This series adds new, dedicated libgomp entry points for compiler-generated
calls that previously reused general OpenMP API functions or shared entry
points for multiple purposes. The motivation is OMPT: when OMPT tracing is
enabled, callbacks need to distinguish the kinds of events being observed
(e.g. implicit vs explicit barriers, reduction critical sections vs user
atomics, masked-filter checks vs user thread-number queries). Reusing
general functions makes this impossible.
The four patches address the four cases identified in the GCC runtime:
1/4 – GOMP_has_masked_thread_num: replaces omp_get_thread_num() == filter
in the masked/master construct lowering.
2/4 – GOMP_loop_static_worksharing / GOMP_distribute_static_worksharing:
replaces the two separate omp_get_thread_num + omp_get_num_threads
(or team equivalents) calls in the static-schedule loop expander
with a single call returning both values packed in an unsigned long
long.
3/4 – GOMP_barrier gains a `kind' parameter so OMPT can report the correct
barrier kind in its sync-region callbacks.
4/4 – GOMP_reduction_start / GOMP_reduction_end: replaces GOMP_atomic_start
/ GOMP_atomic_end in the reduction critical-section path.
Paul-Antoine Arras (4):
openmp: Add GOMP_has_masked_thread_num
openmp: Add GOMP_loop_static_worksharing and
GOMP_distribute_static_worksharing
openmp: Add barrier kind to GOMP_barrier and GOMP_barrier_cancel
openmp: Add GOMP_reduction_start and GOMP_reduction_end
gcc/builtin-types.def | 1 +
gcc/c-family/c-omp.cc | 4 +-
gcc/cp/semantics.cc | 1 +
gcc/fortran/trans-openmp.cc | 4 +-
gcc/fortran/types.def | 1 +
gcc/omp-builtins.def | 17 ++-
gcc/omp-expand.cc | 86 ++++++++++-----
gcc/omp-general.cc | 5 +-
gcc/omp-general.h | 2 +-
gcc/omp-low.cc | 48 ++++++---
gcc/testsuite/c-c++-common/gomp/for-8.c | 67 ++++++++++++
.../c-c++-common/gomp/implicit-barrier-1.c | 102 ++++++++++++++++++
gcc/testsuite/c-c++-common/gomp/masked-1.c | 6 ++
.../c-c++-common/gomp/reduction-builtins-1.c | 47 ++++++++
gcc/testsuite/g++.dg/gomp/barrier-1.C | 2 +-
gcc/testsuite/g++.dg/gomp/master-3.C | 2 +-
gcc/testsuite/gcc.dg/gomp/barrier-1.c | 2 +-
gcc/testsuite/gcc.dg/gomp/master-3.c | 2 +-
.../lastprivate-allocatable-barrier-1.f90 | 33 ++++++
gcc/testsuite/gfortran.dg/gomp/masked-1.f90 | 6 +-
gcc/tree-ssa-alias.cc | 2 +
include/gomp-constants.h | 6 ++
libgomp/atomic.c | 12 +++
libgomp/barrier.c | 4 +-
libgomp/config/gcn/teams.c | 13 +++
libgomp/config/nvptx/teams.c | 12 +++
libgomp/libgomp.map | 9 ++
libgomp/libgomp_g.h | 6 +-
libgomp/parallel.c | 22 ++++
libgomp/teams.c | 13 +++
libgomp/testsuite/libgomp.c/barrier-1.c | 8 +-
31 files changed, 485 insertions(+), 60 deletions(-)
create mode 100644 gcc/testsuite/c-c++-common/gomp/for-8.c
create mode 100644 gcc/testsuite/c-c++-common/gomp/implicit-barrier-1.c
create mode 100644 gcc/testsuite/c-c++-common/gomp/reduction-builtins-1.c
create mode 100644
gcc/testsuite/gfortran.dg/gomp/lastprivate-allocatable-barrier-1.f90
--
2.53.0
