var6 : new_var);
+ x = lang_hooks.decls.omp_clause_assign_op (c, x, var4);
gimplify_and_add (x, &input2_list);
val = rprivb_ref;
if (new_vard != new_var)
val = build_fold_addr_expr_loc (clause_loc, val);
- tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c);
+ gimple_seq tseq = OMP_CLAUSE_REDUCTION_GIMPLE_MERGE (c);
tseq = copy_gimple_seq_and_replace_locals (tseq);
SET_DECL_VALUE_EXPR (new_vard, val);
DECL_HAS_VALUE_EXPR_P (new_vard) = 1;
- DECL_HAS_VALUE_EXPR_P (placeholder) = 0;
+ if (is_for_simd)
+ {
+ SET_DECL_VALUE_EXPR (placeholder, var6);
+ DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
+ }
+ else
+ DECL_HAS_VALUE_EXPR_P (placeholder) = 0;
lower_omp (&tseq, ctx);
if (y)
SET_DECL_VALUE_EXPR (new_vard, y);
@@ -9491,9 +9678,12 @@ lower_omp_for_scan (gimple_seq *body_p,
DECL_HAS_VALUE_EXPR_P (new_vard) = 0;
SET_DECL_VALUE_EXPR (new_vard, NULL_TREE);
}
- SET_DECL_VALUE_EXPR (placeholder, new_var);
- DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
- lower_omp (&tseq, ctx);
+ if (!is_for_simd)
+ {
+ SET_DECL_VALUE_EXPR (placeholder, new_var);
+ DECL_HAS_VALUE_EXPR_P (placeholder) = 1;
+ lower_omp (&tseq, ctx);
+ }
gimple_seq_add_seq (&input2_list, tseq);
x = build_outer_var_ref (var, ctx);
@@ -9532,29 +9722,38 @@ lower_omp_for_scan (gimple_seq *body_p,
else
{
x = build_outer_var_ref (var, ctx);
- gimplify_assign (var2, x, &thr01_list);
+ gimplify_assign (unshare_expr (var4), x, &thr01_list);
x = omp_reduction_init (c, TREE_TYPE (new_var));
- gimplify_assign (var2, unshare_expr (x), &thrn1_list);
- gimplify_assign (var2, x, &thr02_list);
+ gimplify_assign (unshare_expr (var4), unshare_expr (x),
+ &thrn1_list);
+ gimplify_assign (unshare_expr (var4), x, &thr02_list);
- gimplify_assign (var2, rprivam1_ref, &thrn2_list);
+ gimplify_assign (unshare_expr (var4), rprivam1_ref, &thrn2_list);
enum tree_code code = OMP_CLAUSE_REDUCTION_CODE (c);
if (code == MINUS_EXPR)
code = PLUS_EXPR;
- if (ctx->scan_exclusive)
- gimplify_assign (unshare_expr (rprivb_ref), var2, &scan1_list);
- x = build2 (code, TREE_TYPE (new_var), var2, new_var);
- gimplify_assign (var2, x, &scan1_list);
- if (ctx->scan_inclusive)
- gimplify_assign (unshare_expr (rprivb_ref), var2, &scan1_list);
+ if (is_for_simd)
+ gimplify_assign (unshare_expr (rprivb_ref), var5, &scan1_list);
+ else
+ {
+ if (ctx->scan_exclusive)
+ gimplify_assign (unshare_expr (rprivb_ref), var2,
+ &scan1_list);
+ x = build2 (code, TREE_TYPE (new_var), var2, new_var);
+ gimplify_assign (var2, x, &scan1_list);
+ if (ctx->scan_inclusive)
+ gimplify_assign (unshare_expr (rprivb_ref), var2,
+ &scan1_list);
+ }
- gimplify_assign (unshare_expr (rpriva_ref), var2, &mdlist);
+ gimplify_assign (unshare_expr (rpriva_ref), unshare_expr (var4),
+ &mdlist);
- x = build2 (code, TREE_TYPE (new_var), var2, rprivb_ref);
- gimplify_assign (new_var, x, &input2_list);
+ x = build2 (code, TREE_TYPE (new_var), var4, rprivb_ref);
+ gimplify_assign (is_for_simd ? var6 : new_var, x, &input2_list);
gimplify_assign (build_outer_var_ref (var, ctx), rpriva_ref,
&last_list);
@@ -9568,7 +9767,8 @@ lower_omp_for_scan (gimple_seq *body_p,
g = gimple_build_assign (ivar, PLUS_EXPR, ivar, size_one_node);
gimple_seq_add_stmt (&scan1_list, g);
g = gimple_build_assign (ivar, PLUS_EXPR, ivar, size_one_node);
- gimple_seq_add_stmt (gimple_omp_body_ptr (scan_stmt2), g);
+ gimple_seq_add_stmt (gimple_omp_body_ptr (is_for_simd
+ ? scan_stmt4 : scan_stmt2), g);
tree controlb = create_tmp_var (boolean_type_node);
tree controlp = create_tmp_var (ptr_type_node);
@@ -9598,8 +9798,29 @@ lower_omp_for_scan (gimple_seq *body_p,
*cp2 = gimple_omp_for_clauses (new_stmt);
gimple_omp_for_set_clauses (new_stmt, new_clauses2);
- gimple_omp_set_body (scan_stmt1, scan1_list);
- gimple_omp_set_body (input_stmt2, input2_list);
+ if (is_for_simd)
+ {
+ gimple_seq_add_seq (gimple_omp_body_ptr (scan_stmt3), scan1_list);
+ gimple_seq_add_seq (gimple_omp_body_ptr (input_stmt4), input2_list);
+
+ gsi_insert_seq_after (&input3_gsi, gimple_omp_body (input_stmt3),
+ GSI_SAME_STMT);
+ gsi_remove (&input3_gsi, true);
+ gsi_insert_seq_after (&scan3_gsi, gimple_omp_body (scan_stmt3),
+ GSI_SAME_STMT);
+ gsi_remove (&scan3_gsi, true);
+ gsi_insert_seq_after (&input4_gsi, gimple_omp_body (input_stmt4),
+ GSI_SAME_STMT);
+ gsi_remove (&input4_gsi, true);
+ gsi_insert_seq_after (&scan4_gsi, gimple_omp_body (scan_stmt4),
+ GSI_SAME_STMT);
+ gsi_remove (&scan4_gsi, true);
+ }
+ else
+ {
+ gimple_omp_set_body (scan_stmt1, scan1_list);
+ gimple_omp_set_body (input_stmt2, input2_list);
+ }
gsi_insert_seq_after (&input1_gsi, gimple_omp_body (input_stmt1),
GSI_SAME_STMT);
@@ -8,7 +8,7 @@ f1 (int *c, int *d)
for (i = 0; i < 64; i++)
{
d[i] = a;
- #pragma omp scan exclusive (a) /* { dg-message "sorry, unimplemented: '#pragma omp scan' not supported yet" } */
+ #pragma omp scan exclusive (a)
a += c[i];
}
}
@@ -0,0 +1,118 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+int r, a[1024], b[1024];
+
+__attribute__((noipa)) void
+foo (int *a, int *b)
+{
+ #pragma omp for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int s = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, +:s) if (0)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r) simdlen(1)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int s = 0;
+ #pragma omp parallel for simd reduction (inscan, +:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ }
+ r = 0;
+ baz (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ }
+ return 0;
+}
@@ -0,0 +1,120 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+int r, a[1024], b[1024];
+
+#pragma omp declare reduction (foo: int: omp_out += omp_in) initializer (omp_priv = 0)
+
+__attribute__((noipa)) void
+foo (int *a, int *b)
+{
+ #pragma omp for simd reduction (inscan, foo:r) simdlen (1)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int s = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, foo:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b)
+{
+ #pragma omp parallel for simd reduction (inscan, foo:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int s = 0;
+ #pragma omp parallel for simd if (simd: 0) reduction (inscan, foo:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ }
+ r = 0;
+ baz (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ }
+ return 0;
+}
@@ -0,0 +1,91 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+float r = 1.0f, a[1024], b[1024];
+
+__attribute__((noipa)) void
+foo (float *a, float *b)
+{
+ #pragma omp for simd reduction (inscan, *:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r *= a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) float
+bar (void)
+{
+ float s = -__builtin_inff ();
+ #pragma omp parallel for simd reduction (inscan, max:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s = s > a[i] ? s : a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ float s = 1.0f;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (i < 80)
+ a[i] = (i & 1) ? 0.25f : 0.5f;
+ else if (i < 200)
+ a[i] = (i % 3) == 0 ? 2.0f : (i % 3) == 1 ? 4.0f : 1.0f;
+ else if (i < 280)
+ a[i] = (i & 1) ? 0.25f : 0.5f;
+ else if (i < 380)
+ a[i] = (i % 3) == 0 ? 2.0f : (i % 3) == 1 ? 4.0f : 1.0f;
+ else
+ switch (i % 6)
+ {
+ case 0: a[i] = 0.25f; break;
+ case 1: a[i] = 2.0f; break;
+ case 2: a[i] = -1.0f; break;
+ case 3: a[i] = -4.0f; break;
+ case 4: a[i] = 0.5f; break;
+ case 5: a[i] = 1.0f; break;
+ default: a[i] = 0.0f; break;
+ }
+ b[i] = -19.0f;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r * 16384.0f != 0.125f)
+ abort ();
+ float m = -175.25f;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s *= a[i];
+ if (b[i] != s)
+ abort ();
+ else
+ {
+ a[i] = m - ((i % 3) == 1 ? 2.0f : (i % 3) == 2 ? 4.0f : 0.0f);
+ b[i] = -231.75f;
+ m += 0.75f;
+ }
+ }
+ if (bar () != 592.0f)
+ abort ();
+ s = -__builtin_inff ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (s < a[i])
+ s = a[i];
+ if (b[i] != s)
+ abort ();
+ }
+ return 0;
+}
@@ -0,0 +1,182 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+int r, a[1024], b[1024];
+unsigned short r2, b2[1024];
+unsigned char r3, b3[1024];
+
+__attribute__((noipa)) void
+foo (int *a, int *b, unsigned short *b2, unsigned char *b3)
+{
+ #pragma omp for simd reduction (inscan, +:r, r2, r3) if (simd:0)
+ for (int i = 0; i < 1024; i++)
+ {
+ { r += a[i]; r2 += a[i]; r3 += a[i]; }
+ #pragma omp scan inclusive(r, r2, r3)
+ {
+ b[i] = r;
+ b2[i] = r2;
+ b3[i] = r3;
+ }
+ }
+}
+
+__attribute__((noipa)) int
+bar (unsigned short *s2p, unsigned char *s3p)
+{
+ int s = 0;
+ unsigned short s2 = 0;
+ unsigned char s3 = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, +:s, s2, s3) simdlen (1)
+ for (int i = 0; i < 1024; i++)
+ {
+ {
+ s += 2 * a[i];
+ s2 += 2 * a[i];
+ s3 += 2 * a[i];
+ }
+ #pragma omp scan inclusive(s, s2, s3)
+ { b[i] = s; b2[i] = s2; b3[i] = s3; }
+ }
+ *s2p = s2;
+ *s3p = s3;
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b, unsigned short *b2, unsigned char *b3)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r, r2, r3)
+ for (int i = 0; i < 1024; i++)
+ {
+ {
+ r += a[i];
+ r2 += a[i];
+ r3 += a[i];
+ }
+ #pragma omp scan inclusive(r, r2, r3)
+ {
+ b[i] = r;
+ b2[i] = r2;
+ b3[i] = r3;
+ }
+ }
+}
+
+__attribute__((noipa)) int
+qux (unsigned short *s2p, unsigned char *s3p)
+{
+ int s = 0;
+ unsigned short s2 = 0;
+ unsigned char s3 = 0;
+ #pragma omp parallel for simd reduction (inscan, +:s, s2, s3)
+ for (int i = 0; i < 1024; i++)
+ {
+ { s += 2 * a[i]; s2 += 2 * a[i]; s3 += 2 * a[i]; }
+ #pragma omp scan inclusive(s, s2, s3)
+ { b[i] = s; b2[i] = s2; b3[i] = s3; }
+ }
+ *s2p = s2;
+ *s3p = s3;
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ unsigned short s2;
+ unsigned char s3;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ b2[i] = -1;
+ b3[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, b2, b3);
+ if (r != 1024 * 1023 / 2
+ || r2 != (unsigned short) r
+ || r3 != (unsigned char) r)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ else
+ {
+ b[i] = 25;
+ b2[i] = 24;
+ b3[i] = 26;
+ }
+ }
+ if (bar (&s2, &s3) != 1024 * 1023)
+ abort ();
+ if (s2 != (unsigned short) (1024 * 1023)
+ || s3 != (unsigned char) (1024 * 1023))
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ else
+ {
+ b[i] = -1;
+ b2[i] = -1;
+ b3[i] = -1;
+ }
+ }
+ r = 0;
+ r2 = 0;
+ r3 = 0;
+ baz (a, b, b2, b3);
+ if (r != 1024 * 1023 / 2
+ || r2 != (unsigned short) r
+ || r3 != (unsigned char) r)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ else
+ {
+ b[i] = 25;
+ b2[i] = 24;
+ b3[i] = 26;
+ }
+ }
+ s2 = 0;
+ s3 = 0;
+ if (qux (&s2, &s3) != 1024 * 1023)
+ abort ();
+ if (s2 != (unsigned short) (1024 * 1023)
+ || s3 != (unsigned char) (1024 * 1023))
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ }
+ return 0;
+}
@@ -0,0 +1,118 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+int r, a[1024], b[1024];
+
+__attribute__((noipa)) void
+foo (int *a, int *b)
+{
+ #pragma omp for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int s = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, +:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b)
+{
+ #pragma omp parallel for simd simdlen (1) reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int s = 0;
+ #pragma omp parallel for simd if (simd: 0) reduction (inscan, +:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ s += i;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ s += 2 * i;
+ }
+ r = 0;
+ baz (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ s += i;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ s += 2 * i;
+ }
+ return 0;
+}
@@ -0,0 +1,120 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+int r, a[1024], b[1024];
+
+#pragma omp declare reduction (foo: int: omp_out += omp_in) initializer (omp_priv = 0)
+
+__attribute__((noipa)) void
+foo (int *a, int *b)
+{
+ #pragma omp for simd reduction (inscan, foo:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int s = 0;
+ #pragma omp parallel
+ #pragma omp for simd simdlen (1) reduction (inscan, foo:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b)
+{
+ #pragma omp parallel for simd if (simd: 0) reduction (inscan, foo:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int s = 0;
+ #pragma omp parallel for simd reduction (inscan, foo:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ s += i;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ s += 2 * i;
+ }
+ r = 0;
+ baz (a, b);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ s += i;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ s += 2 * i;
+ }
+ return 0;
+}
@@ -0,0 +1,89 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+float r = 1.0f, a[1024], b[1024];
+
+__attribute__((noipa)) void
+foo (float *a, float *b)
+{
+ #pragma omp for simd reduction (inscan, *:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r *= a[i];
+ }
+}
+
+__attribute__((noipa)) float
+bar (void)
+{
+ float s = -__builtin_inff ();
+ #pragma omp parallel for simd reduction (inscan, max:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s = s > a[i] ? s : a[i];
+ }
+ return s;
+}
+
+int
+main ()
+{
+ float s = 1.0f;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (i < 80)
+ a[i] = (i & 1) ? 0.25f : 0.5f;
+ else if (i < 200)
+ a[i] = (i % 3) == 0 ? 2.0f : (i % 3) == 1 ? 4.0f : 1.0f;
+ else if (i < 280)
+ a[i] = (i & 1) ? 0.25f : 0.5f;
+ else if (i < 380)
+ a[i] = (i % 3) == 0 ? 2.0f : (i % 3) == 1 ? 4.0f : 1.0f;
+ else
+ switch (i % 6)
+ {
+ case 0: a[i] = 0.25f; break;
+ case 1: a[i] = 2.0f; break;
+ case 2: a[i] = -1.0f; break;
+ case 3: a[i] = -4.0f; break;
+ case 4: a[i] = 0.5f; break;
+ case 5: a[i] = 1.0f; break;
+ default: a[i] = 0.0f; break;
+ }
+ b[i] = -19.0f;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r * 16384.0f != 0.125f)
+ abort ();
+ float m = -175.25f;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -231.75f;
+ s *= a[i];
+ a[i] = m - ((i % 3) == 1 ? 2.0f : (i % 3) == 2 ? 4.0f : 0.0f);
+ m += 0.75f;
+ }
+ if (bar () != 592.0f)
+ abort ();
+ s = -__builtin_inff ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ if (s < a[i])
+ s = a[i];
+ }
+ return 0;
+}
@@ -0,0 +1,182 @@
+/* { dg-require-effective-target size32plus } */
+/* { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" } */
+/* { dg-additional-options "-mavx" { target avx_runtime } } */
+/* { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } } */
+
+extern void abort (void);
+int r, a[1024], b[1024];
+unsigned short r2, b2[1024];
+unsigned char r3, b3[1024];
+
+__attribute__((noipa)) void
+foo (int *a, int *b, unsigned short *b2, unsigned char *b3)
+{
+ #pragma omp for simd reduction (inscan, +:r, r2, r3)
+ for (int i = 0; i < 1024; i++)
+ {
+ {
+ b[i] = r;
+ b2[i] = r2;
+ b3[i] = r3;
+ }
+ #pragma omp scan exclusive(r, r2, r3)
+ { r += a[i]; r2 += a[i]; r3 += a[i]; }
+ }
+}
+
+__attribute__((noipa)) int
+bar (unsigned short *s2p, unsigned char *s3p)
+{
+ int s = 0;
+ unsigned short s2 = 0;
+ unsigned char s3 = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, +:s, s2, s3)
+ for (int i = 0; i < 1024; i++)
+ {
+ { b[i] = s; b2[i] = s2; b3[i] = s3; }
+ #pragma omp scan exclusive(s, s2, s3)
+ {
+ s += 2 * a[i];
+ s2 += 2 * a[i];
+ s3 += 2 * a[i];
+ }
+ }
+ *s2p = s2;
+ *s3p = s3;
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b, unsigned short *b2, unsigned char *b3)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r, r2, r3) if (simd: 0)
+ for (int i = 0; i < 1024; i++)
+ {
+ {
+ b[i] = r;
+ b2[i] = r2;
+ b3[i] = r3;
+ }
+ #pragma omp scan exclusive(r, r2, r3)
+ {
+ r += a[i];
+ r2 += a[i];
+ r3 += a[i];
+ }
+ }
+}
+
+__attribute__((noipa)) int
+qux (unsigned short *s2p, unsigned char *s3p)
+{
+ int s = 0;
+ unsigned short s2 = 0;
+ unsigned char s3 = 0;
+ #pragma omp parallel for simd simdlen (1) reduction (inscan, +:s, s2, s3)
+ for (int i = 0; i < 1024; i++)
+ {
+ { b[i] = s; b2[i] = s2; b3[i] = s3; }
+ #pragma omp scan exclusive(s, s2, s3)
+ { s += 2 * a[i]; s2 += 2 * a[i]; s3 += 2 * a[i]; }
+ }
+ *s2p = s2;
+ *s3p = s3;
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ unsigned short s2;
+ unsigned char s3;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ b2[i] = -1;
+ b3[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, b2, b3);
+ if (r != 1024 * 1023 / 2
+ || r2 != (unsigned short) r
+ || r3 != (unsigned char) r)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ else
+ {
+ b[i] = 25;
+ b2[i] = 24;
+ b3[i] = 26;
+ }
+ s += i;
+ }
+ if (bar (&s2, &s3) != 1024 * 1023)
+ abort ();
+ if (s2 != (unsigned short) (1024 * 1023)
+ || s3 != (unsigned char) (1024 * 1023))
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ else
+ {
+ b[i] = -1;
+ b2[i] = -1;
+ b3[i] = -1;
+ }
+ s += 2 * i;
+ }
+ r = 0;
+ r2 = 0;
+ r3 = 0;
+ baz (a, b, b2, b3);
+ if (r != 1024 * 1023 / 2
+ || r2 != (unsigned short) r
+ || r3 != (unsigned char) r)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ else
+ {
+ b[i] = 25;
+ b2[i] = 24;
+ b3[i] = 26;
+ }
+ s += i;
+ }
+ s2 = 0;
+ s3 = 0;
+ if (qux (&s2, &s3) != 1024 * 1023)
+ abort ();
+ if (s2 != (unsigned short) (1024 * 1023)
+ || s3 != (unsigned char) (1024 * 1023))
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s
+ || b2[i] != (unsigned short) s
+ || b3[i] != (unsigned char) s)
+ abort ();
+ s += 2 * i;
+ }
+ return 0;
+}
@@ -0,0 +1,154 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } }
+
+extern "C" void abort ();
+
+struct S {
+ inline S ();
+ inline ~S ();
+ inline S (const S &);
+ inline S & operator= (const S &);
+ int s;
+};
+
+S::S () : s (0)
+{
+}
+
+S::~S ()
+{
+}
+
+S::S (const S &x)
+{
+ s = x.s;
+}
+
+S &
+S::operator= (const S &x)
+{
+ s = x.s;
+ return *this;
+}
+
+static inline void
+ini (S &x)
+{
+ x.s = 0;
+}
+
+S r, a[1024], b[1024];
+
+#pragma omp declare reduction (+: S: omp_out.s += omp_in.s)
+#pragma omp declare reduction (plus: S: omp_out.s += omp_in.s) initializer (ini (omp_priv))
+
+__attribute__((noipa)) void
+foo (S *a, S *b)
+{
+ #pragma omp for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r.s += a[i].s;
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) S
+bar (void)
+{
+ S s;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s.s += 2 * a[i].s;
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return S (s);
+}
+
+__attribute__((noipa)) void
+baz (S *a, S *b)
+{
+ #pragma omp parallel for simd simdlen (1) reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r.s += a[i].s;
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) S
+qux (void)
+{
+ S s;
+ #pragma omp parallel for simd if (simd: 0) reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s.s += 2 * a[i].s;
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return S (s);
+}
+
+int
+main ()
+{
+ S s;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i].s = i;
+ b[i].s = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += i;
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ }
+ if (bar ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += 2 * i;
+ if (b[i].s != s.s)
+ abort ();
+ }
+ r.s = 0;
+ baz (a, b);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += i;
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ }
+ if (qux ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += 2 * i;
+ if (b[i].s != s.s)
+ abort ();
+ }
+ return 0;
+}
@@ -0,0 +1,119 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } }
+
+extern "C" void abort ();
+int r, a[1024], b[1024], q;
+
+__attribute__((noipa)) void
+foo (int *a, int *b, int &r)
+{
+ #pragma omp for simd if (0) reduction (inscan, +:r) nowait
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int &s = q;
+ q = 0;
+ #pragma omp parallel
+ #pragma omp for simd simdlen(1) reduction (inscan, +:s) nowait
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b, int &r)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int &s = q;
+ q = 0;
+ #pragma omp parallel for simd reduction (inscan, +:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ }
+ r = 0;
+ baz (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ }
+}
@@ -0,0 +1,122 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } }
+
+extern "C" void abort ();
+int r, a[1024], b[1024], q;
+
+#pragma omp declare reduction (foo: int: omp_out += omp_in) initializer (omp_priv = 0)
+
+__attribute__((noipa)) void
+foo (int *a, int *b, int &r)
+{
+ #pragma omp for simd reduction (inscan, foo:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int &s = q;
+ q = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, foo:s) if (0)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b, int &r)
+{
+ #pragma omp parallel for simd simdlen (1) reduction (inscan, foo:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r += a[i];
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int &s = q;
+ q = 0;
+ #pragma omp parallel for simd reduction (inscan, foo:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s += 2 * a[i];
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ }
+ r = 0;
+ baz (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += i;
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s += 2 * i;
+ if (b[i] != s)
+ abort ();
+ }
+ return 0;
+}
@@ -0,0 +1,153 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } }
+
+extern "C" void abort ();
+
+struct S {
+ inline S ();
+ inline ~S ();
+ inline S (const S &);
+ inline S & operator= (const S &);
+ int s;
+};
+
+S::S () : s (0)
+{
+}
+
+S::~S ()
+{
+}
+
+S::S (const S &x)
+{
+ s = x.s;
+}
+
+S &
+S::operator= (const S &x)
+{
+ s = x.s;
+ return *this;
+}
+
+static inline void
+ini (S &x)
+{
+ x.s = 0;
+}
+
+S r, a[1024], b[1024];
+
+#pragma omp declare reduction (+: S: omp_out.s += omp_in.s)
+#pragma omp declare reduction (plus: S: omp_out.s += omp_in.s) initializer (ini (omp_priv))
+
+__attribute__((noipa)) void
+foo (S *a, S *b, S &r)
+{
+ #pragma omp for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r.s += a[i].s;
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) S
+bar ()
+{
+ S s;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ s.s += 2 * a[i].s;
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (S *a, S *b, S &r)
+{
+ #pragma omp parallel for simd if (simd: 0) reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ r.s += a[i].s;
+ #pragma omp scan inclusive(r)
+ b[i] = r;
+ }
+}
+
+__attribute__((noipa)) S
+qux ()
+{
+ S s;
+ #pragma omp parallel for simd reduction (inscan, plus:s) simdlen(1)
+ for (int i = 0; i < 1024; i++)
+ {
+ s.s += 2 * a[i].s;
+ #pragma omp scan inclusive(s)
+ b[i] = s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ S s;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i].s = i;
+ b[i].s = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, r);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += i;
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ }
+ if (bar ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += 2 * i;
+ if (b[i].s != s.s)
+ abort ();
+ }
+ r.s = 0;
+ baz (a, b, r);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += i;
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ }
+ if (qux ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ s.s += 2 * i;
+ if (b[i].s != s.s)
+ abort ();
+ }
+}
@@ -0,0 +1,161 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { xfail *-*-* } } }
+
+extern "C" void abort ();
+
+template <typename T>
+struct S {
+ inline S ();
+ inline ~S ();
+ inline S (const S &);
+ inline S & operator= (const S &);
+ T s;
+};
+
+template <typename T>
+S<T>::S () : s (0)
+{
+}
+
+template <typename T>
+S<T>::~S ()
+{
+}
+
+template <typename T>
+S<T>::S (const S &x)
+{
+ s = x.s;
+}
+
+template <typename T>
+S<T> &
+S<T>::operator= (const S &x)
+{
+ s = x.s;
+ return *this;
+}
+
+template <typename T>
+static inline void
+ini (S<T> &x)
+{
+ x.s = 0;
+}
+
+S<int> r, a[1024], b[1024];
+
+#pragma omp declare reduction (+: S<int>: omp_out.s += omp_in.s)
+#pragma omp declare reduction (plus: S<int>: omp_out.s += omp_in.s) initializer (ini (omp_priv))
+
+template <typename T>
+__attribute__((noipa)) void
+foo (S<T> *a, S<T> *b)
+{
+ #pragma omp for simd if (0) reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r.s += a[i].s;
+ }
+}
+
+template <typename T>
+__attribute__((noipa)) S<T>
+bar (void)
+{
+ S<T> s;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s.s += 2 * a[i].s;
+ }
+ return S<T> (s);
+}
+
+__attribute__((noipa)) void
+baz (S<int> *a, S<int> *b)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r.s += a[i].s;
+ }
+}
+
+__attribute__((noipa)) S<int>
+qux (void)
+{
+ S<int> s;
+ #pragma omp parallel for simd simdlen(1) reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s.s += 2 * a[i].s;
+ }
+ return S<int> (s);
+}
+
+int
+main ()
+{
+ S<int> s;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i].s = i;
+ b[i].s = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ s.s += i;
+ }
+ if (bar<int> ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ s.s += 2 * i;
+ }
+ r.s = 0;
+ baz (a, b);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ s.s += i;
+ }
+ if (qux ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ s.s += 2 * i;
+ }
+}
@@ -0,0 +1,123 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } }
+
+extern "C" void abort ();
+int r, a[1024], b[1024], q;
+
+template <typename T, typename U>
+__attribute__((noipa)) void
+foo (T a, T b, U r)
+{
+ #pragma omp for simd if (0) reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+template <typename T>
+__attribute__((noipa)) T
+bar ()
+{
+ T &s = q;
+ q = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, +:s) simdlen(1)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+template <typename T>
+__attribute__((noipa)) void
+baz (T *a, T *b, T &r)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r)
+ for (T i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+template <typename T>
+__attribute__((noipa)) int
+qux ()
+{
+ T s = q;
+ q = 0;
+ #pragma omp parallel for simd reduction (inscan, +:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo<int *, int &> (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ s += i;
+ }
+ if (bar<int> () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ s += 2 * i;
+ }
+ r = 0;
+ baz<int> (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ s += i;
+ }
+ if (qux<int &> () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ s += 2 * i;
+ }
+}
@@ -0,0 +1,121 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { target sse2_runtime } } }
+
+extern "C" void abort ();
+int r, a[1024], b[1024], q;
+
+#pragma omp declare reduction (foo: int: omp_out += omp_in) initializer (omp_priv = 0)
+
+__attribute__((noipa)) void
+foo (int *a, int *b, int &r)
+{
+ #pragma omp for simd reduction (inscan, foo:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+__attribute__((noipa)) int
+bar (void)
+{
+ int &s = q;
+ q = 0;
+ #pragma omp parallel
+ #pragma omp for simd reduction (inscan, foo:s) nowait
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (int *a, int *b, int &r)
+{
+ #pragma omp parallel for simd reduction (inscan, foo:r) if (simd: 0)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r += a[i];
+ }
+}
+
+__attribute__((noipa)) int
+qux (void)
+{
+ int &s = q;
+ q = 0;
+ #pragma omp parallel for simd reduction (inscan, foo:s)simdlen(1)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s += 2 * a[i];
+ }
+ return s;
+}
+
+int
+main ()
+{
+ int s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i] = i;
+ b[i] = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = 25;
+ s += i;
+ }
+ if (bar () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -1;
+ s += 2 * i;
+ }
+ r = 0;
+ baz (a, b, r);
+ if (r != 1024 * 1023 / 2)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ else
+ b[i] = -25;
+ s += i;
+ }
+ if (qux () != 1024 * 1023)
+ abort ();
+ s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i] != s)
+ abort ();
+ s += 2 * i;
+ }
+}
@@ -0,0 +1,153 @@
+// { dg-require-effective-target size32plus }
+// { dg-additional-options "-O2 -fopenmp -fdump-tree-vect-details" }
+// { dg-additional-options "-mavx" { target avx_runtime } }
+// { dg-final { scan-tree-dump-times "vectorized \[2-6] loops" 2 "vect" { xfail *-*-* } } }
+
+extern "C" void abort ();
+
+struct S {
+ inline S ();
+ inline ~S ();
+ inline S (const S &);
+ inline S & operator= (const S &);
+ int s;
+};
+
+S::S () : s (0)
+{
+}
+
+S::~S ()
+{
+}
+
+S::S (const S &x)
+{
+ s = x.s;
+}
+
+S &
+S::operator= (const S &x)
+{
+ s = x.s;
+ return *this;
+}
+
+static inline void
+ini (S &x)
+{
+ x.s = 0;
+}
+
+S r, a[1024], b[1024];
+
+#pragma omp declare reduction (+: S: omp_out.s += omp_in.s)
+#pragma omp declare reduction (plus: S: omp_out.s += omp_in.s) initializer (ini (omp_priv))
+
+__attribute__((noipa)) void
+foo (S *a, S *b, S &r)
+{
+ #pragma omp for simd simdlen (1) reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r.s += a[i].s;
+ }
+}
+
+__attribute__((noipa)) S
+bar (void)
+{
+ S s;
+ #pragma omp parallel
+ #pragma omp for simd if (0) reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s.s += 2 * a[i].s;
+ }
+ return s;
+}
+
+__attribute__((noipa)) void
+baz (S *a, S *b, S &r)
+{
+ #pragma omp parallel for simd reduction (inscan, +:r)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = r;
+ #pragma omp scan exclusive(r)
+ r.s += a[i].s;
+ }
+}
+
+__attribute__((noipa)) S
+qux (void)
+{
+ S s;
+ #pragma omp parallel for simd reduction (inscan, plus:s)
+ for (int i = 0; i < 1024; i++)
+ {
+ b[i] = s;
+ #pragma omp scan exclusive(s)
+ s.s += 2 * a[i].s;
+ }
+ return s;
+}
+
+int
+main ()
+{
+ S s;
+ for (int i = 0; i < 1024; ++i)
+ {
+ a[i].s = i;
+ b[i].s = -1;
+ asm ("" : "+g" (i));
+ }
+ #pragma omp parallel
+ foo (a, b, r);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ s.s += i;
+ }
+ if (bar ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ s.s += 2 * i;
+ }
+ r.s = 0;
+ baz (a, b, r);
+ if (r.s != 1024 * 1023 / 2)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ else
+ b[i].s = 25;
+ s.s += i;
+ }
+ if (qux ().s != 1024 * 1023)
+ abort ();
+ s.s = 0;
+ for (int i = 0; i < 1024; ++i)
+ {
+ if (b[i].s != s.s)
+ abort ();
+ s.s += 2 * i;
+ }
+}
Hi! The following patch handles the last yet unsupported scan case, composite #pragma omp {,parallel }for simd ... reduction(inscan, ...) ... where we want to both parallelize and vectorize; in the first worksharing loop use normal scan support we have for #pragma omp simd ... reduction(inscan, ...) ... and just store those into the per-thread array, then do the normal worksharing loop up/down-sweep and then in the second worksharing loop use a simple #pragma omp simd that just combines the two partial sums and does the user's scan phase. Bootstrapped/regtested on x86_64-linux and i686-linux, committed to trunk. 2019-07-06 Jakub Jelinek <jakub@redhat.com> * omp-low.c (struct omp_context): Add for_simd_scan_phase member. (maybe_lookup_ctx): Add forward declaration. (omp_find_scan): Likewise. Walk into body of simd if composited with worksharing loop. (scan_omp_simd_scan): New function. (scan_omp_1_stmt): Call it. (lower_rec_simd_input_clauses): Don't create rvar nor rvar2 if ctx->for_simd_scan_phase. (lower_rec_input_clauses): Do much less work for inscan reductions in ctx->for_simd_scan_phase is_simd regions. (lower_omp_scan): Set is_simd also on simd constructs composited with worksharing loop, unless ctx->for_simd_scan_phase. Never emit a sorry message. Don't change GIMPLE_OMP_SCAN stmts into nops and emit their body after in simd constructs composited with worksharing loop. (lower_omp_for_scan): Handle worksharing loop composited with simd. * c-c++-common/gomp/scan-4.c: Don't expect sorry message. * testsuite/libgomp.c/scan-11.c: New test. * testsuite/libgomp.c/scan-12.c: New test. * testsuite/libgomp.c/scan-13.c: New test. * testsuite/libgomp.c/scan-14.c: New test. * testsuite/libgomp.c/scan-15.c: New test. * testsuite/libgomp.c/scan-16.c: New test. * testsuite/libgomp.c/scan-17.c: New test. * testsuite/libgomp.c/scan-18.c: New test. * testsuite/libgomp.c++/scan-9.C: New test. * testsuite/libgomp.c++/scan-10.C: New test. * testsuite/libgomp.c++/scan-11.C: New test. * testsuite/libgomp.c++/scan-12.C: New test. * testsuite/libgomp.c++/scan-13.C: New test. * testsuite/libgomp.c++/scan-14.C: New test. * testsuite/libgomp.c++/scan-15.C: New test. * testsuite/libgomp.c++/scan-16.C: New test. Jakub