On 23/06/2026 13:07, Tobias Burnus wrote:
Hi PA,
Paul-Antoine Arras:
For static-schedule worksharing loops (`omp for schedule(static)') the
compiler emits two separate calls -- omp_get_thread_num() and
omp_get_num_threads() -- to obtain the thread id and count. Similarly,
for `omp distribute' constructs, it emits omp_get_team_num() and
omp_get_num_teams(). Because these are public OpenMP API routines, OMPT
cannot distinguish them from user-level queries.
Introduce two new libgomp entry points:
- GOMP_loop_static_worksharing(): returns both the thread id and
the thread
count packed into an unsigned long long.
- GOMP_distribute_static_worksharing(): same, but for team id and team
count.
Using a single return value rather than two output-pointer arguments
preserves optimisations in later passes.
After some discussions with a couple of persons, I think it is easier to
read and clearer to use "_Complex int" (complex_integer_type_node)
instead of the 'long long int' + shift operations. Only down side is
that that's a GNU extension.
The complex-packed approach is implemented in the updated patch attached.
Thanks,
--
PA
From c46ba873fe4cfa83809619d89695a4f138748768 Mon Sep 17 00:00:00 2001
From: Paul-Antoine Arras <[email protected]>
Date: Wed, 17 Jun 2026 16:23:00 +0200
Subject: [PATCH v2 2/4] openmp: Add GOMP_loop_static_worksharing and
GOMP_distribute_static_worksharing
For static-schedule worksharing loops (`omp for schedule(static)') the
compiler emits two separate calls -- omp_get_thread_num() and
omp_get_num_threads() -- to obtain the thread id and count. Similarly,
for `omp distribute' constructs, it emits omp_get_team_num() and
omp_get_num_teams(). Because these are public OpenMP API routines, OMPT
cannot distinguish them from user-level queries.
Introduce two new libgomp entry points:
- GOMP_loop_static_worksharing(): returns both the thread id and the thread
count packed into a single complex value.
- GOMP_distribute_static_worksharing(): same, but for team id and team
count.
Using a single return value rather than two output-pointer arguments
preserves optimisations in later passes.
gcc/ChangeLog:
* builtin-types.def (BT_COMPLEX_INT, BT_FN_COMPLEX_INT): New types.
* omp-builtins.def (BUILT_IN_GOMP_LOOP_STATIC_WORKSHARING): New
builtin.
(BUILT_IN_GOMP_DISTRIBUTE_STATIC_WORKSHARING): Likewise.
* omp-expand.cc (expand_omp_for_static_nochunk): Replace separate calls
to public API functions with a single call to
GOMP_loop_static_worksharing or GOMP_distribute_static_worksharing, then
unpack id and count.
(expand_omp_for_static_chunk): Likewise.
gcc/fortran/ChangeLog:
* types.def (BT_COMPLEX_INT): New type.
(BT_FN_COMPLEX_INT): New function type.
libgomp/ChangeLog:
* config/gcn/teams.c (GOMP_distribute_static_worksharing): New
function.
* config/nvptx/teams.c (GOMP_distribute_static_worksharing): Likewise.
* libgomp.map (GOMP_6.0.2): Export GOMP_loop_static_worksharing and
GOMP_distribute_static_worksharing.
* libgomp_g.h (GOMP_loop_static_worksharing): New declaration.
(GOMP_distribute_static_worksharing): Likewise.
* parallel.c (GOMP_loop_static_worksharing): New function.
* teams.c (GOMP_distribute_static_worksharing): Likewise.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/for-8.c: New test.
---
gcc/builtin-types.def | 3 ++
gcc/fortran/types.def | 3 ++
gcc/omp-builtins.def | 6 +++
gcc/omp-expand.cc | 66 ++++++++++++++++--------
gcc/testsuite/c-c++-common/gomp/for-8.c | 67 +++++++++++++++++++++++++
libgomp/config/gcn/teams.c | 12 +++++
libgomp/config/nvptx/teams.c | 11 ++++
libgomp/libgomp.map | 2 +
libgomp/libgomp_g.h | 2 +
libgomp/parallel.c | 12 +++++
libgomp/teams.c | 12 +++++
11 files changed, 174 insertions(+), 22 deletions(-)
create mode 100644 gcc/testsuite/c-c++-common/gomp/for-8.c
diff --git a/gcc/builtin-types.def b/gcc/builtin-types.def
index 365badca2aa..55ed6f0cae1 100644
--- a/gcc/builtin-types.def
+++ b/gcc/builtin-types.def
@@ -61,6 +61,7 @@ DEF_PRIMITIVE_TYPE (BT_VOID, void_type_node)
DEF_PRIMITIVE_TYPE (BT_BOOL, boolean_type_node)
DEF_PRIMITIVE_TYPE (BT_INT, integer_type_node)
DEF_PRIMITIVE_TYPE (BT_UINT, unsigned_type_node)
+DEF_PRIMITIVE_TYPE (BT_COMPLEX_INT, complex_integer_type_node)
DEF_PRIMITIVE_TYPE (BT_LONG, long_integer_type_node)
DEF_PRIMITIVE_TYPE (BT_ULONG, long_unsigned_type_node)
DEF_PRIMITIVE_TYPE (BT_LONGLONG, long_long_integer_type_node)
@@ -235,7 +236,9 @@ DEF_FUNCTION_TYPE_0 (BT_FN_CONST_STRING, BT_CONST_STRING)
DEF_FUNCTION_TYPE_0 (BT_FN_PID, BT_PID)
DEF_FUNCTION_TYPE_0 (BT_FN_INT, BT_INT)
DEF_FUNCTION_TYPE_0 (BT_FN_UINT, BT_UINT)
+DEF_FUNCTION_TYPE_0 (BT_FN_COMPLEX_INT, BT_COMPLEX_INT)
DEF_FUNCTION_TYPE_0 (BT_FN_ULONG, BT_ULONG)
+DEF_FUNCTION_TYPE_0 (BT_FN_ULONGLONG, BT_ULONGLONG)
DEF_FUNCTION_TYPE_0 (BT_FN_FLOAT, BT_FLOAT)
DEF_FUNCTION_TYPE_0 (BT_FN_DOUBLE, BT_DOUBLE)
/* For "long double" we use LONGDOUBLE (not LONG_DOUBLE) to
diff --git a/gcc/fortran/types.def b/gcc/fortran/types.def
index 32c533bc04f..6f98e3e48ff 100644
--- a/gcc/fortran/types.def
+++ b/gcc/fortran/types.def
@@ -47,6 +47,7 @@ along with GCC; see the file COPYING3. If not see
DEF_PRIMITIVE_TYPE (BT_VOID, void_type_node)
DEF_PRIMITIVE_TYPE (BT_BOOL, boolean_type_node)
DEF_PRIMITIVE_TYPE (BT_INT, integer_type_node)
+DEF_PRIMITIVE_TYPE (BT_COMPLEX_INT, complex_integer_type_node)
DEF_PRIMITIVE_TYPE (BT_UINT, unsigned_type_node)
DEF_PRIMITIVE_TYPE (BT_LONG, long_integer_type_node)
DEF_PRIMITIVE_TYPE (BT_ULONGLONG, long_long_unsigned_type_node)
@@ -76,8 +77,10 @@ DEF_POINTER_TYPE (BT_PTR_PTR, BT_PTR)
DEF_FUNCTION_TYPE_0 (BT_FN_BOOL, BT_BOOL)
DEF_FUNCTION_TYPE_0 (BT_FN_PTR, BT_PTR)
DEF_FUNCTION_TYPE_0 (BT_FN_INT, BT_INT)
+DEF_FUNCTION_TYPE_0 (BT_FN_COMPLEX_INT, BT_COMPLEX_INT)
DEF_FUNCTION_TYPE_0 (BT_FN_UINT, BT_UINT)
DEF_FUNCTION_TYPE_0 (BT_FN_VOID, BT_VOID)
+DEF_FUNCTION_TYPE_0 (BT_FN_ULONGLONG, BT_ULONGLONG)
DEF_FUNCTION_TYPE_1 (BT_FN_VOID_PTR, BT_VOID, BT_PTR)
DEF_FUNCTION_TYPE_1 (BT_FN_VOID_PTRPTR, BT_VOID, BT_PTR_PTR)
diff --git a/gcc/omp-builtins.def b/gcc/omp-builtins.def
index c7ddea9880a..82736e77340 100644
--- a/gcc/omp-builtins.def
+++ b/gcc/omp-builtins.def
@@ -499,3 +499,9 @@ DEF_GOMP_BUILTIN (BUILT_IN_GOMP_ERROR, "GOMP_error",
DEF_GOMP_BUILTIN (BUILT_IN_GOMP_HAS_MASKED_THREAD_NUM,
"GOMP_has_masked_thread_num", BT_FN_BOOL_INT,
ATTR_CONST_NOTHROW_LEAF_LIST)
+DEF_GOMP_BUILTIN (BUILT_IN_GOMP_LOOP_STATIC_WORKSHARING,
+ "GOMP_loop_static_worksharing", BT_FN_COMPLEX_INT,
+ ATTR_CONST_NOTHROW_LEAF_LIST)
+DEF_GOMP_BUILTIN (BUILT_IN_GOMP_DISTRIBUTE_STATIC_WORKSHARING,
+ "GOMP_distribute_static_worksharing", BT_FN_COMPLEX_INT,
+ ATTR_CONST_NOTHROW_LEAF_LIST)
diff --git a/gcc/omp-expand.cc b/gcc/omp-expand.cc
index 4d5e81f3fb1..de8889a9649 100644
--- a/gcc/omp-expand.cc
+++ b/gcc/omp-expand.cc
@@ -5190,27 +5190,38 @@ expand_omp_for_static_nochunk (struct omp_region *region,
release_ssa_name (gimple_assign_lhs (g));
}
}
+ /* Fetch the thread/team id and the number of threads/teams in a single
+ call to GOMP_loop_static_worksharing or GOMP_distribute_static_worksharing.
+ The helper returns both values packed into one complex int, with
+ the id as the imaginary part and the count as the real part. Returning
+ (rather than writing through pointers) keeps both values as plain SSA
+ names, which lets later passes - notably IPA-CP propagating constants
+ into the outlined kernel - reason about them. */
+ tree decl;
switch (gimple_omp_for_kind (fd->for_stmt))
{
case GF_OMP_FOR_KIND_FOR:
- nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS);
- threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM);
+ decl = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_STATIC_WORKSHARING);
break;
case GF_OMP_FOR_KIND_DISTRIBUTE:
- nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS);
- threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM);
+ decl = builtin_decl_explicit (BUILT_IN_GOMP_DISTRIBUTE_STATIC_WORKSHARING);
break;
default:
gcc_unreachable ();
}
- nthreads = build_call_expr (nthreads, 0);
- nthreads = fold_convert (itype, nthreads);
- nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE,
- true, GSI_SAME_STMT);
- threadid = build_call_expr (threadid, 0);
- threadid = fold_convert (itype, threadid);
- threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE,
+ {
+ tree packed = build_call_expr (decl, 0);
+ packed = force_gimple_operand_gsi (&gsi, packed, true, NULL_TREE,
true, GSI_SAME_STMT);
+ threadid = fold_build1 (IMAGPART_EXPR, integer_type_node, packed);
+ threadid = fold_convert (itype, threadid);
+ threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ nthreads = fold_build1 (REALPART_EXPR, integer_type_node, packed);
+ nthreads = fold_convert (itype, nthreads);
+ nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ }
n1 = fd->loop.n1;
n2 = fd->loop.n2;
@@ -5944,27 +5955,38 @@ expand_omp_for_static_chunk (struct omp_region *region,
release_ssa_name (gimple_assign_lhs (g));
}
}
+ /* Fetch the thread/team id and the number of threads/teams in a single
+ call to GOMP_loop_static_worksharing or GOMP_distribute_static_worksharing.
+ The helper returns both values packed into one complex int, with
+ the id as the imaginary part and the count as the real part. Returning
+ (rather than writing through pointers) keeps both values as plain SSA
+ names, which lets later passes - notably IPA-CP propagating constants
+ into the outlined kernel - reason about them. */
+ tree decl;
switch (gimple_omp_for_kind (fd->for_stmt))
{
case GF_OMP_FOR_KIND_FOR:
- nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_THREADS);
- threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_THREAD_NUM);
+ decl = builtin_decl_explicit (BUILT_IN_GOMP_LOOP_STATIC_WORKSHARING);
break;
case GF_OMP_FOR_KIND_DISTRIBUTE:
- nthreads = builtin_decl_explicit (BUILT_IN_OMP_GET_NUM_TEAMS);
- threadid = builtin_decl_explicit (BUILT_IN_OMP_GET_TEAM_NUM);
+ decl = builtin_decl_explicit (BUILT_IN_GOMP_DISTRIBUTE_STATIC_WORKSHARING);
break;
default:
gcc_unreachable ();
}
- nthreads = build_call_expr (nthreads, 0);
- nthreads = fold_convert (itype, nthreads);
- nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE,
- true, GSI_SAME_STMT);
- threadid = build_call_expr (threadid, 0);
- threadid = fold_convert (itype, threadid);
- threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE,
+ {
+ tree packed = build_call_expr (decl, 0);
+ packed = force_gimple_operand_gsi (&gsi, packed, true, NULL_TREE,
true, GSI_SAME_STMT);
+ threadid = fold_build1 (IMAGPART_EXPR, integer_type_node, packed);
+ threadid = fold_convert (itype, threadid);
+ threadid = force_gimple_operand_gsi (&gsi, threadid, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ nthreads = fold_build1 (REALPART_EXPR, integer_type_node, packed);
+ nthreads = fold_convert (itype, nthreads);
+ nthreads = force_gimple_operand_gsi (&gsi, nthreads, true, NULL_TREE,
+ true, GSI_SAME_STMT);
+ }
n1 = fd->loop.n1;
n2 = fd->loop.n2;
diff --git a/gcc/testsuite/c-c++-common/gomp/for-8.c b/gcc/testsuite/c-c++-common/gomp/for-8.c
new file mode 100644
index 00000000000..a894aaaaed1
--- /dev/null
+++ b/gcc/testsuite/c-c++-common/gomp/for-8.c
@@ -0,0 +1,67 @@
+/* { dg-do compile } */
+/* { dg-additional-options "-fdump-tree-ompexp" } */
+
+/* Check that the static loop and distribute expanders fetch the thread/team
+ id and count through a single GOMP_loop_static_worksharing /
+ GOMP_distribute_static_worksharing call rather than separate
+ omp_get_thread_num/omp_get_num_threads or omp_get_team_num/omp_get_num_teams
+ calls. */
+
+void bar (int);
+
+/* Static schedule without a chunk size goes through
+ expand_omp_for_static_nochunk. */
+
+void
+f1 (int n)
+{
+ int i;
+ #pragma omp for schedule(static)
+ for (i = 0; i < n; ++i)
+ bar (i);
+}
+
+/* Static schedule with a chunk size goes through
+ expand_omp_for_static_chunk. */
+
+void
+f2 (int n)
+{
+ int i;
+ #pragma omp for schedule(static, 4)
+ for (i = 0; i < n; ++i)
+ bar (i);
+}
+
+/* Distribute without a chunk size goes through
+ expand_omp_for_static_nochunk. */
+
+void
+f3 (int n)
+{
+ int i;
+#pragma omp teams
+#pragma omp distribute
+ for (i = 0; i < n; ++i)
+ bar (i);
+}
+
+/* Distribute with a chunk size goes through
+ expand_omp_for_static_chunk. */
+
+void
+f4 (int n)
+{
+ int i;
+#pragma omp teams
+#pragma omp distribute dist_schedule(static, 4)
+ for (i = 0; i < n; ++i)
+ bar (i);
+}
+
+/* { dg-final { scan-tree-dump "GOMP_loop_static_worksharing" "ompexp" } } */
+/* { dg-final { scan-tree-dump "GOMP_distribute_static_worksharing" "ompexp" } } */
+/* { dg-final { scan-tree-dump-not "omp_get_num_threads" "ompexp" } } */
+/* { dg-final { scan-tree-dump-not "omp_get_thread_num" "ompexp" } } */
+/* { dg-final { scan-tree-dump-not "omp_get_num_teams" "ompexp" } } */
+/* { dg-final { scan-tree-dump-not "omp_get_team_num" "ompexp" } } */
diff --git a/libgomp/config/gcn/teams.c b/libgomp/config/gcn/teams.c
index 793d4bb7838..5aa8b15502a 100644
--- a/libgomp/config/gcn/teams.c
+++ b/libgomp/config/gcn/teams.c
@@ -38,6 +38,18 @@ GOMP_teams_reg (void (*fn) (void *), void *data, unsigned int num_teams,
(void) thread_limit;
}
+/* For a distribute construct with static schedule, return the team ID and
+ number of teams packed into a single complexvalue. */
+
+_Complex int
+GOMP_distribute_static_worksharing (void)
+{
+ int __lds *gomp_team_num = (int __lds *) GOMP_TEAM_NUM;
+ unsigned tid = *gomp_team_num;
+ unsigned nteams = gomp_num_teams_var + 1;
+ return nteams + tid * 1I;
+}
+
int
omp_get_num_teams (void)
{
diff --git a/libgomp/config/nvptx/teams.c b/libgomp/config/nvptx/teams.c
index af6bf6c5710..9763f6f4108 100644
--- a/libgomp/config/nvptx/teams.c
+++ b/libgomp/config/nvptx/teams.c
@@ -41,6 +41,17 @@ GOMP_teams_reg (void (*fn) (void *), void *data, unsigned int num_teams,
(void) thread_limit;
}
+/* For a distribute construct with static schedule, return the team ID and
+ number of teams packed into a single complexvalue. */
+
+_Complex int
+GOMP_distribute_static_worksharing (void)
+{
+ unsigned tid = __gomp_team_num;
+ unsigned nteams = gomp_num_teams_var + 1;
+ return nteams + tid * 1I;
+}
+
int
omp_get_num_teams (void)
{
diff --git a/libgomp/libgomp.map b/libgomp/libgomp.map
index 5faeb27b74b..4aa518ecdae 100644
--- a/libgomp/libgomp.map
+++ b/libgomp/libgomp.map
@@ -485,6 +485,8 @@ GOMP_6.0.1 {
GOMP_6.0.2 {
global:
GOMP_has_masked_thread_num;
+ GOMP_loop_static_worksharing;
+ GOMP_distribute_static_worksharing;
} GOMP_6.0.1;
OACC_2.0 {
diff --git a/libgomp/libgomp_g.h b/libgomp/libgomp_g.h
index d36d9d36020..6dc1fd46fce 100644
--- a/libgomp/libgomp_g.h
+++ b/libgomp/libgomp_g.h
@@ -294,6 +294,7 @@ extern unsigned GOMP_parallel_reductions (void (*) (void *), void *, unsigned,
extern bool GOMP_cancel (int, bool);
extern bool GOMP_cancellation_point (int);
extern bool GOMP_has_masked_thread_num (int);
+extern __complex__ int GOMP_loop_static_worksharing (void);
/* task.c */
@@ -374,6 +375,7 @@ extern void *GOMP_target_map_indirect_ptr (void *);
extern void GOMP_teams_reg (void (*) (void *), void *, unsigned, unsigned,
unsigned);
+extern __complex__ int GOMP_distribute_static_worksharing (void);
/* allocator.c */
diff --git a/libgomp/parallel.c b/libgomp/parallel.c
index 175554de4c9..92472ecb4f0 100644
--- a/libgomp/parallel.c
+++ b/libgomp/parallel.c
@@ -271,6 +271,18 @@ GOMP_cancel (int which, bool do_cancel)
return true;
}
+/* For a worksharing-loop construct with static schedule, return the thread ID
+ and number of threads packed into a single complex value. */
+
+_Complex int
+GOMP_loop_static_worksharing (void)
+{
+ struct gomp_team *team = gomp_thread ()->ts.team;
+ unsigned tid = gomp_thread ()->ts.team_id;
+ unsigned nthreads = team ? team->nthreads : 1;
+ return nthreads + tid * 1I;
+}
+
/* Return true if the current thread number equals TID.
Used to implement the masked construct's filter clause. */
diff --git a/libgomp/teams.c b/libgomp/teams.c
index bc80eb4f2ca..7038623b2df 100644
--- a/libgomp/teams.c
+++ b/libgomp/teams.c
@@ -58,6 +58,18 @@ GOMP_teams_reg (void (*fn) (void *), void *data, unsigned int num_teams,
}
}
+/* For a distribute construct with static schedule, return the team ID and
+ number of teams packed into a single complex value. */
+
+_Complex int
+GOMP_distribute_static_worksharing (void)
+{
+ struct gomp_thread *thr = gomp_thread ();
+ unsigned tid = thr->team_num;
+ unsigned nteams = thr->num_teams + 1;
+ return nteams + tid * 1I;
+}
+
int
omp_get_num_teams (void)
{
--
2.53.0
