This patch adds support for custom allocators on private/firstprivate
clauses for task (and taskloop) constructs. Private didn't need anything
special, but firstprivate if it is passed by reference needs the GOMP_alloc
calls in the copyfn and GOMP_free in the task body.
2020-11-14 Jakub Jelinek <jakub@redhat.com>
* gimplify.c (gimplify_omp_for): Add OMP_CLAUSE_ALLOCATE_ALLOCATOR
decls as firstprivate on task clauses even when allocate clause
decl is not lastprivate.
* omp-low.c (install_var_field): Don't dereference omp_is_reference
types if mask is 33 rather than 1.
(scan_sharing_clauses): Populate allocate_map even for task
constructs. For now remove it back for variables mentioned in
reduction and in_reduction clauses on task/taskloop constructs
or on VLA task firstprivates. For firstprivate on task construct,
install the var field into field_map with by_ref and 33 instead
of false and 1 if mentioned in allocate clause.
(lower_private_allocate): Set TREE_THIS_NOTRAP on the created
MEM_REF.
(lower_rec_input_clauses): Handle allocate for task firstprivatized
non-VLA variables.
(create_task_copyfn): Likewise.
* testsuite/libgomp.c-c++-common/allocate-1.c (struct S): New type.
(foo): Add tests for non-VLA private and firstprivate clauses on
omp task.
(bar): Likewise. Remove taking of address from private/firstprivate
variables.
* testsuite/libgomp.c++/allocate-1.C (struct S): New type.
(foo): Add p, q, px and s arguments. Add tests for array reductions
and for non-VLA private and firstprivate clauses on omp task.
(bar): Removed.
(main): Adjust foo caller. Don't call bar.
Document status quo re PR94358 "[OMP] Privatize internal array variables
introduced by the Fortran FE".
libgomp/
PR fortran/94358
* testsuite/libgomp.oacc-fortran/pr94358-1.f90: New.
Co-authored-by: Thomas Schwinge <thomas@codesourcery.com>
This adds allocate clause support for array section reductions.
Furthermore, it fixes one bug that would cause inscan reductions with
allocate to be rejected by C, and for now just ignores allocate for
inscan/task reductions, that will need slightly more work.
2020-11-13 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-low.c (scan_sharing_clauses): For now remove for reduction
clauses with inscan or task modifiers decl from allocate_map.
(lower_private_allocate): Handle TYPE_P (new_var).
(lower_rec_input_clauses): Handle allocate clause for C/C++ array
reductions.
gcc/c/
* c-typeck.c (c_finish_omp_clauses): Don't clear
OMP_CLAUSE_REDUCTION_INSCAN unless reduction_seen == -2.
libgomp/
* testsuite/libgomp.c-c++-common/allocate-1.c (foo): Add tests
for array reductions.
(main): Adjust foo callers.
For now, task/taskloop constructs aren't handled and C/C++ array reductions
and reductions with task or inscan modifiers need further work.
Instead of calling omp_alloc/omp_free (where the former doesn't have
alignment argument and omp_aligned_alloc is 5.1 only feature), this calls
GOMP_alloc/GOMP_free, so that the library can fail if it would fall back
into NULL (exception is zero length allocations).
2020-11-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* builtin-types.def (BT_FN_PTR_SIZE_SIZE_PTRMODE): New function type.
* omp-builtins.def (BUILT_IN_GOACC_DECLARE): Move earlier.
(BUILT_IN_GOMP_ALLOC, BUILT_IN_GOMP_FREE): New builtins.
* gimplify.c (gimplify_scan_omp_clauses): Force allocator into a
decl if it is not NULL, INTEGER_CST or decl.
(gimplify_adjust_omp_clauses): Clear GOVD_EXPLICIT on explicit clauses
which are being removed. Remove allocate clauses for variables not seen
if they are private, firstprivate or linear too. Call
omp_notice_variable on the allocator otherwise.
(gimplify_omp_for): Handle iterator vars mentioned in allocate clauses
similarly to non-is_gimple_reg iterators.
* omp-low.c (struct omp_context): Add allocate_map field.
(delete_omp_context): Delete it.
(scan_sharing_clauses): Fill it from allocate clauses. Remove it
if mentioned also in shared clause.
(lower_private_allocate): New function.
(lower_rec_input_clauses): Handle allocate clause for privatized
variables, except for task/taskloop, C/C++ array reductions for now
and task/inscan variables.
(lower_send_shared_vars): Don't consider variables in allocate_map
as shared.
* omp-expand.c (expand_omp_for_generic, expand_omp_for_static_nochunk,
expand_omp_for_static_chunk): Use expand_omp_build_assign instead of
gimple_build_assign + gsi_insert_after.
* builtins.c (builtin_fnspec): Handle BUILTIN_GOMP_ALLOC and
BUILTIN_GOMP_FREE.
* tree-ssa-ccp.c (evaluate_stmt): Handle BUILTIN_GOMP_ALLOC.
* tree-ssa-dce.c (mark_stmt_if_obviously_necessary): Handle
BUILTIN_GOMP_ALLOC.
(mark_all_reaching_defs_necessary_1): Handle BUILTIN_GOMP_ALLOC
and BUILTIN_GOMP_FREE.
(propagate_necessity): Likewise.
gcc/fortran/
* f95-lang.c (ATTR_ALLOC_WARN_UNUSED_RESULT_SIZE_2_NOTHROW_LIST):
Define.
(gfc_init_builtin_functions): Add alloc_size and warn_unused_result
attributes to __builtin_GOMP_alloc.
* types.def (BT_PTRMODE): New primitive type.
(BT_FN_VOID_PTR_PTRMODE, BT_FN_PTR_SIZE_SIZE_PTRMODE): New function
types.
libgomp/
* libgomp.map (GOMP_alloc, GOMP_free): Export at GOMP_5.0.1.
* omp.h.in (omp_alloc): Add malloc and alloc_size attributes.
* libgomp_g.h (GOMP_alloc, GOMP_free): Declare.
* allocator.c (omp_aligned_alloc): New for now static function,
add alignment argument and handle it.
(omp_alloc): Reimplement using omp_aligned_alloc.
(GOMP_alloc, GOMP_free): New functions.
(omp_free): Add ialias.
* testsuite/libgomp.c-c++-common/allocate-1.c: New test.
* testsuite/libgomp.c++/allocate-1.C: New test.
This patch implements some parts of the target variable mapping changes
specified in OpenMP 5.0, including base-pointer attachment/detachment
behavior for array section list-items in map clauses, and ordering of
map clauses according to map kind.
2020-11-10 Chung-Lin Tang <cltang@codesourcery.com>
gcc/c-family/ChangeLog:
* c-common.h (c_omp_adjust_map_clauses): New declaration.
* c-omp.c (struct map_clause): Helper type for c_omp_adjust_map_clauses.
(c_omp_adjust_map_clauses): New function.
gcc/c/ChangeLog:
* c-parser.c (c_parser_omp_target_data): Add use of
new c_omp_adjust_map_clauses function. Add GOMP_MAP_ATTACH_DETACH as
handled map clause kind.
(c_parser_omp_target_enter_data): Likewise.
(c_parser_omp_target_exit_data): Likewise.
(c_parser_omp_target): Likewise.
* c-typeck.c (handle_omp_array_sections): Adjust COMPONENT_REF case to
use GOMP_MAP_ATTACH_DETACH map kind for C_ORT_OMP region type.
(c_finish_omp_clauses): Adjust bitmap checks to allow struct decl and
same struct field access to co-exist on OpenMP construct.
gcc/cp/ChangeLog:
* parser.c (cp_parser_omp_target_data): Add use of
new c_omp_adjust_map_clauses function. Add GOMP_MAP_ATTACH_DETACH as
handled map clause kind.
(cp_parser_omp_target_enter_data): Likewise.
(cp_parser_omp_target_exit_data): Likewise.
(cp_parser_omp_target): Likewise.
* semantics.c (handle_omp_array_sections): Adjust COMPONENT_REF case to
use GOMP_MAP_ATTACH_DETACH map kind for C_ORT_OMP region type. Fix
interaction between reference case and attach/detach.
(finish_omp_clauses): Adjust bitmap checks to allow struct decl and
same struct field access to co-exist on OpenMP construct.
gcc/ChangeLog:
* gimplify.c (is_or_contains_p): New static helper function.
(omp_target_reorder_clauses): New function.
(gimplify_scan_omp_clauses): Add use of omp_target_reorder_clauses to
reorder clause list according to OpenMP 5.0 rules. Add handling of
GOMP_MAP_ATTACH_DETACH for OpenMP cases.
* omp-low.c (is_omp_target): New static helper function.
(scan_sharing_clauses): Add scan phase handling of GOMP_MAP_ATTACH/DETACH
for OpenMP cases.
(lower_omp_target): Add lowering handling of GOMP_MAP_ATTACH/DETACH for
OpenMP cases.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/clauses-2.c: Remove dg-error cases now valid.
* gfortran.dg/gomp/map-2.f90: Likewise.
* c-c++-common/gomp/map-5.c: New testcase.
libgomp/ChangeLog:
* libgomp.h (enum gomp_map_vars_kind): Adjust enum values to be bit-flag
usable.
* oacc-mem.c (acc_map_data): Adjust gomp_map_vars argument flags to
'GOMP_MAP_VARS_OPENACC | GOMP_MAP_VARS_ENTER_DATA'.
(goacc_enter_datum): Likewise for call to gomp_map_vars_async.
(goacc_enter_data_internal): Likewise.
* target.c (gomp_map_vars_internal):
Change checks of GOMP_MAP_VARS_ENTER_DATA to use bit-and (&). Adjust use
of gomp_attach_pointer for OpenMP cases.
(gomp_exit_data): Add handling of GOMP_MAP_DETACH.
(GOMP_target_enter_exit_data): Add handling of GOMP_MAP_ATTACH.
* testsuite/libgomp.c-c++-common/ptr-attach-1.c: New testcase.
Avoid code duplication, and better test what we expect to happen.
libgomp/
PR target/85486
* testsuite/libgomp.oacc-c-c++-common/pr85486-2.c: Simplify and enhance.
* testsuite/libgomp.oacc-c-c++-common/pr85486-3.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/pr85486.c: Likewise.
This changes makes 'dg-warning', 'dg-error', 'dg-bogus', 'dg-message' behave as
expected, and also enables use of relative line numbers as well as 'dg-line'.
libgomp/
PR testsuite/80219
PR testsuite/85303
* testsuite/lib/libgomp.exp (libgomp_init): Set
'gcc_warning_prefix', 'gcc_error_prefix'.
This marks all variants of declare variant also declare target if the base
functions are called directly in target regions or declare target functions.
2020-10-28 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-offload.c (omp_declare_target_tgt_fn_r): Handle direct calls to
declare variant base functions.
libgomp/
* testsuite/libgomp.c/target-42.c: New test.
With the patch I've posted today to fix up declare variant LTO handling,
Tobias reported the patch still doesn't work, and there are two
reasons for that.
One is that when the base function is marked implicitly as declare target,
we don't mark also implicitly the variants. I'll need to ask on omp-lang
about details for that, but generally the compiler should do it some way.
The other one is that the way base_delay is written, it will always
call the usleep function, which is undesirable for nvptx. While the
compiler will replace all direct calls to base_delay to nvptx_delay,
the base_delay definition which calls usleep stays.
2020-10-28 Jakub Jelinek <jakub@redhat.com>
Tom de Vries <tdevries@suse.de>
PR testsuite/81690
* testsuite/libgomp.c/usleep.h: New file.
* testsuite/libgomp.c/target-32.c: Include usleep.h.
(main): Use tgt_usleep instead of usleep.
* testsuite/libgomp.c/thread-limit-2.c: Include usleep.h.
(main): Use tgt_usleep instead of usleep.
> I've tried to add the saving/restoring next to ipa refs saving/restoring, as
> the declare variant alt stuff is kind of extension of those, unfortunately
> following doesn't compile, because I need to also write or read a tree there
> (ctx is a portion of DECL_ATTRIBUTES of the base function), but the ipa refs
> write/read back functions don't have arguments that can be used for that.
This patch adds the streaming out and in of those omp_declare_variant_alt
hash table on the side data for the declare_variant_alt cgraph_nodes and
treats for LTO purposes the declare_variant_alt nodes (which have no body)
as if they contained a body that calls all the possible variants.
After IPA all the calls to these magic declare_variant_alt calls are
replaced with call to one of the variant depending on which one has the
highest score in the context.
2020-10-28 Jakub Jelinek <jakub@redhat.com>
PR lto/96680
gcc/
* lto-streamer.h (omp_lto_output_declare_variant_alt,
omp_lto_input_declare_variant_alt): Declare variant.
* symtab.c (symtab_node::get_partitioning_class): Return
SYMBOL_DUPLICATE for declare_variant_alt nodes.
* passes.c (ipa_write_summaries): Add declare_variant_alt to
partition.
* lto-cgraph.c (output_refs): Call omp_lto_output_declare_variant_alt
on declare_variant_alt nodes.
(input_refs): Call omp_lto_input_declare_variant_alt on
declare_variant_alt nodes.
* lto-streamer-out.c (output_function): Don't call
collect_block_tree_leafs if DECL_INITIAL is error_mark_node.
(lto_output): Call output_function even for declare_variant_alt
nodes.
* omp-general.c (omp_lto_output_declare_variant_alt,
omp_lto_input_declare_variant_alt): New functions.
gcc/lto/
* lto-common.c (lto_fixup_prevailing_decls): Don't use
LTO_NO_PREVAIL on TREE_LIST's TREE_PURPOSE.
* lto-partition.c (lto_balanced_map): Treat declare_variant_alt
nodes like definitions.
libgomp/
* testsuite/libgomp.c/declare-variant-1.c: New test.
> Therefore, I think until omp_get_initial_device () value is changed, we
The following so far untested patch implements that change.
OpenMP 4.5 said for omp_get_initial_device:
The value of the device number is implementation defined. If it is between 0 and one less than
omp_get_num_devices() then it is valid for use with all device constructs and routines; if it is
outside that range, then it is only valid for use with the device memory routines and not in the
device clause.
and OpenMP 5.0 similarly, but OpenMP 5.1 says:
The value of the device number is the value returned by the omp_get_num_devices routine.
As the new value is compatible with what has been required earlier, I think
we can change it already now.
2020-10-22 Jakub Jelinek <jakub@redhat.com>
* icv.c (omp_get_initial_device): Remove including corresponding
ialias.
* icv-device.c (omp_get_initial_device): New function. Return
gomp_get_num_devices (). Add ialias.
* target.c (resolve_device): Don't fail with
OMP_TARGET_OFFLOAD=mandatory if device_id is equal to
gomp_get_num_devices ().
(omp_target_alloc, omp_target_free, omp_target_is_present,
omp_target_memcpy, omp_target_memcpy_rect, omp_target_associate_ptr,
omp_target_disassociate_ptr, omp_pause_resource): Use
gomp_get_num_devices () instead of GOMP_DEVICE_HOST_FALLBACK on the
first use in the functions, in uses dominated by the
gomp_get_num_devices call use num_devices_openmp instead.
* libgomp.texi (omp_get_initial_device): Document.
* config/gcn/icv-device.c (omp_get_initial_device): New function.
Add ialias.
* config/nvptx/icv-device.c (omp_get_initial_device): Likewise.
* testsuite/libgomp.c/target-40.c: New test.
> the patch also breaks bootstrap on both i386-pc-solaris2.11 and
> sparc-sun-solaris2.11:
>
> /vol/gcc/src/hg/master/local/libgomp/env.c: In function 'initialize_env':
> /vol/gcc/src/hg/master/local/libgomp/env.c:414:16: error: 'new_offload' may be used uninitialized in this function [-Werror=maybe-uninitialized]
> 414 | *offload = new_offload;
> | ~~~~~~~~~^~~~~~~~~~~~~
> /vol/gcc/src/hg/master/local/libgomp/env.c:384:30: note: 'new_offload' was declared here
> 384 | enum gomp_target_offload_t new_offload;
> | ^~~~~~~~~~~
I can't reproduce that, but I fail to see why we need two separate
variables, one with actual value and one tracking if the value is valid.
So, I'm going with:
2020-10-21 Jakub Jelinek <jakub@redhat.com>
* env.c (parse_target_offload): Change new_offload var type to int,
preinitialize to -1, remove found var and test new_offload != -1
instead of found.
> On 10/20/20 2:11 PM, Tobias Burnus wrote:
>
> > Unfortunately, the committed patch
> > (r11-4121-g1bfc07d150790fae93184a79a7cce897655cb37b)
> > causes build errors.
> >
> > The error seems to be provoked by function cloning – as the code
> > itself looks fine:
> > ...
> > struct gomp_device_descr *devices_s
> > = malloc (num_devices * sizeof (struct gomp_device_descr));
> > ...
> > for (i = 0; i < num_devices; i++)
> > if (!(devices[i].capabilities & GOMP_OFFLOAD_CAP_OPENMP_400))
> > devices_s[num_devices_after_openmp++] = devices[i];
>
> gomp_target_init.part.0 ()
> {
> ...
> <bb 2>
> devices_s_1 = malloc (0);
> ...
> num_devices.16_67 = num_devices;
> ...
> if (num_devices.16_67 > 0)
> goto <bb 3>; [89.00%]
> else
> goto <bb 18>; [11.00%]
>
> Which seems to have an ordering problem.
This patch fixes the warning that breaks the bootstrap.
2020-10-20 Jakub Jelinek <jakub@redhat.com>
* target.c (gomp_target_init): Inside of the function, use automatic
variables corresponding to num_devices, num_devices_openmp and devices
global variables and update the globals only at the end of the
function.
This implements support for the OMP_TARGET_OFFLOAD environment variable
introduced in the OpenMP 5.0 standard, which controls how offloading
is handled. It may be set to MANDATORY (abort if offloading cannot be
performed), DISABLED (no offloading to devices) or DEFAULT (offload to
device if possible, fall back to host if not).
2020-10-20 Kwok Cheung Yeung <kcy@codesourcery.com>
libgomp/
* env.c (gomp_target_offload_var): New.
(parse_target_offload): New.
(handle_omp_display_env): Print value of OMP_TARGET_OFFLOAD.
(initialize_env): Parse OMP_TARGET_OFFLOAD.
* libgomp.h (gomp_target_offload_t): New.
(gomp_target_offload_var): New.
* libgomp.texi (OMP_TARGET_OFFLOAD): New section.
* target.c (resolve_device): Generate error if device not found and
offloading is mandatory.
(gomp_target_fallback): Generate error if offloading is mandatory.
(GOMP_target): Add argument in call to gomp_target_fallback.
(GOMP_target_ext): Likewise.
(gomp_target_data_fallback): Generate error if offloading is mandatory.
(GOMP_target_data): Add argument in call to gomp_target_data_fallback.
(GOMP_target_data_ext): Likewise.
(gomp_target_task_fn): Add argument in call to gomp_target_fallback.
(gomp_target_init): Return early if offloading is disabled.
This patch implements the omp_get_supported_active_levels runtime routine
from the OpenMP 5.0 specification, which returns the maximum number of
active nested parallel regions supported by this implementation. The
current maximum (set using the omp_set_max_active_levels routine or the
OMP_MAX_ACTIVE_LEVELS environment variable) cannot exceed this number.
2020-10-13 Kwok Cheung Yeung <kcy@codesourcery.com>
libgomp/
* env.c (gomp_max_active_levels_var): Initialize to
gomp_supported_active_levels.
(initialize_env): Limit gomp_max_active_levels_var to be at most
equal to gomp_supported_active_levels.
* fortran.c (omp_get_supported_active_levels): Add ialias_redirect.
(omp_get_supported_active_levels_): New.
* icv.c (omp_set_max_active_levels): Limit gomp_max_active_levels_var
to at most equal to gomp_supported_active_levels.
(omp_get_supported_active_levels): New.
* libgomp.h (gomp_supported_active_levels): New.
* libgomp.map (OMP_5.0.1): Add omp_get_supported_active_levels and
omp_get_supported_active_levels_.
* libgomp.texi (omp_get_supported_active_levels): New.
(omp_set_max_active_levels): Update. Add reference to
omp_get_supported_active_levels.
* omp.h.in (omp_get_supported_active_levels): New.
* omp_lib.f90.in (omp_get_supported_active_levels): New.
* omp_lib.h.in (omp_get_supported_active_levels): New.
* testsuite/libgomp.c/lib-2.c (main): Check omp_get_max_active_levels
against omp_get_supported_active_levels.
* testsuite/libgomp.fortran/lib4.f90 (lib4): Likewise.
AIX caches shared objects in archives with read-other permission.
libgomp and libatomic might be in use during the build or testing, which
may cause archiver operations on them to fail. This patch adjusts the
Makefile fragments to delete the library archives before creating fresh
archives containing both the 32 bit and 64 bit shared objects.
libatomic/ChangeLog:
2020-10-11 Clement Chigot <clement.chigot@atos.net>
* config/t-aix: Delete and recreate libatomic before creating
FAT library.
libgomp/ChangeLog:
2020-10-11 Clement Chigot <clement.chigot@atos.net>
* config/t-aix: Delete and recreate libgomp before creating
FAT library.
Using this patch, when using GOMP_DEBUG=1 and launching a kernel in
GOMP_OFFLOAD_run (used by the omp implementation), we see the kernel launch
dimensions:
...
GOMP_OFFLOAD_run: kernel main$_omp_fn$0: \
launch [(teams: 1), 1, 1] [(lanes: 32), (threads: 1), 1]
...
Build on x86_64-linux with nvptx accelerator, tested libgomp.
libgomp/ChangeLog:
2020-10-08 Tom de Vries <tdevries@suse.de>
PR libgomp/81802
* plugin/plugin-nvptx.c (GOMP_OFFLOAD_run): Report launch
dimensions.
Consider test-case test.c, with VLA A:
...
int main (void) {
int N = 1000;
int A[N];
#pragma acc declare copy(A)
return 0;
}
...
compiled using:
...
$ gcc test.c -fopenacc -S -fdump-tree-all
...
At original, we have:
...
#pragma acc declare map(tofrom:A);
...
but at gimple, we have a map (to:A.1), but not a map (from:A.1):
...
int[0:D.2074] * A.1;
{
int A[0:D.2074] [value-expr: *A.1];
saved_stack.2 = __builtin_stack_save ();
try
{
A.1 = __builtin_alloca_with_align (D.2078, 32);
#pragma omp target oacc_declare map(to:(*A.1) [len: D.2076])
}
finally
{
__builtin_stack_restore (saved_stack.2);
}
}
...
This is caused by the following incompatibility. When storing the desired
from clause in oacc_declare_returns, we use 'A.1' as the key:
...
10898 oacc_declare_returns->put (decl, c);
(gdb) call debug_generic_expr (decl)
A.1
(gdb) call debug_generic_expr (c)
map(from:(*A.1))
...
but when looking it up, we use 'A' as the key:
...
(gdb)
1471 tree *c = oacc_declare_returns->get (t);
(gdb) call debug_generic_expr (t)
A
...
Fix this by extracing the 'A.1' lookup key from 'A' using the decl-expr.
In addition, unshare the looked up value, to fix avoid running into
an "incorrect sharing of tree nodes" error.
Using these two fixes, we get our desired:
...
finally
{
+ #pragma omp target oacc_declare map(from:(*A.1))
__builtin_stack_restore (saved_stack.2);
}
...
Build on x86_64-linux with nvptx accelerator, tested libgomp.
gcc/ChangeLog:
2020-10-06 Tom de Vries <tdevries@suse.de>
PR middle-end/90861
* gimplify.c (gimplify_bind_expr): Handle lookup in
oacc_declare_returns using key with decl-expr.
libgomp/ChangeLog:
2020-10-06 Tom de Vries <tdevries@suse.de>
PR middle-end/90861
* testsuite/libgomp.oacc-c-c++-common/declare-vla.c: Remove xfail.
When running the libgomp testsuite on x86_64-linux with nvptx accelerator on
the test-case included in this patch, we run into:
...
FAIL: libgomp.fortran/pr95654.f90 -O3 -fomit-frame-pointer -funroll-loops \
-fpeel-loops -ftracer -finline-functions execution test
...
The test-case is a minimal version of this FAIL:
...
FAIL: libgomp.fortran/pr66199-5.f90 -O3 -fomit-frame-pointer -funroll-loops \
-fpeel-loops -ftracer -finline-functions execution test
...
but that one has stopped failing at commit c2ebf4f10d "openmp: Add support
for non-rect simd and improve collapsed simd support".
The problem is that ftracer duplicates a block containing GOMP_SIMT_VOTE_ANY.
That is, before ftracer we have (dropping the GOMP_SIMT_ prefix):
...
bb4(ENTER_ALLOC)
*----------+
| \
| \
| v
| *
v bb8
*<------------*
bb5(VOTE_ANY)
*-------------+
| |
| |
| |
| |
| v
| *
v bb7(XCHG_IDX)
*<------------*
bb6(EXIT)
...
The XCHG_IDX internal-fn does inter-SIMT-lane communication, which for nvptx
maps onto shfl, an operator which has the requirement that the warp executing
the operator is convergent. The warp diverges at bb4, and
reconverges at bb5, and does not diverge by going to bb7, so the shfl is
indeed executed by a convergent warp.
After ftracer, we have:
...
bb4(ENTER_ALLOC)
*----------+
| \
| \
| \
| \
v v
* *
bb5(VOTE_ANY) bb8(VOTE_ANY)
* *
|\ /|
| \ +--------+ |
| \/ |
| /\ |
| / +----------v
|/ *
v bb7(XCHG_IDX)
*<--------------*
bb6(EXIT)
...
The warp diverges again at bb5, but does not reconverge again before bb6, so
the shfl is executed by a divergent warp, which causes the FAIL.
Fix this by making ftracer ignore blocks containing ENTER_ALLOC, VOTE_ANY and
EXIT, effectively treating the SIMT region conservatively.
An argument can be made that the test needs to be added in a more
generic place, like gimple_can_duplicate_bb_p or some such, and that ftracer
then needs to use the generic test. But that's a discussion with a much
broader scope, so I'm leaving that for another patch.
Bootstrapped and reg-tested on x86_64-linux.
Build on x86_64-linux with nvptx accelerator, tested with libgomp.
gcc/ChangeLog:
PR fortran/95654
* tracer.c (ignore_bb_p): Ignore GOMP_SIMT_ENTER_ALLOC,
GOMP_SIMT_VOTE_ANY and GOMP_SIMT_EXIT.
libgomp/ChangeLog:
2020-10-05 Tom de Vries <tdevries@suse.de>
PR fortran/95654
* testsuite/libgomp.fortran/pr95654.f90: New test.
Accelerators with fixed thread-counts will break if nested teams are expected
to have multiple threads each.
libgomp/ChangeLog:
2020-09-29 Andrew Stubbs <ams@codesourcery.com>
* parallel.c (gomp_resolve_num_threads): Ignore nest_var on nvptx
and amdgcn targets.
Both GCN and NVPTX allow nested parallel regions, but the barrier
implementation did not allow the nested teams to run independently of each
other (due to hardware limitations). This patch fixes that, under the
assumption that each thread will create a new subteam of one thread, by
simply not using barriers when there's no other thread to synchronise.
libgomp/ChangeLog:
* config/gcn/bar.c (gomp_barrier_wait_end): Skip the barrier if the
total number of threads is one.
(gomp_team_barrier_wake): Likewise.
(gomp_team_barrier_wait_end): Likewise.
(gomp_team_barrier_wait_cancel_end): Likewise.
* config/nvptx/bar.c (gomp_barrier_wait_end): Likewise.
(gomp_team_barrier_wake): Likewise.
(gomp_team_barrier_wait_end): Likewise.
(gomp_team_barrier_wait_cancel_end): Likewise.
* testsuite/libgomp.c-c++-common/nested-parallel-unbalanced.c: New test.
