This is the mail archive of the mailing list for the GCC project.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[committed 0/4] (Partial) OpenMP 5.0 support for GCC 9


The OpenMP 5.0 specification, ,
has been just released a few minutes ago and to celebrate that, I've merged
gomp-5_0-branch into trunk after bootstrapping/regtesting it on x86_64-linux and

Because the amount of changes in OpenMP 5.0 is much bigger than in any of the earlier
releases of the standard, unfortunately the whole spec isn't implemented at this point,
not even for C/C++.  So, let me start by listing features that are implemented.
Unless otherwise stated, the implementation is for now for C/C++ only, Fortran to follow
after C/C++ is fully done.

New OpenMP 5.0 features in this patchset:

- task reductions, including task modifier on parallel/worksharing construct reduction
- != conditions in OpenMP loops
- C++1[147] range for loops in worksharing loop, taskloop and distribute
  (and combined/composite constructs)
- allow private or lastprivate clauses for iterator variable(s) on simd construct or
  combined/composite constructs including simd
- iterators in depend clause
- support for lvalue expressions in depend clauses (note, some expressions in depend
  clauses that got allowed very recently are still unsupported)
- mutexinoutset dependence kind (right now this is implemented in the runtime library
  as less efficient inout, but can be improved later solely on in the runtime library)
- depobj construct, depobj dependence kind and omp_depend_t
- depend clause on taskwait construct
- host teams construct (the library implementation still needs work, so that it is
  actually beneficial on NUMA setups, see below)
- cancel if clause modifier
- if and nontemporal clauses on simd (both are parsed only right now, unless I figure
  out how to propagate it through IL to the vectorizer quickly, if will either force
  no simd at all, or will cause duplication of the loop; nontemporal either will
  use nontemporal stores, or for GCC 9 will do nothing)
- defaultmap clause extensions
- hint and memory order clauses on atomic (hint is parsed and ignored by the
  implementation for now, memory order fully implemented)
- memory order clauses on flush
- support for new combined #pragma omp parallel master
  and #pragma omp {,parallel }master taskloop{, simd} constructs
- affinity display support, omp_{[sg]et_affinity_format,{display,capture}_affinity},
  OMP_DISPLAY_AFFINITY and OMP_AFFINITY_FORMAT env vars (this is implemented also
  for Fortran)
- omp_pause_resource{,_all} support, omp_pause_resource_t type (for now the
  runtime library is able to free resources on the host only; this is implemented
  also for Fortran)
- worksharing loop schedules now default to nonmonotonic with the exception of
  static schedules, nonmonotonic allowed on static, runtime and auto,
  omp_sched_monotonic modifier and OMP_SCHEDULE env var parsing changes (note,
  the runtime library is told if monotonic or nonmonotonic schedule is used
  in a backwards compatible way, but the runtime library ATM doesn't take
  advantage of nonmonotonic schedules, everything is still effectively monotonic)
- allow only use_device_ptr clause(s) on target data construct
- change data sharing for readonly variables without mutable members, they are
  no longer predetermined shared (this actually changed in earlier OpenMP standard
  releases, but was considered a mistake; for 5.0 it was decided it isn't going to
  be reverted; this makes a difference mainly when using default(none))
- allow atomic constructs in simd regions (note, for now this causes the vectorizer
  to fail to vectorize such regions)
- allow comma in between (name) and hint clause on critical construct
- device routine prototype changes (void * to const void * arguments)
- omp_lock_hint* to omp_sync_hint* changes
- partial requires construct support, only atomic_default_mem_order fully implemented

Features I'll still try to implement for GCC 9:

- make sure all expressions in the OpenMP grammar within clauses are
  assignment-expression, with the exception of array section expressions
- verify taskloop construct cancellation works
- stop diagnosing threadprivate directive after first use to allow definitions of
  variables with constructors followed by threadprivate, rather than the currently
  required workaround of extern declaration of it, followed by threadprivate directive
  followed by definition

New OpenMP 5.0 features that won't be available in GCC 9, are planned for GCC 10
or later versions as time permits:

- requires directive other than atomic_default_mem_order (parsing is implemented, but
  the rest is not)
- inscan reduction clause modifier and scan directive
- lastprivate clause with conditional modifier
- OMP_TARGET_OFFLOAD env var and target-offload-var ICV
- max-active-levels-var / nested-var ICV, omp_[sg]et_nested and OMP_NESTED changes
- omp_get_supported_active_levels API addition
- array shaping support, array sections with non-unit strides in to/from/depend clauses
- metadirective support
- declare variant support
- non-rectangular loop support, support for not perfectly nested loops
- order(concurrent) clause support
- loop construct support
- affinity clause support
- detach clause support, omp_fulfill_event
- OpenMP allocator support, omp_alloc/omp_init_allocator etc., allocate directive and clause
- use_device_addr clause support
- ancestor modifier on device clause, reverse offloading
- implicit declare target support
- lvalue expressions in map/to/from clauses
- nested declare target support
- various target variable mapping changes
- declare mapper directive support
- omp_get_device_num API addition
- OMPT support
- OMPD support
- OpenMP 5.0 Fortran support (and finish up the remaining missing Fortran 4.5 features)

My short term todo list:

- requires directive other than *atomic* (currently parsed and then ignored); remove or sorry for GCC9?
- inscan modifier (currently parsed in the clause only and then ignored); remove for GCC9?
- add testsuite coverage for ordered and doacross loop with task reductions
- finish up cancelled parallel handling of worksharing reductions
- simd if (perhaps force simdlen 1 for GCC9 or duplicate loop)
- simd nontemporal (try to actually use nontemporal stores)
- host teams runtime (either implement for real for NUMA, or always use 1 team for GCC 9)
- lastprivate conditional (currently parsed and then ignored); remove or sorry for GCC9?
- check what clauses are not handled in tree-nested.c, add testsuite coverage for those
- check auto schedule what we default to, we should default to nonmonotonic
- check omp_init_*lock_with hint state
- testsuite coverage for taskloop construct cancellation
- implement mutexinoutset better than inout in the runtime (GCC 10?)


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]