Yesterday when working on numa_domains, I've noticed because of a bug
in my patch a hang on a large NUMA machine. I've fixed the bug, but
also discovered that the hang was a result of making wrong assumptions
about strtoul/strtoull. All the uses were for portability setting
errno = 0 before the calls and treating non-zero errno after the call
as invalid input, but for the case where there are no valid digits at
all strtoul may set errno to EINVAL, but doesn't have to and with
glibc doesn't do that. So, this patch goes through all the strtoul calls
and next to errno != 0 checks adds also endptr == startptr check.
Haven't done it in places where we immediately reject strtoul returning 0
the same as we reject errno != 0, because strtoul must return 0 in the
case where it sets endptr to the start pointer. In some spots the code
was using errno = 0; x = strtoul (p, &p, 10); if (errno) { /*invalid*/ }
and those spots had to be changed to
errno = 0; x = strtoul (p, &end, 10); if (errno || end == p) { /*invalid*/ }
p = end;
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* env.c (parse_schedule): For strtoul or strtoull calls which don't
clearly reject return value 0 as invalid handle the case where end
pointer is the same as first argument as invalid.
(parse_unsigned_long_1): Likewise.
(parse_one_place): Likewise.
(parse_places_var): Likewise.
(parse_stacksize): Likewise.
(parse_spincount): Likewise.
(parse_affinity): Likewise.
(parse_gomp_openacc_dim): Likewise. Avoid strict aliasing violation.
Make code valid C89.
* config/linux/affinity.c (gomp_affinity_find_last_cache_level):
For strtoul calls which don't clearly reject return value 0 as
invalid handle the case where end pointer is the same as first
argument as invalid.
(gomp_affinity_init_level_1): Likewise.
(gomp_affinity_init_numa_domains): Likewise.
* config/rtems/proc.c (parse_thread_pools): Likewise.
When writing the places-*.c tests, I've noticed that we mishandle threads
abstract name with specified num-places if num-places isn't a multiple of
number of hw threads in a core. It then happily ignores the maximum count
and overwrites for the remaining hw threads in a core further places that
haven't been allocated.
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* config/linux/affinity.c (gomp_affinity_init_level_1): For level 1
after creating count places clean up and return immediately.
* testsuite/libgomp.c/places-6.c: New test.
* testsuite/libgomp.c/places-7.c: New test.
* testsuite/libgomp.c/places-8.c: New test.
* testsuite/libgomp.c/places-9.c: New test.
* testsuite/libgomp.c/places-10.c: New test.
This adds support for numa_domains abstract name in OMP_PLACES, also new
in OpenMP 5.1.
Way to test this is
OMP_PLACES=numa_domains OMP_DISPLAY_ENV=true LD_PRELOAD=.libs/libgomp.so.1 /bin/true
and see what it prints on OMP_PLACES line.
For non-NUMA machines it should print a single place that covers all CPUs,
for NUMA machine one place for each NUMA node with corresponding CPUs.
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* env.c (parse_places_var): Handle numa_domains as level 5.
* config/linux/affinity.c (gomp_affinity_init_numa_domains): New
function.
(gomp_affinity_init_level): Use it instead of
gomp_affinity_init_level_1 for level == 5.
* testsuite/libgomp.c/places-5.c: New test.
This patch implements support for ll_caches abstract name in OMP_PLACES,
which stands for places where logical cpus in each place share the last
level cache.
This seems to work fine for me on x86 and kernel sources show that it is
in common code, but on some machines on CompileFarm the files I'm using,
i.e.
/sys/devices/system/cpu/cpuN/cache/indexN/level
/sys/devices/system/cpu/cpuN/cache/indexN/shared_cpu_list
don't exist, is that because they have too old kernel and newer kernels
are fine or should I implement some fallback methods (which)?
E.g. on gcc112.fsffrance.org I see just shared_cpu_map and not shared_cpu_list
(with shared_cpu_map being harder to parse) and on another box I didn't even
see the cache subdirectories.
Way to test this is
OMP_PLACES=ll_caches OMP_DISPLAY_ENV=true LD_PRELOAD=.libs/libgomp.so.1 /bin/true
and see what it prints on OMP_PLACES line.
2021-10-15 Jakub Jelinek <jakub@redhat.com>
* env.c (parse_places_var): Handle ll_caches as level 4.
* config/linux/affinity.c (gomp_affinity_find_last_cache_level): New
function.
(gomp_affinity_init_level_1): Handle level 4 as logical cpus sharing
last level cache.
(gomp_affinity_init_level): Likewise.
* testsuite/libgomp.c/places-1.c: New test.
* testsuite/libgomp.c/places-2.c: New test.
* testsuite/libgomp.c/places-3.c: New test.
* testsuite/libgomp.c/places-4.c: New test.
This patch releases the device lock on a sanity-checking error path in
transfer combining (cbuf) handling in libgomp:target.c. This shouldn't
happen when handling well-formed mapping clauses, but erroneous clauses
can currently cause a hang if the condition triggers.
2021-12-10 Julian Brown <julian@codesourcery.com>
libgomp/
* target.c (gomp_copy_host2dev): Release device lock on cbuf
error path.
This adds the Fortran testsuite coverage of
omp_{get_max,set_num}_threads and omp_{s,g}et_teams_thread_limit
libgomp/
* testsuite/libgomp.fortran/icv-3.f90: New.
* testsuite/libgomp.fortran/icv-4.f90: New.
This patch adds documentation for these new OpenMP 5.1 APIs as well as
two new environment variables - OMP_NUM_TEAMS and OMP_TEAMS_THREAD_LIMIT.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (omp_get_max_teams, omp_get_teams_thread_limit,
omp_set_num_teams, omp_set_teams_thread_limit, OMP_NUM_TEAMS,
OMP_TEAMS_THREAD_LIMIT): Document.
When building libgomp documentation, I see
makeinfo --split-size=5000000 -I ../../../libgomp/../gcc/doc/include -I ../../../libgomp -o libgomp.info ../../../libgomp/libgomp.texi
../../../libgomp/libgomp.texi:503: warning: node next `omp_get_default_device' in menu `omp_get_device_num' and in sectioning `omp_get_dynamic' differ
../../../libgomp/libgomp.texi:528: warning: node prev `omp_get_dynamic' in menu `omp_get_device_num' and in sectioning `omp_get_default_device' differ
../../../libgomp/libgomp.texi:560: warning: node next `omp_get_initial_device' in menu `omp_get_level' and in sectioning `omp_get_device_num' differ
../../../libgomp/libgomp.texi:587: warning: node next `omp_get_device_num' in menu `omp_get_dynamic' and in sectioning `omp_get_level' differ
../../../libgomp/libgomp.texi:587: warning: node prev `omp_get_device_num' in menu `omp_get_default_device' and in sectioning `omp_get_initial_device' differ
../../../libgomp/libgomp.texi:615: warning: node prev `omp_get_level' in menu `omp_get_initial_device' and in sectioning `omp_get_device_num' differ
warnings. This patch fixes those.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (omp_get_device_num): Move @node before omp_get_dynamic
to avoid makeinfo warnings.
This adds (C/C++ only) testsuite coverage for these new OpenMP 5.1 APIs.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
* testsuite/libgomp.c-c++-common/icv-3.c: New test.
* testsuite/libgomp.c-c++-common/icv-4.c: New test.
As reported, the alloc-9.c test and alloc-{1,2,3}.F* and alloc-11.f90
tests fail on powerpc64-linux with -m32.
The reason why it fails just there is that malloc doesn't guarantee there
128-bit alignment (historically glibc guaranteed 2 * sizeof (void *)
alignment from malloc).
There are two separate issues.
One is a thinko on my side.
In this part of alloc-9.c test (copied to alloc-11.f90), we have
2 allocators, a with pool size 1024B and alignment 16B and default fallback
and a2 with pool size 512B and alignment 32B and a as fallback allocator.
We start at no allocations in both at line 194 and do:
p = (int *) omp_alloc (sizeof (int), a2);
// This succeeds in a2 and needs 4+overhead bytes (which includes the 32B alignment)
p = (int *) omp_realloc (p, 420, a, a2);
// This allocates 420 bytes+overhead in a, with 16B alignment and deallocates the above
q = (int *) omp_alloc (sizeof (int), a);
// This allocates 4+overhead bytes in a, with 16B alignment
q = (int *) omp_realloc (q, 420, a2, a);
// This allocates 420+overhead in a2 with 32B alignment
q = (int *) omp_realloc (q, 768, a2, a2);
// This attempts to reallocate, but as there are elevated alignment
// requirements doesn't try to just realloc (even if it wanted to try that
// a2 is almost full, with 512-420-overhead bytes left in it), so it
// tries to alloc in a2, but there is no space left in the pool, falls
// back to a, which already has 420+overhead bytes allocated in it and
// 1024-420-overhead bytes left and so fails too and fails to default
// non-pool allocator that allocates it, but doesn't guarantee alignment
// higher than malloc guarantees.
// But, the test expected 16B alignment.
So, I've slightly lowered the allocation sizes in that part of the test
420->320 and 768 -> 568, so that the last test still fails to allocate
in a2 (568 > 512-320-overhead) but succeeds in a as fallback, which was
the intent of the test.
Another thing is that alloc-1.F90 seems to be transcription of
libgomp.c-c++-common/alloc-1.c into Fortran, but alloc-1.c had:
q = (int *) omp_alloc (768, a2);
if ((((uintptr_t) q) % 16) != 0)
abort ();
q[0] = 7;
q[767 / sizeof (int)] = 8;
r = (int *) omp_alloc (512, a2);
if ((((uintptr_t) r) % __alignof (int)) != 0)
abort ();
there but Fortran has:
cq = omp_alloc (768_c_size_t, a2)
if (mod (transfer (cq, intptr), 16_c_intptr_t) /= 0) stop 12
call c_f_pointer (cq, q, [768 / c_sizeof (i)])
q(1) = 7
q(768 / c_sizeof (i)) = 8
cr = omp_alloc (512_c_size_t, a2)
if (mod (transfer (cr, intptr), 16_c_intptr_t) /= 0) stop 13
I'm changing the latter to 4_c_intptr_t because other spots in the
testcase do that, Fortran sadly doesn't have c_alignof, but strictly
speaking it isn't correct, __alignof (int) could be on some architectures
smaller than 4.
So probably alloc-1.F90 etc. should also have
! { dg-additional-sources alloc-7.c }
! { dg-prune-output "command-line option '-fintrinsic-modules-path=.*' is valid for Fortran but not for C" }
and use get__alignof_int.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
PR libgomp/102628
PR libgomp/102668
* testsuite/libgomp.c-c++-common/alloc-9.c (main): Decrease
allocation sizes from 420 to 320 and from 768 to 568.
* testsuite/libgomp.fortran/alloc-11.f90: Likewise.
* testsuite/libgomp.fortran/alloc-1.F90: Change expected alignment
for cr from 16 to 4.
> * testsuite/libgomp.c++/scan-10.C: Add option -fvect-cost-model=cheap.
I don't think this is the right thing to do.
This just means that at some point between 2013 when -fsimd-cost-model has
been introduced and now -fsimd-cost-model= option at least partially stopped
working properly.
As documented, -fsimd-cost-model= overrides the -fvect-cost-model= setting
for OpenMP simd loops (loop->force_vectorize is true) if specified differently
from default.
In tree-vectorizer.h we have:
static inline bool
unlimited_cost_model (loop_p loop)
{
if (loop != NULL && loop->force_vectorize
&& flag_simd_cost_model != VECT_COST_MODEL_DEFAULT)
return flag_simd_cost_model == VECT_COST_MODEL_UNLIMITED;
return (flag_vect_cost_model == VECT_COST_MODEL_UNLIMITED);
}
and use it in various places, but we also just use flag_vect_cost_model
in lots of places (and in one spot use flag_simd_cost_model, not sure if
we are sure it is a force_vectorize loop or what).
So, IMHO we should change the above inline function to
loop_cost_model and let it return the cost model and then just
reimplement unlimited_cost_model as
return loop_cost_model (loop) == VECT_COST_MODEL_UNLIMITED;
and then adjust the direct uses of the flag and revert these changes.
2021-10-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree-vectorizer.h (loop_cost_model): New function.
(unlimited_cost_model): Use it.
* tree-vect-loop.c (vect_analyze_loop_costing): Use loop_cost_model
call instead of flag_vect_cost_model.
* tree-vect-data-refs.c (vect_enhance_data_refs_alignment): Likewise.
(vect_prune_runtime_alias_test_list): Likewise. Also use it instead
of flag_simd_cost_model.
gcc/testsuite/
* gcc.dg/gomp/simd-2.c: Remove option -fvect-cost-model=cheap.
* gcc.dg/gomp/simd-3.c: Likewise.
libgomp/
* testsuite/libgomp.c/scan-11.c: Remove option -fvect-cost-model=cheap.
* testsuite/libgomp.c/scan-12.c: Likewise.
* testsuite/libgomp.c/scan-13.c: Likewise.
* testsuite/libgomp.c/scan-14.c: Likewise.
* testsuite/libgomp.c/scan-15.c: Likewise.
* testsuite/libgomp.c/scan-16.c: Likewise.
* testsuite/libgomp.c/scan-17.c: Likewise.
* testsuite/libgomp.c/scan-18.c: Likewise.
* testsuite/libgomp.c/scan-19.c: Likewise.
* testsuite/libgomp.c/scan-20.c: Likewise.
* testsuite/libgomp.c/scan-21.c: Likewise.
* testsuite/libgomp.c/scan-22.c: Likewise.
* testsuite/libgomp.c++/scan-9.C: Likewise.
* testsuite/libgomp.c++/scan-10.C: Likewise.
* testsuite/libgomp.c++/scan-11.C: Likewise.
* testsuite/libgomp.c++/scan-12.C: Likewise.
* testsuite/libgomp.c++/scan-13.C: Likewise.
* testsuite/libgomp.c++/scan-14.C: Likewise.
* testsuite/libgomp.c++/scan-15.C: Likewise.
* testsuite/libgomp.c++/scan-16.C: Likewise.
This issue was observed in rs6000 specific PR102658 as well.
I've looked into it a bit, it's caused by the "conditional store replacement" which
is originally disabled without vectorization as below code.
/* If either vectorization or if-conversion is disabled then do
not sink any stores. */
if (param_max_stores_to_sink == 0
|| (!flag_tree_loop_vectorize && !flag_tree_slp_vectorize)
|| !flag_tree_loop_if_convert)
return false;
The new change makes the innermost loop look like
for (int c1 = 0; c1 <= 1499; c1 += 1) {
if (c1 <= 500) {
S_10(c0, c1);
} else {
S_9(c0, c1);
}
S_11(c0, c1);
}
and can not be splitted as:
for (int c1 = 0; c1 <= 500; c1 += 1)
S_10(c0, c1);
for (int c1 = 501; c1 <= 1499; c1 += 1)
S_9(c0, c1);
So instead of disabling vectorization, could we just disable this cs replacement
with parameter "--param max-stores-to-sink=0"?
I tested this proposal on ppc64le, it should work as well.
2021-10-11 Kewen Lin <linkw@linux.ibm.com>
libgomp/ChangeLog:
* testsuite/libgomp.graphite/force-parallel-8.c: Add --param max-stores-to-sink=0.
The variable omp_atv_sequential was replaced by omp_atv_serialized in OpenMP
5.1. This was already implemented by Jakub (C/C++, commit ea82325afe) and
Tobias (Fortran, commit fff15bad1a).
This patch adds two tests to check if omp_atv_serialized is available (one test
for C/C++ and one for Fortran). Besides that omp_atv_sequential is marked as
deprecated in C/C++ and Fortran for OpenMP 5.1.
libgomp/ChangeLog:
* allocator.c (omp_init_allocator): Replace omp_atv_sequential with
omp_atv_serialized.
* omp.h.in: Add deprecated flag for omp_atv_sequential.
* omp_lib.f90.in: Add deprecated flag for omp_atv_sequential.
* testsuite/libgomp.c-c++-common/alloc-10.c: New test.
* testsuite/libgomp.fortran/alloc-12.f90: New test.
OpenMP 5.1 adds env vars and functions to set and query new ICVs used
as fallback if thread_limit or num_teams clauses aren't specified on
teams construct.
The following patch implements those, though further work will be needed:
1) OpenMP 5.1 also changed the num_teams clause, so that it can specify
both lower and upper limit for how many teams should be created and
changed the meaning when only one expression is provided, instead of
num_teams(expr) in 5.0 meaning num_teams(1:expr) in 5.1, it now means
num_teams(expr:expr), i.e. while previously we could create 1 to expr
teams, in 5.1 we have some low limit by default equal to the single
expression provided and may not create fewer teams.
For host teams (which we don't currently implement efficiently for
NUMA hosts) we trivially satisfy it now by always honoring what the
user asked for, but for the offloading teams I think we'll need to
rethink the APIs; currently teams construct is just a call that returns
and possibly lowers the number of teams; and whenever possible we try
to evaluate num_teams/thread_limit already on the target construct
and the GOMP_teams call just sets the number of teams to the minimum
of provided and requested teams; for some cases e.g. where target
is not combined with teams and num_teams expression calls some functions
etc., we need to call those functions in the target region and so it is
late to figure number of teams, but also hw could just limit what it
is willing to create; in that case I'm afraid we need to run the target
body multiple times and arrange for omp_get_team_num () returning the
right values
2) we need to finally implement the NUMA handling for GOMP_teams_reg
3) I now realize I haven't added some testcase coverage, will do that
incrementally
4) libgomp.texi needs updates for these new APIs, but also others like
the allocator
2021-10-11 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-low.c (omp_runtime_api_call): Handle omp_get_max_teams,
omp_[sg]et_teams_thread_limit and omp_set_num_teams.
libgomp/
* omp.h.in (omp_set_num_teams, omp_get_max_teams,
omp_set_teams_thread_limit, omp_get_teams_thread_limit): Declare.
* omp_lib.f90.in (omp_set_num_teams, omp_get_max_teams,
omp_set_teams_thread_limit, omp_get_teams_thread_limit): Declare.
* omp_lib.h.in (omp_set_num_teams, omp_get_max_teams,
omp_set_teams_thread_limit, omp_get_teams_thread_limit): Declare.
* libgomp.h (gomp_nteams_var, gomp_teams_thread_limit_var): Declare.
* libgomp.map (OMP_5.1): Export omp_get_max_teams{,_},
omp_get_teams_thread_limit{,_}, omp_set_num_teams{,_,_8_} and
omp_set_teams_thread_limit{,_,_8_}.
* icv.c (omp_set_num_teams, omp_get_max_teams,
omp_set_teams_thread_limit, omp_get_teams_thread_limit): New
functions.
* env.c (gomp_nteams_var, gomp_teams_thread_limit_var): Define.
(omp_display_env): Print OMP_NUM_TEAMS and OMP_TEAMS_THREAD_LIMIT.
(initialize_env): Handle OMP_NUM_TEAMS and OMP_TEAMS_THREAD_LIMIT env
vars.
* teams.c (GOMP_teams_reg): If thread_limit is not specified, use
gomp_teams_thread_limit_var as fallback if not zero. If num_teams
is not specified, use gomp_nteams_var.
* fortran.c (omp_set_num_teams, omp_get_max_teams,
omp_set_teams_thread_limit, omp_get_teams_thread_limit): Add
ialias_redirect.
(omp_set_num_teams_, omp_set_num_teams_8_, omp_get_max_teams_,
omp_set_teams_thread_limit_, omp_set_teams_thread_limit_8_,
omp_get_teams_thread_limit_): New functions.
Related to this is the addition of structured-block-sequence in OpenMP 5.1,
which doesn't change anything for Fortran, but for C/C++ allows multiple
statements instead of just one possibly compound around the separating
directives (section and scan).
I've also made some updates to the OpenMP 5.1 support list in libgomp.texi.
2021-10-09 Jakub Jelinek <jakub@redhat.com>
gcc/c/
* c-parser.c (c_parser_omp_structured_block_sequence): New function.
(c_parser_omp_scan_loop_body): Use it.
(c_parser_omp_sections_scope): Likewise.
gcc/cp/
* parser.c (cp_parser_omp_structured_block): Remove disallow_omp_attrs
argument.
(cp_parser_omp_structured_block_sequence): New function.
(cp_parser_omp_scan_loop_body): Use it.
(cp_parser_omp_sections_scope): Likewise.
gcc/testsuite/
* c-c++-common/gomp/sections1.c (foo): Don't expect errors on
multiple statements in between section directive(s). Add testcases
for invalid no statements in between section directive(s).
* gcc.dg/gomp/sections-2.c (foo): Don't expect errors on
multiple statements in between section directive(s).
* g++.dg/gomp/sections-2.C (foo): Likewise.
* g++.dg/gomp/attrs-6.C (foo): Add testcases for multiple
statements in between section directive(s).
(bar): Add testcases for multiple statements in between scan
directive.
* g++.dg/gomp/attrs-7.C (bar): Adjust expected error recovery.
libgomp/
* libgomp.texi (OpenMP 5.1): Mention implemented support for
structured block sequences in C/C++. Mention support for
unconstrained/reproducible modifiers on order clause.
Mention partial (C/C++ only) support of extentensions to atomics
construct. Mention partial (C/C++ on clause only) support of
align/allocator modifiers on allocate clause.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/order-reproducible-1.f90: New test
based on libgomp.c-c++-common/order-reproducible-1.c.
* testsuite/libgomp.fortran/order-reproducible-2.f90: Likewise.
* testsuite/libgomp.fortran/my-usleep.c: New test.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/alloc-10.f90: Fix alignment check.
* testsuite/libgomp.fortran/alloc-7.f90: Fix array access.
* testsuite/libgomp.fortran/alloc-8.f90: Likewise.
* testsuite/libgomp.fortran/alloc-11.f90: New test for omp_realloc,
based on libgomp.c-c++-common/alloc-9.c.
While OpenMP 5.1 implies order(concurrent) is the same thing as
order(reproducible:concurrent), this is going to change in OpenMP 5.2, where
essentially order(concurrent) means nothing is stated on whether it is
reproducible or unconstrained (and is determined by other means, e.g. for/do
with schedule static or runtime with static being selected is implicitly
reproducible, distribute with dist_schedule static is implicitly reproducible,
loop is implicitly reproducible) and when the modifier is specified explicitly,
it overrides the implicit behavior either way.
And, when order(reproducible:concurrent) is used with e.g. schedule(dynamic)
or some other schedule that is by definition not reproducible, it is
implementation's duty to ensure it is reproducible, either by remembering how
it scheduled some loop and then replaying the same schedule when seeing loops
with the same directive/schedule/number of iterations, or by overriding the
schedule to some reproducible one.
This patch doesn't implement the 5.2 wording just yet, but in the FEs
differentiates between the 3 states - no explicit modifier, explicit reproducible
or explicit unconstrainted, so that the middle-end can easily switch any time.
Instead it follows the 5.1 wording where both order(concurrent) (implicit or
explicit) or order(reproducible:concurrent) imply reproducibility.
And, it implements the easier method, when for/do should be reproducible, it
just chooses static schedule. order(concurrent) implies no OpenMP APIs in the
loop body nor threadprivate vars, so the exact scheduling isn't (easily at least)
observable.
2021-10-01 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.h (OMP_CLAUSE_ORDER_REPRODUCIBLE): Define.
* tree-pretty-print.c (dump_omp_clause) <case OMP_CLAUSE_ORDER>: Print
reproducible: for OMP_CLAUSE_ORDER_REPRODUCIBLE.
* omp-general.c (omp_extract_for_data): If OMP_CLAUSE_ORDER is seen
without OMP_CLAUSE_ORDER_UNCONSTRAINED, overwrite sched_kind to
OMP_CLAUSE_SCHEDULE_STATIC.
gcc/c-family/
* c-omp.c (c_omp_split_clauses): Also copy
OMP_CLAUSE_ORDER_REPRODUCIBLE.
gcc/c/
* c-parser.c (c_parser_omp_clause_order): Set
OMP_CLAUSE_ORDER_REPRODUCIBLE for explicit reproducible: modifier.
gcc/cp/
* parser.c (cp_parser_omp_clause_order): Set
OMP_CLAUSE_ORDER_REPRODUCIBLE for explicit reproducible: modifier.
gcc/fortran/
* gfortran.h (gfc_omp_clauses): Add order_reproducible bitfield.
* dump-parse-tree.c (show_omp_clauses): Print REPRODUCIBLE: for it.
* openmp.c (gfc_match_omp_clauses): Set order_reproducible for
explicit reproducible: modifier.
* trans-openmp.c (gfc_trans_omp_clauses): Set
OMP_CLAUSE_ORDER_REPRODUCIBLE for order_reproducible.
(gfc_split_omp_clauses): Also copy order_reproducible.
gcc/testsuite/
* gfortran.dg/gomp/order-5.f90: Adjust scan-tree-dump-times regexps.
libgomp/
* testsuite/libgomp.c-c++-common/order-reproducible-1.c: New test.
* testsuite/libgomp.c-c++-common/order-reproducible-2.c: New test.
This patch adds alloc_align attribute to omp_aligned_{,c}alloc so that if
the first argument is constant, GCC can assume requested alignment.
Additionally, it adds testsuite coverage for omp_realloc which I haven't
managed to write in the patch from yesterday.
2021-10-01 Jakub Jelinek <jakub@redhat.com>
* omp.h.in (omp_aligned_alloc, omp_aligned_calloc): Add
__alloc_align__ (1) attribute.
* testsuite/libgomp.c-c++-common/alloc-9.c: New test.
libgomp/
* testsuite/libgomp.fortran/alloc-7.f90: Add dg-prune-output
for -fintrinsic-modules-path= warning of the C compiler.
* testsuite/libgomp.fortran/alloc-9.f90: Likewise.
* testsuite/libgomp.fortran/alloc-10.f90: Likewise.
gcc/ChangeLog:
* omp-low.c (omp_runtime_api_call): Add omp_aligned_{,c}alloc and
omp_{c,re}alloc, fix omp_alloc/omp_free.
libgomp/ChangeLog:
* libgomp.texi (OpenMP 5.1): Set implementation status to Y for
omp_aligned_{,c}alloc and omp_{c,re}alloc routines.
* omp_lib.f90.in (omp_aligned_alloc, omp_aligned_calloc, omp_calloc,
omp_realloc): Add.
* omp_lib.h.in (omp_aligned_alloc, omp_aligned_calloc, omp_calloc,
omp_realloc): Add.
* testsuite/libgomp.fortran/alloc-10.f90: New test.
* testsuite/libgomp.fortran/alloc-6.f90: New test.
* testsuite/libgomp.fortran/alloc-7.c: New test.
* testsuite/libgomp.fortran/alloc-7.f90: New test.
* testsuite/libgomp.fortran/alloc-8.f90: New test.
* testsuite/libgomp.fortran/alloc-9.f90: New test.
This patch adds new OpenMP 5.1 allocator entrypoints and in addition to that
fixes an omp_alloc bug which is hard to test for - if the first allocator
fails but has a larger alignment trait and has a fallback allocator, either
the default behavior or a user fallback, then the extra alignment will be used
even in the fallback allocation, rather than just starting with whatever
alignment has been requested (in GOMP_alloc or the minimum one in omp_alloc).
Jonathan's comment on IRC this morning made me realize that I should add
alloc_align attributes to 2 of the prototypes and I still need to add testsuite
coverage for omp_realloc, will do that in a follow-up.
2021-09-30 Jakub Jelinek <jakub@redhat.com>
* omp.h.in (omp_aligned_alloc, omp_calloc, omp_aligned_calloc,
omp_realloc): New prototypes.
(omp_alloc): Move after omp_free prototype, add __malloc__ (omp_free)
attribute.
* allocator.c: Include string.h.
(omp_aligned_alloc): No longer static, add ialias. Add new_alignment
variable and use it instead of alignment so that when retrying the old
alignment is used again. Don't retry if new alignment is the same
as old alignment, unless allocator had pool size.
(omp_alloc, GOMP_alloc, GOMP_free): Use ialias_call.
(omp_aligned_calloc, omp_calloc, omp_realloc): New functions.
* libgomp.map (OMP_5.0.2): Export omp_aligned_alloc, omp_calloc,
omp_aligned_calloc and omp_realloc.
* testsuite/libgomp.c-c++-common/alloc-4.c (main): Add
omp_aligned_alloc, omp_calloc and omp_aligned_calloc tests.
* testsuite/libgomp.c-c++-common/alloc-5.c: New test.
* testsuite/libgomp.c-c++-common/alloc-6.c: New test.
* testsuite/libgomp.c-c++-common/alloc-7.c: New test.
* testsuite/libgomp.c-c++-common/alloc-8.c: New test.
The depend type is a struct with two pointer members for C/C++ - but for
Fortran OpenMP requires an integer type with kind = omp_depend_kind. Thus,
libgomp's configure checks that an integer type/kind with size 2*sizeof(void*)
is available. However, this integer type/kind is not needed when building without
Fortran support. Thus, only check this when Fortran is enabled.
libgomp/
PR libgomp/96661
* configure.ac: Only check for int-type = 2*size_t support when
building with Fortran support.
* configure: Regenerate.
Fix-up for recent commit 00f6de9c69
"Fortran: Fix assumed-size to assumed-rank passing [PR94070]",
and commit da1f6391b7
"libgomp.oacc-fortran/privatized-ref-2.f90: Fix dg-note".
Due to use of '#if !ACC_MEM_SHARED' conditionals in
'libgomp.oacc-fortran/if-1.f90', 'target { ! openacc_host_selected }'
needs some special care (ignoring the pre-existing mismatch of
'ACC_MEM_SHARED' vs. 'openacc_host_selected').
As seen with GCN offloading, we need to revert to another bit of the
original code in 'libgomp.oacc-fortran/privatized-ref-2.f90'.
libgomp/
* testsuite/libgomp.oacc-fortran/if-1.f90: Adjust.
* testsuite/libgomp.oacc-fortran/privatized-ref-2.f90: Likewise.
This patch implements the new hybrid forward threader and replaces the
embedded VRP threader with it.
With all the pieces that have gone in, the implementation of the hybrid
threader is straightforward: convert the current state into
SSA imports that the solver will understand, and let the path solver
precompute ranges and relations for the path. After this setup is done,
we can use the range_query API to solve gimple statements in the threader.
The forward threader is now engine agnostic so there are no changes to
the threader per se.
I have put the hybrid bits in tree-ssa-threadedge.*, instead of VRP,
because they will also be used in the evrp removal of the DOM/threader,
which is my next task.
Most of the patch, is actually test changes. I have gone through every
single one and verified that we're correct. Most were trivial dump
file name changes, but others required going through the IL an
certifying that the different IL was expected.
For example, in pr59597.c, we have one less thread because the
ASSERT_EXPR was getting in the way, and making it seem like things were
not crossing loops. The hybrid threader sees the correct representation
of the IL, and avoids threading this one case.
The final numbers are a 12.16% improvement in jump threads immediately
after VRP, and a 0.82% improvement in overall jump threads. The
performance drop is 0.6% (plus the 1.43% hit from moving the embedded
threader into its own pass). As I've said, I'd prefer to keep the
threader in its own pass, but if this is an issue, we can address this
with a shared ranger when VRP is replaced with an evrp instance
(upcoming).
Note, that these numbers are slightly different than what I originally
posted. A few correctness tweaks, plus restricting loop threads, made
the difference. That being said, I was aiming for par. A 12% gain is
just gravy ;-). When we merge the threaders, we should see even better
numbers-- and we'll have the benefit of an entire release stress testing
the solver.
As I mentioned in my introductory note, paths ending in MEM_REF
conditional are missing. In reality, this didn't make a difference, as
it was so rare. However, as a follow-up, I will distill a test and add
a suitable PR to keep us honest.
There is a one-line change to libgomp/team.c silencing a new used
uninitialized warning. As my previous work with the threaders has
shown, warnings flare up after each improvement to jump threading. I
expect this to be no different. I've promised Jakub to investigate
fully, so I will analyze and add the appropriate PR for the warning
experts.
Oh yeah, the new pass dump is called vrp-threader[12] to match each
VRP[12] pass. However, there's no reason for it to either be named
vrp-threader, or for it to live in tree-vrp.c.
Tested on x86-64 Linux.
OK?
p.s. "Did I say 5 weeks? My bad, I meant 5 months."
gcc/ChangeLog:
* passes.def (pass_vrp_threader): New.
* tree-pass.h (make_pass_vrp_threader): Add make_pass_vrp_threader.
* tree-ssa-threadedge.c (hybrid_jt_state::register_equivs_stmt): New.
(hybrid_jt_simplifier::hybrid_jt_simplifier): New.
(hybrid_jt_simplifier::simplify): New.
(hybrid_jt_simplifier::compute_ranges_from_state): New.
* tree-ssa-threadedge.h (class hybrid_jt_state): New.
(class hybrid_jt_simplifier): New.
* tree-vrp.c (execute_vrp): Remove ASSERT_EXPR based jump
threader.
(class hybrid_threader): New.
(hybrid_threader::hybrid_threader): New.
(hybrid_threader::~hybrid_threader): New.
(hybrid_threader::before_dom_children): New.
(hybrid_threader::after_dom_children): New.
(execute_vrp_threader): New.
(class pass_vrp_threader): New.
(make_pass_vrp_threader): New.
libgomp/ChangeLog:
* team.c: Initialize start_data.
* testsuite/libgomp.graphite/force-parallel-4.c: Adjust.
* testsuite/libgomp.graphite/force-parallel-8.c: Adjust.
gcc/testsuite/ChangeLog:
* gcc.dg/torture/pr55107.c: Adjust.
* gcc.dg/tree-ssa/phi_on_compare-1.c: Adjust.
* gcc.dg/tree-ssa/phi_on_compare-2.c: Adjust.
* gcc.dg/tree-ssa/phi_on_compare-3.c: Adjust.
* gcc.dg/tree-ssa/phi_on_compare-4.c: Adjust.
* gcc.dg/tree-ssa/pr21559.c: Adjust.
* gcc.dg/tree-ssa/pr59597.c: Adjust.
* gcc.dg/tree-ssa/pr61839_1.c: Adjust.
* gcc.dg/tree-ssa/pr61839_3.c: Adjust.
* gcc.dg/tree-ssa/pr71437.c: Adjust.
* gcc.dg/tree-ssa/ssa-dom-thread-11.c: Adjust.
* gcc.dg/tree-ssa/ssa-dom-thread-16.c: Adjust.
* gcc.dg/tree-ssa/ssa-dom-thread-18.c: Adjust.
* gcc.dg/tree-ssa/ssa-dom-thread-2a.c: Adjust.
* gcc.dg/tree-ssa/ssa-dom-thread-4.c: Adjust.
* gcc.dg/tree-ssa/ssa-thread-14.c: Adjust.
* gcc.dg/tree-ssa/ssa-vrp-thread-1.c: Adjust.
* gcc.dg/tree-ssa/vrp106.c: Adjust.
* gcc.dg/tree-ssa/vrp55.c: Adjust.
In my last commit, r12-3897-g00f6de9c69119594f7dad3bd525937c94c8200d0,
which inlined array-size code, I had to update the expected output. However,
in doing so, I accidentally (copy'n'paste) changed dg-note into dg-message.
libgomp/
* testsuite/libgomp.oacc-fortran/privatized-ref-2.f90: Change
dg-message back to dg-note.
This code inlines the size0 and size1 libgfortran calls, the former is still
used by libgfortan itself (and by old code). Besides permitting more
optimizations, it also permits to handle assumed-rank dummies better: If the
dummy argument is a nonpointer/nonallocatable, an assumed-size actual arg is
repesented by having ubound == -1 for the last dimension. However, for
allocatable/pointers, this value can also exist. Hence, the dummy arg attr
has to be honored.
For that reason, when calling an assumed-rank procedure with nonpointer,
nonallocatable dummy arguments, the bounds have to be updated to avoid
the case ubound == -1 for the last dimension.
PR fortran/94070
gcc/fortran/ChangeLog:
* trans-array.c (gfc_tree_array_size): New function to
find size inline (whole array or one dimension).
(array_parameter_size): Use it, take stmt_block as arg.
(gfc_conv_array_parameter): Update call.
* trans-array.h (gfc_tree_array_size): Add prototype.
* trans-decl.c (gfor_fndecl_size0, gfor_fndecl_size1): Remove
these global vars.
(gfc_build_intrinsic_function_decls): Remove their initialization.
* trans-expr.c (gfc_conv_procedure_call): Update
bounds of pointer/allocatable actual args to nonallocatable/nonpointer
dummies to be one based.
* trans-intrinsic.c (gfc_conv_intrinsic_shape): Fix case for
assumed rank with allocatable/pointer dummy.
(gfc_conv_intrinsic_size): Update to use inline function.
* trans.h (gfor_fndecl_size0, gfor_fndecl_size1): Remove var decl.
libgfortran/ChangeLog:
* intrinsics/size.c (size0, size1): Comment that now not
used by newer compiler code.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-fortran/privatized-ref-2.f90: Update
expected dg-note output.
gcc/testsuite/ChangeLog:
* gfortran.dg/c-interop/cf-out-descriptor-6.f90: Remove xfail.
* gfortran.dg/c-interop/size.f90: Remove xfail.
* gfortran.dg/intrinsic_size_3.f90: Update scan-tree-dump-times.
* gfortran.dg/transpose_optimization_2.f90: Likewise.
* gfortran.dg/size_optional_dim_1.f90: Add scan-tree-dump-not.
* gfortran.dg/assumed_rank_22.f90: New test.
* gfortran.dg/assumed_rank_22_aux.c: New test.
It turned out that enabling the -Wmissing-include-dirs for libcpp did output
too many warnings – at least as run with -B and similar options during the
GCC build and warning for internal include dirs like finclude, unlikely of
relevance to for a real-world user.
This patch now only warns for -I and -J by default but permits to get the
full warnings including libcpp ones with -Wmissing-include-dirs. It
additionally documents this in the manual.
With that change, the -Wno-missing-include-dirs could be removed
from libgfortran's configure and libgomp's testsuite always cflags.
This reverts those bits of the previous
commit r12-3722-g417ea5c02cef7f000e66d1af22b066c2c1cda047
Additionally, it turned out that all call to load_file called exit
explicitly - except for the main file via gfc_init -> gfc_new_file. The
latter also output a file not existing fatal error, such that two errors
where printed. Now exit is called in line with the other users of
load_file.
Finally, when compileing with "nonexisting/file.f90", first a warning that
"nonexisting" does not exist as include path was printed before the file
not found error was printed. Now the directory in which the physical file
is located is added silently, relying on the file-not-found diagnostic for
those.
PR fortran/55534
gcc/ChangeLog:
* doc/invoke.texi (-Wno-missing-include-dirs.): Document Fortran
behavior.
gcc/fortran/ChangeLog:
* cpp.c (gfc_cpp_register_include_paths, gfc_cpp_post_options):
Add new bool verbose_missing_dir_warn argument.
* cpp.h (gfc_cpp_post_options): Update prototype.
* f95-lang.c (gfc_init): Remove duplicated file-not found diag.
* gfortran.h (gfc_check_include_dirs): Takes bool
verbose_missing_dir_warn arg.
(gfc_new_file): Returns now void.
* options.c (gfc_post_options): Update to warn for -I and -J,
only, by default but for all when user requested.
* scanner.c (gfc_do_check_include_dir):
(gfc_do_check_include_dirs, gfc_check_include_dirs): Take bool
verbose warn arg and update to avoid printing the same message
twice or never.
(load_file): Fix indent.
(gfc_new_file): Return void and exit when load_file failed
as all other load_file users do.
libgfortran/ChangeLog:
* configure.ac (AM_FCFLAGS): Revert r12-3722 by removing
-Wno-missing-include-dirs.
* configure: Regenerate.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/fortran.exp (ALWAYS_CFLAGS): Revert
r12-3722 by removing -Wno-missing-include-dirs.
* testsuite/libgomp.oacc-fortran/fortran.exp (ALWAYS_CFLAGS): Likewise.
gcc/testsuite/ChangeLog:
* gfortran.dg/include_14.f90: Add -J testcase and update dg-output.
* gfortran.dg/include_15.f90: Likewise.
* gfortran.dg/include_16.f90: Likewise.
* gfortran.dg/include_17.f90: Likewise.
* gfortran.dg/include_18.f90: Likewise.
* gfortran.dg/include_19.f90: Likewise.
As the allocate-2.c testcase shows, this change isn't 100% backwards compatible,
one could have allocate and/or align functions that return an OpenMP allocator
handle and previously it would call those functions and now would use those
names as keywords for the modifiers. But it allows specify extra alignment
requirements for the allocations.
2021-09-22 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.h (OMP_CLAUSE_ALLOCATE_ALIGN): Define.
* tree.c (omp_clause_num_ops): Change number of OMP_CLAUSE_ALLOCATE
arguments from 2 to 3.
* tree-pretty-print.c (dump_omp_clause): Print allocator() around
allocate clause allocator and print align if present.
* omp-low.c (scan_sharing_clauses): Force allocate_map entry even
for omp_default_mem_alloc if align modifier is present. If align
modifier is present, use TREE_LIST to encode both allocator and
align.
(lower_private_allocate, lower_rec_input_clauses, create_task_copyfn):
Handle align modifier on allocator clause if present.
gcc/c-family/
* c-omp.c (c_omp_split_clauses): Copy over OMP_CLAUSE_ALLOCATE_ALIGN.
gcc/c/
* c-parser.c (c_parser_omp_clause_allocate): Parse allocate clause
modifiers.
gcc/cp/
* parser.c (cp_parser_omp_clause_allocate): Parse allocate clause
modifiers.
* semantics.c (finish_omp_clauses) <OMP_CLAUSE_ALLOCATE>: Perform
semantic analysis of OMP_CLAUSE_ALLOCATE_ALIGN.
* pt.c (tsubst_omp_clauses) <case OMP_CLAUSE_ALLOCATE>: Handle
also OMP_CLAUSE_ALLOCATE_ALIGN.
gcc/testsuite/
* c-c++-common/gomp/allocate-6.c: New test.
* c-c++-common/gomp/allocate-7.c: New test.
* g++.dg/gomp/allocate-4.C: New test.
libgomp/
* testsuite/libgomp.c-c++-common/allocate-2.c: New test.
* testsuite/libgomp.c-c++-common/allocate-3.c: New test.
gcc/fortran/ChangeLog:
PR fortran/55534
* cpp.c: Define GCC_C_COMMON_C for #include "options.h" to make
cpp_reason_option_codes available.
(gfc_cpp_register_include_paths): Make static, set pfile's
warn_missing_include_dirs and move before caller.
(gfc_cpp_init_cb): New, cb code moved from ...
(gfc_cpp_init_0): ... here.
(gfc_cpp_post_options): Call gfc_cpp_init_cb.
(cb_cpp_diagnostic_cpp_option): New. As implemented in c-family
to match CppReason flags to -W... names.
(cb_cpp_diagnostic): Use it to replace single special case.
* cpp.h (gfc_cpp_register_include_paths): Remove as now static.
* gfortran.h (gfc_check_include_dirs): New prototype.
(gfc_add_include_path): Add new bool arg.
* options.c (gfc_init_options): Don't set -Wmissing-include-dirs.
(gfc_post_options): Set it here after commandline processing. Call
gfc_add_include_path with defer_warn=false.
(gfc_handle_option): Call it with defer_warn=true.
* scanner.c (gfc_do_check_include_dir, gfc_do_check_include_dirs,
gfc_check_include_dirs): New. Diagnostic moved from ...
(add_path_to_list): ... here, which came before cmdline processing.
Take additional bool defer_warn argument.
(gfc_add_include_path): Take additional defer_warn arg.
* scanner.h (struct gfc_directorylist): Reorder for alignment issues,
add new 'bool warn'.
libgfortran/ChangeLog:
PR fortran/55534
* configure.ac (AM_FCFLAGS): Add -Wno-missing-include-dirs.
* configure: Regenerate.
libgomp/ChangeLog:
PR fortran/55534
* testsuite/libgomp.fortran/fortran.exp: Add -Wno-missing-include-dirs
to ALWAYS_CFLAGS.
* testsuite/libgomp.oacc-fortran/fortran.exp: Likewise.
gcc/testsuite/ChangeLog:
* gfortran.dg/include_6.f90: Change dg-error to
dg-warning and update pattern.
* gfortran.dg/include_14.f90: New test.
* gfortran.dg/include_15.f90: New test.
* gfortran.dg/include_16.f90: New test.
* gfortran.dg/include_17.f90: New test.
* gfortran.dg/include_18.f90: New test.
* gfortran.dg/include_19.f90: New test.
* gfortran.dg/include_20.f90: New test.
* gfortran.dg/include_21.f90: New test.
OpenMP 5.1 allows default(private) or default(firstprivate) even in C/C++,
but it behaves the same way as in Fortran only for variables not declared at
namespace or file scope. For the namespace/file scope variables it instead
behaves as default(none).
2021-09-18 Jakub Jelinek <jakub@redhat.com>
gcc/
* gimplify.c (omp_default_clause): For C/C++ default({,first}private),
if file/namespace scope variable doesn't have predetermined sharing,
treat it as if there was default(none).
gcc/c/
* c-parser.c (c_parser_omp_clause_default): Handle private and
firstprivate arguments, adjust diagnostics on unknown argument.
gcc/cp/
* parser.c (cp_parser_omp_clause_default): Handle private and
firstprivate arguments, adjust diagnostics on unknown argument.
* cp-gimplify.c (cxx_omp_finish_clause): Handle OMP_CLAUSE_PRIVATE.
gcc/testsuite/
* c-c++-common/gomp/default-2.c: New test.
* c-c++-common/gomp/default-3.c: New test.
* g++.dg/gomp/default-1.C: New test.
libgomp/
* testsuite/libgomp.c++/default-1.C: New test.
* testsuite/libgomp.c-c++-common/default-1.c: New test.
* libgomp.texi (OpenMP 5.1): Mark "private and firstprivate argument
to default clause in C and C++" as implemented.
This patch implements an algorithm to lay out local data-share (LDS)
space. It currently works for AMD GCN. At the moment, LDS is used for
three things:
1. Gang-private variables
2. Reduction temporaries (accumulators)
3. Broadcasting for worker partitioning
After the patch is applied, (2) and (3) are placed at preallocated
locations in LDS, and (1) continues to be handled by the backend (as it
is at present prior to this patch being applied). LDS now looks like this:
+--------------+ (gang-private size + 1024, = 1536)
| free space |
| ... |
| - - - - - - -|
| worker bcast |
+--------------+
| reductions |
+--------------+ <<< -mgang-private-size=<number> (def. 512)
| gang-private |
| vars |
+--------------+ (32)
| low LDS vars |
+--------------+ LDS base
So, gang-private space is fixed at a constant amount at compile time
(which can be increased with a command-line switch if necessary
for some given code). The layout algorithm takes out a slice of the
remainder of usable space for reduction vars, and uses the rest for
worker partitioning.
The partitioning algorithm works as follows.
1. An "adjacency" set is built up for each basic block that might
do a broadcast. This is calculated by starting at each such block,
and doing a recursive DFS walk over successors to find the next
block (or blocks) that *also* does a broadcast
(dfs_broadcast_reachable_1).
2. The adjacency set is inverted to get adjacent predecessor blocks also.
3. Blocks that will perform a broadcast are sorted by size of that
broadcast: the biggest blocks are handled first.
4. A splay tree structure is used to calculate the spans of LDS memory
that are already allocated by the blocks adjacent to this one
(merge_ranges{,_1}.
5. The current block's broadcast space is allocated from the first free
span not allocated in the splay tree structure calculated above
(first_fit_range). This seems to work quite nicely and efficiently
with the splay tree structure.
6. Continue with the next-biggest broadcast block until we're done.
In this way, "adjacent" broadcasts will not use the same piece of
LDS memory.
PR96334 "openacc: Unshare reduction temporaries for GCN" got merged in:
The GCN backend uses tree nodes like MEM((__lds TYPE *) <constant>)
for reduction temporaries. Unlike e.g. var decls and SSA names, these
nodes cannot be shared during gimplification, but are so in some
circumstances. This is detected when appropriate --enable-checking
options are used. This patch unshares such nodes when they are reused
more than once.
gcc/
* config/gcn/gcn-protos.h
(gcn_goacc_create_worker_broadcast_record): Update prototype.
* config/gcn/gcn-tree.c (gcn_goacc_get_worker_red_decl): Use
preallocated block of LDS memory. Do not cache/share decls for
reduction temporaries between invocations.
(gcn_goacc_reduction_teardown): Unshare VAR on second use.
(gcn_goacc_create_worker_broadcast_record): Add OFFSET parameter
and return temporary LDS space at that offset. Return pointer in
"sender" case.
* config/gcn/gcn.c (acc_lds_size, gang_private_hwm, lds_allocs):
New global vars.
(ACC_LDS_SIZE): Define as acc_lds_size.
(gcn_init_machine_status): Don't initialise lds_allocated,
lds_allocs, reduc_decls fields of machine function struct.
(gcn_option_override): Handle default size for gang-private
variables and -mgang-private-size option.
(gcn_expand_prologue): Use LDS_SIZE instead of LDS_SIZE-1 when
initialising M0_REG.
(gcn_shared_mem_layout): New function.
(gcn_print_lds_decl): Update comment. Use global lds_allocs map and
gang_private_hwm variable.
(TARGET_GOACC_SHARED_MEM_LAYOUT): Define target hook.
* config/gcn/gcn.h (machine_function): Remove lds_allocated,
lds_allocs, reduc_decls. Add reduction_base, reduction_limit.
* config/gcn/gcn.opt (gang_private_size_opt): New global.
(mgang-private-size=): New option.
* doc/tm.texi.in (TARGET_GOACC_SHARED_MEM_LAYOUT): Place
documentation hook.
* doc/tm.texi: Regenerate.
* omp-oacc-neuter-broadcast.cc (targhooks.h, diagnostic-core.h):
Add includes.
(build_sender_ref): Handle sender_decl being pointer.
(worker_single_copy): Add PLACEMENT and ISOLATE_BROADCASTS
parameters. Pass placement argument to
create_worker_broadcast_record hook invocations. Handle
sender_decl being pointer and isolate_broadcasts inserting extra
barriers.
(blk_offset_map_t): Add typedef.
(neuter_worker_single): Add BLK_OFFSET_MAP parameter. Pass
preallocated range to worker_single_copy call.
(dfs_broadcast_reachable_1): New function.
(idx_decl_pair_t, used_range_vec_t): New typedefs.
(sort_size_descending): New function.
(addr_range): New class.
(splay_tree_compare_addr_range, splay_tree_free_key)
(first_fit_range, merge_ranges_1, merge_ranges): New functions.
(execute_omp_oacc_neuter_broadcast): Rename to...
(oacc_do_neutering): ... this. Add BOUNDS_LO, BOUNDS_HI
parameters. Arrange layout of shared memory for broadcast
operations.
(execute_omp_oacc_neuter_broadcast): New function.
(pass_omp_oacc_neuter_broadcast::gate): Remove num_workers==1
handling from here. Enable pass for all OpenACC routines in order
to call shared memory-layout hook.
* target.def (create_worker_broadcast_record): Add OFFSET
parameter.
(shared_mem_layout): New hook.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/broadcast-many.c: Update.
Besides the C++ FE changes, I've noticed that the C FE didn't reject
#pragma omp atomic capture compare
{ v = x; x = y; }
and other forms of atomic swap, this patch fixes that too. And the
c-family/ routine needed quite a few changes so that the new code
in it works fine with both FEs.
2021-09-17 Jakub Jelinek <jakub@redhat.com>
gcc/c-family/
* c-omp.c (c_finish_omp_atomic): Avoid creating
TARGET_EXPR if test is true, use create_tmp_var_raw instead of
create_tmp_var and add a zero initializer to TARGET_EXPRs that
had NULL initializer. When omitting operands after v = x,
use type of v rather than type of x. Fix type of vtmp
TARGET_EXPR.
gcc/c/
* c-parser.c (c_parser_omp_atomic): Reject atomic swap if capture
is true.
gcc/cp/
* cp-tree.h (finish_omp_atomic): Add r and weak arguments.
* parser.c (cp_parser_omp_atomic): Update function comment for
OpenMP 5.1 atomics, parse OpenMP 5.1 atomics and fail, compare and
weak clauses.
* semantics.c (finish_omp_atomic): Add r and weak arguments, handle
them, handle COND_EXPRs.
* pt.c (tsubst_expr): Adjust for COND_EXPR forms that
finish_omp_atomic can now produce.
gcc/testsuite/
* c-c++-common/gomp/atomic-18.c: Expect same diagnostics in C++ as in
C.
* c-c++-common/gomp/atomic-25.c: Drop c effective target.
* c-c++-common/gomp/atomic-26.c: Likewise.
* c-c++-common/gomp/atomic-27.c: Likewise.
* c-c++-common/gomp/atomic-28.c: Likewise.
* c-c++-common/gomp/atomic-29.c: Likewise.
* c-c++-common/gomp/atomic-30.c: Likewise. Adjust expected diagnostics
for C++ when it differs from C.
(foo): Change return type from double to void.
* g++.dg/gomp/atomic-5.C: Adjust expected diagnostics wording.
* g++.dg/gomp/atomic-20.C: New test.
libgomp/
* testsuite/libgomp.c-c++-common/atomic-19.c: Drop c effective target.
Use /* */ comments instead of //.
* testsuite/libgomp.c-c++-common/atomic-20.c: Likewise.
* testsuite/libgomp.c-c++-common/atomic-21.c: Likewise.
* testsuite/libgomp.c++/atomic-16.C: New test.
* testsuite/libgomp.c++/atomic-17.C: New test.
This patch implements OpenMP 5.1 atomics (with clarifications from upcoming 5.2).
The most important changes are that it is now possible to write (for C/C++,
for Fortran it was possible before already) min/max atomics and more importantly
compare and exchange in various forms.
Also, acq_rel is now allowed on read/write and acq_rel/acquire are allowed on
update, and there are new compare, weak and fail clauses.
2021-09-10 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree-core.h (enum omp_memory_order): Add OMP_MEMORY_ORDER_MASK,
OMP_FAIL_MEMORY_ORDER_UNSPECIFIED, OMP_FAIL_MEMORY_ORDER_RELAXED,
OMP_FAIL_MEMORY_ORDER_ACQUIRE, OMP_FAIL_MEMORY_ORDER_RELEASE,
OMP_FAIL_MEMORY_ORDER_ACQ_REL, OMP_FAIL_MEMORY_ORDER_SEQ_CST and
OMP_FAIL_MEMORY_ORDER_MASK enumerators.
(OMP_FAIL_MEMORY_ORDER_SHIFT): Define.
* gimple-pretty-print.c (dump_gimple_omp_atomic_load,
dump_gimple_omp_atomic_store): Print [weak] for weak atomic
load/store.
* gimple.h (enum gf_mask): Change GF_OMP_ATOMIC_MEMORY_ORDER
to 6-bit mask, adjust GF_OMP_ATOMIC_NEED_VALUE value and add
GF_OMP_ATOMIC_WEAK.
(gimple_omp_atomic_weak_p, gimple_omp_atomic_set_weak): New inline
functions.
* tree.h (OMP_ATOMIC_WEAK): Define.
* tree-pretty-print.c (dump_omp_atomic_memory_order): Adjust for
fail memory order being encoded in the same enum and also print
fail clause if present.
(dump_generic_node): Print weak clause if OMP_ATOMIC_WEAK.
* gimplify.c (goa_stabilize_expr): Add target_expr and rhs arguments,
handle pre_p == NULL case as a test mode that only returns value
but doesn't change gimplify nor change anything otherwise, adjust
recursive calls, add MODIFY_EXPR, ADDR_EXPR, COND_EXPR, TARGET_EXPR
and CALL_EXPR handling, adjust COMPOUND_EXPR handling for
__builtin_clear_padding calls, for !rhs gimplify as lvalue rather
than rvalue.
(gimplify_omp_atomic): Adjust goa_stabilize_expr caller. Handle
COND_EXPR rhs. Set weak flag on gimple load/store for
OMP_ATOMIC_WEAK.
* omp-expand.c (omp_memory_order_to_fail_memmodel): New function.
(omp_memory_order_to_memmodel): Adjust for fail clause encoded
in the same enum.
(expand_omp_atomic_cas): New function.
(expand_omp_atomic_pipeline): Use omp_memory_order_to_fail_memmodel
function.
(expand_omp_atomic): Attempt to optimize atomic compare and exchange
using expand_omp_atomic_cas.
gcc/c-family/
* c-common.h (c_finish_omp_atomic): Add r and weak arguments.
* c-omp.c: Include gimple-fold.h.
(c_finish_omp_atomic): Add r and weak arguments. Add support for
OpenMP 5.1 atomics.
gcc/c/
* c-parser.c (c_parser_conditional_expression): If omp_atomic_lhs and
cond.value is >, < or == with omp_atomic_lhs as one of the operands,
don't call build_conditional_expr, instead build a COND_EXPR directly.
(c_parser_binary_expression): Avoid calling parser_build_binary_op
if omp_atomic_lhs even in more cases for >, < or ==.
(c_parser_omp_atomic): Update function comment for OpenMP 5.1 atomics,
parse OpenMP 5.1 atomics and fail, compare and weak clauses, allow
acq_rel on atomic read/write and acq_rel/acquire clauses on update.
* c-typeck.c (build_binary_op): For flag_openmp only handle
MIN_EXPR/MAX_EXPR.
gcc/cp/
* parser.c (cp_parser_omp_atomic): Allow acq_rel on atomic read/write
and acq_rel/acquire clauses on update.
* semantics.c (finish_omp_atomic): Adjust c_finish_omp_atomic caller.
gcc/testsuite/
* c-c++-common/gomp/atomic-17.c (foo): Add tests for atomic read,
write or update with acq_rel clause and atomic update with acquire clause.
* c-c++-common/gomp/atomic-18.c (foo): Adjust expected diagnostics
wording, remove tests moved to atomic-17.c.
* c-c++-common/gomp/atomic-21.c: Expect only 2 omp atomic release and
2 omp atomic acq_rel directives instead of 4 omp atomic release.
* c-c++-common/gomp/atomic-25.c: New test.
* c-c++-common/gomp/atomic-26.c: New test.
* c-c++-common/gomp/atomic-27.c: New test.
* c-c++-common/gomp/atomic-28.c: New test.
* c-c++-common/gomp/atomic-29.c: New test.
* c-c++-common/gomp/atomic-30.c: New test.
* c-c++-common/goacc-gomp/atomic.c: Expect 1 omp atomic release and
1 omp atomic_acq_rel instead of 2 omp atomic release directives.
* gcc.dg/gomp/atomic-5.c: Adjust expected error diagnostic wording.
* g++.dg/gomp/atomic-18.C:Expect 4 omp atomic release and
1 omp atomic_acq_rel instead of 5 omp atomic release directives.
libgomp/
* testsuite/libgomp.c-c++-common/atomic-19.c: New test.
* testsuite/libgomp.c-c++-common/atomic-20.c: New test.
* testsuite/libgomp.c-c++-common/atomic-21.c: New test.
... to avoid a regression with recent
commit 090f0d78f1
"openmp: Improve expand_omp_atomic_pipeline":
unresolved symbol __atomic_compare_exchange_1
collect2: error: ld returned 1 exit status
mkoffload: fatal error: [...]/gcc/x86_64-pc-linux-gnu-accel-nvptx-none-gcc returned 1 exit status
libgomp/
* testsuite/libgomp.c/target-43.c: '-latomic' for nvptx offloading.
Intel MIC (emulated) offloading execution failure remains to be analyzed.
libgomp/
* testsuite/libgomp.c/address-space-1.c: New file.
Co-authored-by: Jakub Jelinek <jakub@redhat.com>
..., and use that to improve XFAILing for Intel MIC offloading execution
instead of compilation in 'libgomp.c-c++-common/target-45.c',
'libgomp.fortran/target10.f90'.
gcc/
* config/i386/i386-options.c (ix86_omp_device_kind_arch_isa)
<omp_device_arch> [ACCEL_COMPILER]: Match "intel_mic".
* config/i386/t-omp-device (omp-device-properties-i386) <arch>:
Add "intel_mic".
libgomp/
* testsuite/lib/libgomp.exp
(check_effective_target_offload_target_intelmic): Remove 'proc'.
(check_effective_target_offload_device_intel_mic): New 'proc'.
* testsuite/libgomp.c-c++-common/on_device_arch.h
(device_arch_intel_mic, on_device_arch_intel_mic): New.
* testsuite/libgomp.c-c++-common/target-45.c: Use that for
'dg-xfail-run-if'.
* testsuite/libgomp.fortran/target10.f90: Likewise.
This patch adds support for the 'strict' modifier on grainsize/num_tasks
clauses, an OpenMP 5.1 feature supported in C/C++ since commit
r12-3066-g3bc75533d1f87f0617be6c1af98804f9127ec637
gcc/fortran/ChangeLog:
* dump-parse-tree.c (show_omp_clauses): Handle 'strict' modifier
on grainsize/num_tasks
* gfortran.h (gfc_omp_clauses): Add grainsize_strict
and num_tasks_strict.
* trans-openmp.c (gfc_trans_omp_clauses, gfc_split_omp_clauses):
Handle 'strict' modifier on grainsize/num_tasks.
* openmp.c (gfc_match_omp_clauses): Likewise.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/taskloop-4-a.f90: New test.
* testsuite/libgomp.fortran/taskloop-4.f90: New test.
* testsuite/libgomp.fortran/taskloop-5-a.f90: New test.
* testsuite/libgomp.fortran/taskloop-5.f90: New test.
With strict: modifier on these clauses, the standard is explicit about
how many iterations (and which) each generated task of taskloop directive
should contain. For num_tasks it actually matches what we were already
implementing, but for grainsize it does not (and even violates the old
rule - without strict it requires that the number of iterations (unspecified
which exactly) handled by each generated task is >= grainsize argument and
< 2 * grainsize argument, with strict: it requires that each generated
task handles exactly == grainsize argument iterations, except for the
generated task handling the last iteration which can handles <= grainsize
iterations).
The following patch implements it for C and C++.
2021-08-23 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.h (OMP_CLAUSE_GRAINSIZE_STRICT): Define.
(OMP_CLAUSE_NUM_TASKS_STRICT): Define.
* tree-pretty-print.c (dump_omp_clause) <case OMP_CLAUSE_GRAINSIZE,
case OMP_CLAUSE_NUM_TASKS>: Print strict: modifier.
* omp-expand.c (expand_task_call): Use GOMP_TASK_FLAG_STRICT in iflags
if either grainsize or num_tasks clause has the strict modifier.
gcc/c/
* c-parser.c (c_parser_omp_clause_num_tasks,
c_parser_omp_clause_grainsize): Parse the optional strict: modifier.
gcc/cp/
* parser.c (cp_parser_omp_clause_num_tasks,
cp_parser_omp_clause_grainsize): Parse the optional strict: modifier.
include/
* gomp-constants.h (GOMP_TASK_FLAG_STRICT): Define.
libgomp/
* taskloop.c (GOMP_taskloop): Handle GOMP_TASK_FLAG_STRICT.
* testsuite/libgomp.c-c++-common/taskloop-4.c (main): Fix up comment.
* testsuite/libgomp.c-c++-common/taskloop-5.c: New test.
... and add a minimum amount of offloading testing.
(Leaving aside that 'fwrite' to 'stderr' probably wouldn't work anyway) the
'fwrite' calls in 'libgomp/error.c:GOMP_warning', 'libgomp/error.c:GOMP_error'
drag in 'isatty', which isn't provided by my nvptx newlib build at present, so
we get, for example:
[...]
FAIL: libgomp.c/../libgomp.c-c++-common/declare_target-1.c (test for excess errors)
Excess errors:
unresolved symbol isatty
mkoffload: fatal error: [...]/build-gcc/./gcc/x86_64-pc-linux-gnu-accel-nvptx-none-gcc returned 1 exit status
[...]
..., and many more.
Fix up for recent commit 0d973c0a0d
"openmp: Implement the error directive".
libgomp/
* config/nvptx/error.c (fwrite, exit): Override, too.
* testsuite/libgomp.c-c++-common/error-1.c: Add a minimum amount
of offloading testing.
* testsuite/libgomp.fortran/error-1.f90: Likewise.
This patch implements the error directive. Depending on clauses it is either
a compile time diagnostics (in that case diagnosed right away) or runtime
diagnostics (libgomp API call that diagnoses at runtime), and either fatal
or warning (error or warning at compile time or fatal error vs. error at
runtime) and either has no message or user supplied message (this kind of
e.g. deprecated attribute). The directive is also stand-alone directive
when at runtime while utility (thus disappears from the IL as if it wasn't
there for parsing like nothing directive) at compile time.
There are some clarifications in the works ATM, so this patch doesn't yet
require that for compile time diagnostics the user message must be a constant
string literal, there are uncertainities on what exactly is valid argument
of message clause (whether just const char * type, convertible to const char *,
qualified/unqualified const char * or char * or what else) and what to do
in templates. Currently even in templates it is diagnosed right away for
compile time diagnostics, if we'll need to substitute it, we'd need to queue
something into the IL, have pt.c handle it and diagnose only later.
2021-08-20 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-builtins.def (BUILT_IN_GOMP_WARNING, BUILT_IN_GOMP_ERROR): New
builtins.
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_ERROR.
* c-pragma.c (omp_pragmas): Add error directive.
* c-omp.c (omp_directives): Uncomment error directive entry.
gcc/c/
* c-parser.c (c_parser_omp_error): New function.
(c_parser_pragma): Handle PRAGMA_OMP_ERROR.
gcc/cp/
* parser.c (cp_parser_handle_statement_omp_attributes): Determine if
PRAGMA_OMP_ERROR directive is C_OMP_DIR_STANDALONE.
(cp_parser_omp_error): New function.
(cp_parser_pragma): Handle PRAGMA_OMP_ERROR.
gcc/fortran/
* types.def (BT_FN_VOID_CONST_PTR_SIZE): New DEF_FUNCTION_TYPE_2.
* f95-lang.c (ATTR_COLD_NORETURN_NOTHROW_LEAF_LIST): Define.
gcc/testsuite/
* c-c++-common/gomp/error-1.c: New test.
* c-c++-common/gomp/error-2.c: New test.
* c-c++-common/gomp/error-3.c: New test.
* g++.dg/gomp/attrs-1.C (bar): Add error directive test.
* g++.dg/gomp/attrs-2.C (bar): Add error directive test.
* g++.dg/gomp/attrs-13.C: New test.
* g++.dg/gomp/error-1.C: New test.
libgomp/
* libgomp.map (GOMP_5.1): Add GOMP_error and GOMP_warning.
* libgomp_g.h (GOMP_warning, GOMP_error): Declare.
* error.c (GOMP_warning, GOMP_error): New functions.
* testsuite/libgomp.c-c++-common/error-1.c: New test.
This patch adds the Fortran interface for omp_alloc/omp_free
and the omp_target_* memory routines, which were added in
OpenMP 5.0 for C/C++ but only OpenMP 5.1 added them for Fortran.
Those functions use BIND(C), i.e. on the libgomp side, the same
interface as for C/C++ is used.
Note: By using BIND(C) in omp_lib.h, files including this file
no longer compiler with -std=f95 but require at least -std=f2003.
libgomp/ChangeLog:
* omp_lib.f90.in (omp_alloc, omp_free, 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): Add interface.
* omp_lib.h.in (omp_alloc, omp_free, 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): Add interface.
* testsuite/libgomp.fortran/alloc-1.F90: Remove local
interface block for omp_alloc + omp_free.
* testsuite/libgomp.fortran/alloc-4.f90: Likewise.
* testsuite/libgomp.fortran/refcount-1.f90: New test.
* testsuite/libgomp.fortran/target-12.f90: New test.
As has been clarified, it is intentional that nothing directive is accepted
in substatements of selection and looping statements and after labels and
is handled as if the directive just isn't there, so that
void
foo (int x)
{
if (x)
#pragma omp metadirective when (...:nothing) when (...:parallel)
bar ();
}
behaves consistently; declarative and stand-alone directives aren't allowed
at that point, but constructs are parsed with the following statement as
the construct body and nothing or missing default on metadirective therefore
should handle the following statement as part of the if substatement instead
of having nothing as the substatement and bar done unconditionally after the
if.
2021-08-18 Jakub Jelinek <jakub@redhat.com>
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_NOTHING.
* c-pragma.c (omp_pragmas): Add nothing directive.
* c-omp.c (omp_directives): Uncomment nothing directive entry.
gcc/c/
* c-parser.c (c_parser_omp_nothing): New function.
(c_parser_pragma): Handle PRAGMA_OMP_NOTHING.
gcc/cp/
* parser.c (cp_parser_omp_nothing): New function.
(cp_parser_pragma): Handle PRAGMA_OMP_NOTHING.
gcc/testsuite/
* c-c++-common/gomp/nothing-1.c: New test.
* g++.dg/gomp/attrs-1.C (bar): Add nothing directive test.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* g++.dg/gomp/attrs-9.C: Likewise.
libgomp/
* testsuite/libgomp.c-c++-common/nothing-1.c: New test.
Fortran version to commit e45483c7c4,
which implemented OpenMP's scope construct for C and C++.
Most testcases are based on the C testcases; it also contains some
testcases which existed previously but had no Fortran equivalent.
gcc/fortran/ChangeLog:
* dump-parse-tree.c (show_omp_node, show_code_node): Handle
EXEC_OMP_SCOPE.
* gfortran.h (enum gfc_statement): Add ST_OMP_(END_)SCOPE.
(enum gfc_exec_op): Add EXEC_OMP_SCOPE.
* match.h (gfc_match_omp_scope): New.
* openmp.c (OMP_SCOPE_CLAUSES): Define
(gfc_match_omp_scope): New.
(gfc_match_omp_cancellation_point, gfc_match_omp_end_nowait):
Improve error diagnostic.
(omp_code_to_statement): Handle ST_OMP_SCOPE.
(gfc_resolve_omp_directive): Handle EXEC_OMP_SCOPE.
* parse.c (decode_omp_directive, next_statement,
gfc_ascii_statement, parse_omp_structured_block,
parse_executable): Handle OpenMP's scope construct.
* resolve.c (gfc_resolve_blocks): Likewise
* st.c (gfc_free_statement): Likewise
* trans-openmp.c (gfc_trans_omp_scope): New.
(gfc_trans_omp_directive): Call it.
* trans.c (trans_code): handle EXEC_OMP_SCOPE.
libgomp/ChangeLog:
* testsuite/libgomp.fortran/scope-1.f90: New test.
* testsuite/libgomp.fortran/task-reduction-16.f90: New test.
gcc/testsuite/ChangeLog:
* gfortran.dg/gomp/scan-1.f90:
* gfortran.dg/gomp/cancel-1.f90: New test.
* gfortran.dg/gomp/cancel-4.f90: New test.
* gfortran.dg/gomp/loop-4.f90: New test.
* gfortran.dg/gomp/nesting-1.f90: New test.
* gfortran.dg/gomp/nesting-2.f90: New test.
* gfortran.dg/gomp/nesting-3.f90: New test.
* gfortran.dg/gomp/nowait-1.f90: New test.
* gfortran.dg/gomp/reduction-task-1.f90: New test.
* gfortran.dg/gomp/reduction-task-2.f90: New test.
* gfortran.dg/gomp/reduction-task-2a.f90: New test.
* gfortran.dg/gomp/reduction-task-3.f90: New test.
* gfortran.dg/gomp/scope-1.f90: New test.
* gfortran.dg/gomp/scope-2.f90: New test.
This patch implements the OpenMP 5.1 scope construct, which is similar
to worksharing constructs in many regards, but isn't one of them.
The body of the construct is encountered by all threads though, it can
be nested in itself or intermixed with taskgroup and worksharing etc.
constructs can appear inside of it (but it can't be nested in
worksharing etc. constructs). The main purpose of the construct
is to allow reductions (normal and task ones) without the need to
close the parallel and reopen another one.
If it doesn't have task reductions, it can be implemented without
any new library support, with nowait it just does the privatizations
at the start if any and reductions before the end of the body, with
without nowait emits a normal GOMP_barrier{,_cancel} at the end too.
For task reductions, we need to ensure only one thread initializes
the task reduction library data structures and other threads copy from that,
so a new GOMP_scope_start routine is added to the library for that.
It acts as if the start of the scope construct is a nowait worksharing
construct (that is ok, it can't be nested in other worksharing
constructs and all threads need to encounter the start in the same
order) which does the task reduction initialization, but as the body
can have other scope constructs and/or worksharing constructs, that is
all where we use this dummy worksharing construct. With task reductions,
the construct must not have nowait and ends with a GOMP_barrier{,_cancel},
followed by task reductions followed by GOMP_workshare_task_reduction_unregister.
Only C/C++ FE support is done.
2021-08-17 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.def (OMP_SCOPE): New tree code.
* tree.h (OMP_SCOPE_BODY, OMP_SCOPE_CLAUSES): Define.
* tree-nested.c (convert_nonlocal_reference_stmt,
convert_local_reference_stmt, convert_gimple_call): Handle
GIMPLE_OMP_SCOPE.
* tree-pretty-print.c (dump_generic_node): Handle OMP_SCOPE.
* gimple.def (GIMPLE_OMP_SCOPE): New gimple code.
* gimple.c (gimple_build_omp_scope): New function.
(gimple_copy): Handle GIMPLE_OMP_SCOPE.
* gimple.h (gimple_build_omp_scope): Declare.
(gimple_has_substatements): Handle GIMPLE_OMP_SCOPE.
(gimple_omp_scope_clauses, gimple_omp_scope_clauses_ptr,
gimple_omp_scope_set_clauses): New inline functions.
(CASE_GIMPLE_OMP): Add GIMPLE_OMP_SCOPE.
* gimple-pretty-print.c (dump_gimple_omp_scope): New function.
(pp_gimple_stmt_1): Handle GIMPLE_OMP_SCOPE.
* gimple-walk.c (walk_gimple_stmt): Likewise.
* gimple-low.c (lower_stmt): Likewise.
* gimplify.c (is_gimple_stmt): Handle OMP_MASTER.
(gimplify_scan_omp_clauses): For task reductions, handle OMP_SCOPE
like ORT_WORKSHARE constructs. Adjust diagnostics for %<scope%>
allowing task reductions. Reject inscan reductions on scope.
(omp_find_stores_stmt): Handle GIMPLE_OMP_SCOPE.
(gimplify_omp_workshare, gimplify_expr): Handle OMP_SCOPE.
* tree-inline.c (remap_gimple_stmt): Handle GIMPLE_OMP_SCOPE.
(estimate_num_insns): Likewise.
* omp-low.c (build_outer_var_ref): Look through GIMPLE_OMP_SCOPE
contexts if var isn't privatized there.
(check_omp_nesting_restrictions): Handle GIMPLE_OMP_SCOPE.
(scan_omp_1_stmt): Likewise.
(maybe_add_implicit_barrier_cancel): Look through outer
scope constructs.
(lower_omp_scope): New function.
(lower_omp_task_reductions): Handle OMP_SCOPE.
(lower_omp_1): Handle GIMPLE_OMP_SCOPE.
(diagnose_sb_1, diagnose_sb_2): Likewise.
* omp-expand.c (expand_omp_single): Support also GIMPLE_OMP_SCOPE.
(expand_omp): Handle GIMPLE_OMP_SCOPE.
(omp_make_gimple_edges): Likewise.
* omp-builtins.def (BUILT_IN_GOMP_SCOPE_START): New built-in.
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_SCOPE.
* c-pragma.c (omp_pragmas): Add scope construct.
* c-omp.c (omp_directives): Uncomment scope directive entry.
gcc/c/
* c-parser.c (OMP_SCOPE_CLAUSE_MASK): Define.
(c_parser_omp_scope): New function.
(c_parser_omp_construct): Handle PRAGMA_OMP_SCOPE.
gcc/cp/
* parser.c (OMP_SCOPE_CLAUSE_MASK): Define.
(cp_parser_omp_scope): New function.
(cp_parser_omp_construct, cp_parser_pragma): Handle PRAGMA_OMP_SCOPE.
* pt.c (tsubst_expr): Handle OMP_SCOPE.
gcc/testsuite/
* c-c++-common/gomp/nesting-2.c (foo): Add scope and masked
construct tests.
* c-c++-common/gomp/scan-1.c (f3): Add scope construct test..
* c-c++-common/gomp/cancel-1.c (f2): Add scope and masked
construct tests.
* c-c++-common/gomp/reduction-task-2.c (bar): Add scope construct
test. Adjust diagnostics for the addition of scope.
* c-c++-common/gomp/loop-1.c (f5): Add master, masked and scope
construct tests.
* c-c++-common/gomp/clause-dups-1.c (f1): Add scope construct test.
* gcc.dg/gomp/nesting-1.c (f1, f2, f3): Add scope construct tests.
* c-c++-common/gomp/scope-1.c: New test.
* c-c++-common/gomp/scope-2.c: New test.
* g++.dg/gomp/attrs-1.C (bar): Add scope construct tests.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* gfortran.dg/gomp/reduction4.f90: Adjust expected diagnostics.
* gfortran.dg/gomp/reduction7.f90: Likewise.
libgomp/
* Makefile.am (libgomp_la_SOURCES): Add scope.c
* Makefile.in: Regenerated.
* libgomp_g.h (GOMP_scope_start): Declare.
* libgomp.map: Add GOMP_scope_start@@GOMP_5.1.
* scope.c: New file.
* testsuite/libgomp.c-c++-common/scope-1.c: New test.
* testsuite/libgomp.c-c++-common/task-reduction-16.c: New test.
... for 'gcc/gimplify.c:gimplify_scan_omp_clauses' changes in recent
commit d0befed793 "openmp: Add support
for OpenMP 5.1 masked construct".
libgomp/
* testsuite/libgomp.oacc-c-c++-common/static-variable-1.c: Adjust.
This construct has been introduced as a replacement for master
construct, but unlike that construct is slightly more general,
has an optional clause which allows to choose which thread
will be the one running the region, it can be some other thread
than the master (primary) thread with number 0, or it could be no
threads or multiple threads (then of course one needs to be careful
about data races).
It is way too early to deprecate the master construct though, we don't
even have OpenMP 5.0 fully implemented, it has been deprecated in 5.1,
will be also in 5.2 and removed in 6.0. But even then it will likely
be a good idea to just -Wdeprecated warn about it and still accept it.
The patch also contains something I should have done much earlier,
for clauses that accept some integral expression where we only care
about the value, forces during gimplification that value into
either a min invariant (as before), SSA_NAME or a fresh temporary,
but never e.g. a user VAR_DECL, so that for those clauses we don't
need to worry about adjusting it.
2021-08-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.def (OMP_MASKED): New tree code.
* tree-core.h (enum omp_clause_code): Add OMP_CLAUSE_FILTER.
* tree.h (OMP_MASKED_BODY, OMP_MASKED_CLAUSES, OMP_MASKED_COMBINED,
OMP_CLAUSE_FILTER_EXPR): Define.
* tree.c (omp_clause_num_ops): Add OMP_CLAUSE_FILTER entry.
(omp_clause_code_name): Likewise.
(walk_tree_1): Handle OMP_CLAUSE_FILTER.
* tree-nested.c (convert_nonlocal_omp_clauses,
convert_local_omp_clauses): Handle OMP_CLAUSE_FILTER.
(convert_nonlocal_reference_stmt, convert_local_reference_stmt,
convert_gimple_call): Handle GIMPLE_OMP_MASTER.
* tree-pretty-print.c (dump_omp_clause): Handle OMP_CLAUSE_FILTER.
(dump_generic_node): Handle OMP_MASTER.
* gimple.def (GIMPLE_OMP_MASKED): New gimple code.
* gimple.c (gimple_build_omp_masked): New function.
(gimple_copy): Handle GIMPLE_OMP_MASKED.
* gimple.h (gimple_build_omp_masked): Declare.
(gimple_has_substatements): Handle GIMPLE_OMP_MASKED.
(gimple_omp_masked_clauses, gimple_omp_masked_clauses_ptr,
gimple_omp_masked_set_clauses): New inline functions.
(CASE_GIMPLE_OMP): Add GIMPLE_OMP_MASKED.
* gimple-pretty-print.c (dump_gimple_omp_masked): New function.
(pp_gimple_stmt_1): Handle GIMPLE_OMP_MASKED.
* gimple-walk.c (walk_gimple_stmt): Likewise.
* gimple-low.c (lower_stmt): Likewise.
* gimplify.c (is_gimple_stmt): Handle OMP_MASTER.
(gimplify_scan_omp_clauses): Handle OMP_CLAUSE_FILTER. For clauses
that take one expression rather than decl or constant, force
gimplification of that into a SSA_NAME or temporary unless min
invariant.
(gimplify_adjust_omp_clauses): Handle OMP_CLAUSE_FILTER.
(gimplify_expr): Handle OMP_MASKED.
* tree-inline.c (remap_gimple_stmt): Handle GIMPLE_OMP_MASKED.
(estimate_num_insns): Likewise.
* omp-low.c (scan_sharing_clauses): Handle OMP_CLAUSE_FILTER.
(check_omp_nesting_restrictions): Handle GIMPLE_OMP_MASKED. Adjust
diagnostics for existence of masked construct.
(scan_omp_1_stmt, lower_omp_master, lower_omp_1, diagnose_sb_1,
diagnose_sb_2): Handle GIMPLE_OMP_MASKED.
* omp-expand.c (expand_omp_synch, expand_omp, omp_make_gimple_edges):
Likewise.
gcc/c-family/
* c-pragma.h (enum pragma_kind): Add PRAGMA_OMP_MASKED.
(enum pragma_omp_clause): Add PRAGMA_OMP_CLAUSE_FILTER.
* c-pragma.c (omp_pragmas_simd): Add masked construct.
* c-common.h (enum c_omp_clause_split): Add C_OMP_CLAUSE_SPLIT_MASKED
enumerator.
(c_finish_omp_masked): Declare.
* c-omp.c (c_finish_omp_masked): New function.
(c_omp_split_clauses): Handle combined masked constructs.
gcc/c/
* c-parser.c (c_parser_omp_clause_name): Parse filter clause name.
(c_parser_omp_clause_filter): New function.
(c_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_FILTER.
(OMP_MASKED_CLAUSE_MASK): Define.
(c_parser_omp_masked): New function.
(c_parser_omp_parallel): Handle parallel masked.
(c_parser_omp_construct): Handle PRAGMA_OMP_MASKED.
* c-typeck.c (c_finish_omp_clauses): Handle OMP_CLAUSE_FILTER.
gcc/cp/
* parser.c (cp_parser_omp_clause_name): Parse filter clause name.
(cp_parser_omp_clause_filter): New function.
(cp_parser_omp_all_clauses): Handle PRAGMA_OMP_CLAUSE_FILTER.
(OMP_MASKED_CLAUSE_MASK): Define.
(cp_parser_omp_masked): New function.
(cp_parser_omp_parallel): Handle parallel masked.
(cp_parser_omp_construct, cp_parser_pragma): Handle PRAGMA_OMP_MASKED.
* semantics.c (finish_omp_clauses): Handle OMP_CLAUSE_FILTER.
* pt.c (tsubst_omp_clauses): Likewise.
(tsubst_expr): Handle OMP_MASKED.
gcc/testsuite/
* c-c++-common/gomp/clauses-1.c (bar): Add tests for combined masked
constructs with clauses.
* c-c++-common/gomp/clauses-5.c (foo): Add testcase for filter clause.
* c-c++-common/gomp/clause-dups-1.c (f1): Likewise.
* c-c++-common/gomp/masked-1.c: New test.
* c-c++-common/gomp/masked-2.c: New test.
* c-c++-common/gomp/masked-combined-1.c: New test.
* c-c++-common/gomp/masked-combined-2.c: New test.
* c-c++-common/goacc/uninit-if-clause.c: Remove xfails.
* g++.dg/gomp/block-11.C: New test.
* g++.dg/gomp/tpl-masked-1.C: New test.
* g++.dg/gomp/attrs-1.C (bar): Add tests for masked construct and
combined masked constructs with clauses in attribute syntax.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* gcc.dg/gomp/nesting-1.c (f1, f2): Add tests for masked construct
nesting.
* gfortran.dg/goacc/host_data-tree.f95: Allow also SSA_NAMEs in if
clause.
* gfortran.dg/goacc/kernels-tree.f95: Likewise.
libgomp/
* testsuite/libgomp.c-c++-common/masked-1.c: New test.
In OpenMP 5.1 "master thread" was changed to "primary thread" and
the proc_bind clause and the OMP_PROC_BIND environment variable
now take 'primary' as argument as alias for 'master', while the
latter is deprecated.
This commit accepts 'primary' and adds the named constant
omp_proc_bind_primary and changes 'master thread' in the
documentation; however, given that not even OpenMP 5.0 is
fully supported, omp_display_env and the dumps currently
still output 'master' and there is no deprecation warning
when using the 'master' in the proc_bind clause.
gcc/c/ChangeLog:
* c-parser.c (c_parser_omp_clause_proc_bind): Accept
'primary' as alias for 'master'.
gcc/cp/ChangeLog:
* parser.c (cp_parser_omp_clause_proc_bind): Accept
'primary' as alias for 'master'.
gcc/fortran/ChangeLog:
* gfortran.h (gfc_omp_proc_bind_kind): Add OMP_PROC_BIND_PRIMARY.
* dump-parse-tree.c (show_omp_clauses): Add TODO comment to
change 'master' to 'primary' in proc_bind for OpenMP 5.1.
* intrinsic.texi (OMP_LIB): Mention OpenMP 5.1; add
omp_proc_bind_primary.
* openmp.c (gfc_match_omp_clauses): Accept
'primary' as alias for 'master'.
* trans-openmp.c (gfc_trans_omp_clauses): Handle
OMP_PROC_BIND_PRIMARY.
gcc/ChangeLog:
* tree-core.h (omp_clause_proc_bind_kind): Add
OMP_CLAUSE_PROC_BIND_PRIMARY.
* tree-pretty-print.c (dump_omp_clause): Add TODO comment to
change 'master' to 'primary' in proc_bind for OpenMP 5.1.
libgomp/ChangeLog:
* env.c (parse_bind_var): Accept 'primary' as alias for
'master'.
(omp_display_env): Add TODO comment to
change 'master' to 'primary' in proc_bind for OpenMP 5.1.
* libgomp.texi: Change 'master thread' to 'primary thread'
in line with OpenMP 5.1.
(omp_get_proc_bind): Add omp_proc_bind_primary and note that
omp_proc_bind_master is an alias of it.
(OMP_PROC_BIND): Mention 'PRIMARY'.
* omp.h.in (__GOMP_DEPRECATED_5_1): Define.
(omp_proc_bind_primary): Add.
(omp_proc_bind_master): Deprecate for OpenMP 5.1.
* omp_lib.f90.in (omp_proc_bind_primary): Add.
(omp_proc_bind_master): Deprecate for OpenMP 5.1.
* omp_lib.h.in (omp_proc_bind_primary): Add.
* testsuite/libgomp.c/affinity-1.c: Check that
'primary' works and is identical to 'master'.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/pr61486-2.c: Duplicate one proc_bind(master)
testcase and test proc_bind(primary) instead.
* gfortran.dg/gomp/affinity-1.f90: Likewise.
This patch implements the omp_get_device_num library routine, specified in
OpenMP 5.0.
GOMP_DEVICE_NUM_VAR is a macro symbol which defines name of a "device number"
variable, is defined on the device-side libgomp, has it's address returned to
host-side libgomp during device initialization, and the host libgomp then
sets its value to the designated device number.
libgomp/ChangeLog:
* icv-device.c (omp_get_device_num): New API function, host side.
* fortran.c (omp_get_device_num_): New interface function.
* libgomp-plugin.h (GOMP_DEVICE_NUM_VAR): Define macro symbol.
* libgomp.map (OMP_5.0.2): New version space with omp_get_device_num,
omp_get_device_num_.
* libgomp.texi (omp_get_device_num): Add documentation for new API
function.
* omp.h.in (omp_get_device_num): Add declaration.
* omp_lib.f90.in (omp_get_device_num): Likewise.
* omp_lib.h.in (omp_get_device_num): Likewise.
* target.c (gomp_load_image_to_device): If additional entry for device
number exists at end of returned entries from 'load_image_func' hook,
copy the assigned device number over to the device variable.
* config/gcn/icv-device.c (GOMP_DEVICE_NUM_VAR): Define static global.
(omp_get_device_num): New API function, device side.
* plugin/plugin-gcn.c ("symcat.h"): Add include.
(GOMP_OFFLOAD_load_image): Add addresses of device GOMP_DEVICE_NUM_VAR
at end of returned 'target_table' entries.
* config/nvptx/icv-device.c (GOMP_DEVICE_NUM_VAR): Define static global.
(omp_get_device_num): New API function, device side.
* plugin/plugin-nvptx.c ("symcat.h"): Add include.
(GOMP_OFFLOAD_load_image): Add addresses of device GOMP_DEVICE_NUM_VAR
at end of returned 'target_table' entries.
* testsuite/lib/libgomp.exp
(check_effective_target_offload_target_intelmic): New function for
testing for intelmic offloading.
* testsuite/libgomp.c-c++-common/target-45.c: New test.
* testsuite/libgomp.fortran/target10.f90: New test.
This is a new interface which is easily implemented using the
already existing code for the handling of the OMP_DISPLAY_ENV
environment variable.
libgomp/
* env.c (wait_policy, stacksize): New static variables,
move out of handle_omp_display_env.
(omp_display_env): New function. The meat of the old
handle_omp_display_env function.
(handle_omp_display_env): Change to not take parameters
and instead use the global variables. Only perform
parsing, defer to omp_display_env for the implementation.
(initialize_env): Remove local variables wait_policy and
stacksize. Don't pass parameters to handle_omp_display_env.
* fortran.c: Add ialias_redirect for omp_display_env.
(omp_display_env_, omp_display_env_8_): New functions.
* libgomp.map (OMP_5.1): New version. Add omp_display_env,
omp_display_env_, and omp_display_env_8_.
* omp.h.in: Declare omp_display_env.
* omp_lib.f90.in: Likewise.
* omp_lib.h.in: Likewise.
The host data might not be computed yet (by an earlier asynchronous compute
region, for example.
libgomp/
* target.c (gomp_coalesce_buf_add): Update comment.
(gomp_copy_host2dev, gomp_map_vars_internal): Don't expect to see
'aq && cbuf'.
(gomp_map_vars_internal): Only 'if (!aq)', do
'gomp_coalesce_buf_add'.
* testsuite/libgomp.oacc-c-c++-common/async-data-1-2.c: Remove
XFAIL.
Co-Authored-By: Julian Brown <julian@codesourcery.com>
This patch fixes several places in libgomp/target.c where "ephemeral" data
(on the stack or in temporary heap locations) may be used as the source of
an asynchronous host-to-device copy that may not complete before the host
data disappears.
An existing, but flawed, workaround for this problem in the AMD GCN
libgomp offloading plugin is currently present on mainline, and was
posted for the og9 branch here:
https://gcc.gnu.org/legacy-ml/gcc-patches/2019-08/msg00901.html
and previous versions of this patch were posted here (for mainline/og9):
https://gcc.gnu.org/legacy-ml/gcc-patches/2019-11/msg01482.htmlhttps://gcc.gnu.org/legacy-ml/gcc-patches/2019-09/msg01026.html
libgomp/
* libgomp.h (gomp_copy_host2dev): Update prototype.
* oacc-mem.c (memcpy_tofrom_device, update_dev_host): Add new
argument to gomp_copy_host2dev (false).
* plugin/plugin-gcn.c (struct copy_data): Remove free_src field.
(copy_data): Don't free src.
(queue_push_copy): Remove free_src handling.
(GOMP_OFFLOAD_dev2dev): Update call to queue_push_copy.
(GOMP_OFFLOAD_openacc_async_host2dev): Remove source-data
snapshotting.
(GOMP_OFFLOAD_openacc_async_dev2host): Update call to
queue_push_copy.
* target.c (goacc_device_copy_async): Add SRCADDR_ORIG parameter.
(gomp_copy_host2dev): Add EPHEMERAL parameter. Snapshot source
data when true, and set up deferred freeing of temporary buffer.
(gomp_copy_dev2host): Update call to goacc_device_copy_async.
(gomp_map_vars_existing, gomp_map_pointer, gomp_attach_pointer)
(gomp_detach_pointer, gomp_map_vars_internal, gomp_update): Update
calls to gomp_copy_host2dev with appropriate ephemeral argument.
* testsuite/libgomp.oacc-c-c++-common/async-data-1-1.c: Remove
XFAIL.
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
libgomp/
* testsuite/libgomp.oacc-c-c++-common/async-data-1-1.c: New file.
* testsuite/libgomp.oacc-c-c++-common/async-data-1-2.c: Likewise.
Co-Authored-By: Tom de Vries <tom@codesourcery.com>
... as noticed with GCN offloading.
Fix-up for r271346 (commit 5fae049dc2)
"OpenACC Profiling Interface (incomplete)".
libgomp/
* testsuite/libgomp.oacc-c-c++-common/acc_prof-init-1.c: Clarify
sequencing of 'async' data copying vs. profiling events.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-parallel-1.c:
Likewise.
With yesterday's commit 9f2bc5077d "[gcn]
Work-around libgomp 'error: array subscript 0 is outside array bounds of
‘__lds struct gomp_thread * __lds[0]’ [-Werror=array-bounds]' [PR101484]",
I did defuse the "unexpected" '-Werror=array-bounds' diagnostics that we see
as of commit a110855667 "Correct handling of
variable offset minus constant in -Warray-bounds [PR100137]". However, these
'#pragma GCC diagnostic [...]' directives cause some code generation changes
(that seems unexpected, problematic!), which results in a lot (ten thousands)
of 'GCN team arena exhausted' run-time diagnostics, also leading to a few
FAILs:
PASS: libgomp.c/../libgomp.c-c++-common/for-11.c (test for excess errors)
[-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/for-11.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/for-12.c (test for excess errors)
[-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/for-12.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/for-3.c (test for excess errors)
[-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/for-3.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/for-5.c (test for excess errors)
[-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/for-5.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/for-6.c (test for excess errors)
[-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/for-6.c execution test
PASS: libgomp.c/../libgomp.c-c++-common/for-9.c (test for excess errors)
[-PASS:-]{+FAIL:+} libgomp.c/../libgomp.c-c++-common/for-9.c execution test
Same for 'libgomp.c++'.
It remains to be analyzed how '#pragma GCC diagnostic [...]' directives can
cause code generation changes; for now I'm working around the "unexpected"
'-Werror=array-bounds' diagnostics differently.
Overall, still awaiting a different solution, of course.
libgomp/
PR target/101484
* configure.tgt [amdgcn*-*-*] (XCFLAGS): Add
'-Wno-error=array-bounds'.
* config/gcn/team.c: Remove '-Werror=array-bounds' work-around.
* libgomp.h [__AMDGCN__]: Likewise.
... seen as of commit a110855667 "Correct
handling of variable offset minus constant in -Warray-bounds [PR100137]".
Awaiting a different solution, of course.
libgomp/
PR target/101484
* config/gcn/team.c: Apply '-Werror=array-bounds' work-around.
* libgomp.h [__AMDGCN__]: Likewise.
sem.h is included in between # pragma GCC visibility push(hidden)
and # pragma GCC visibility pop and includes limits.h there, which
since the introduction of sysconf declaration in recent glibcs
in there causes trouble. libgomp assumes it is compiled by gcc,
so we don't really need to include limits.h there and can use
-__INT_MAX__ - 1 instead (which clang and icc support too for years).
2021-07-13 Jakub Jelinek <jakub@redhat.com>
Florian Weimer <fweimer@redhat.com>
* config/linux/sem.h: Don't include limits.h.
(SEM_WAIT): Define to -__INT_MAX__ - 1 instead of INT_MIN.
* config/linux/affinity.c: Include limits.h.
As the testcase shows, the special treatment of && and || reduction combiners
where we expand them as omp_out = (omp_out != 0) && (omp_in != 0) (or with ||)
is not needed just for &&/|| on floating point or complex types, but for all
&&/|| reductions - when expanded as omp_out = omp_out && omp_in (not in C but
GENERIC) it is actually gimplified into NOP_EXPRs to bool from both operands,
which turns non-zero values multiple of 2 into 0 rather than 1.
This patch just treats all &&/|| the same and furthermore uses bool type
instead of int for the comparisons.
2021-07-01 Jakub Jelinek <jakub@redhat.com>
PR middle-end/94366
gcc/
* omp-low.c (lower_rec_input_clauses): Rename is_fp_and_or to
is_truth_op, set it for TRUTH_*IF_EXPR regardless of new_var's type,
use boolean_type_node instead of integer_type_node as NE_EXPR type.
(lower_reduction_clauses): Likewise.
libgomp/
* testsuite/libgomp.c-c++-common/pr94366.c: New test.
As -foffload={options,targets,targets=options} is very convoluted,
it has been split into -foffload=targets (supporting the old syntax
for backward compatibilty) and -foffload-options={options,target=options}.
Only the new syntax is documented.
Additionally, -foffload=default is supported, which can reset the
devices after -foffload=disable / -foffload=targets to the default,
if needed.
gcc/ChangeLog:
PR other/67300
* common.opt (-foffload=): Update description.
(-foffload-options=): New.
* doc/invoke.texi (C Language Options): Document
-foffload and -foffload-options.
* gcc.c (check_offload_target_name): New, split off from
handle_foffload_option.
(check_foffload_target_names): New.
(handle_foffload_option): Handle -foffload=default.
(driver_handle_option): Update for -foffload-options.
* lto-opts.c (lto_write_options): Use -foffload-options
instead of -foffload.
* lto-wrapper.c (merge_and_complain, append_offload_options):
Likewise.
* opts.c (common_handle_option): Likewise.
libgomp/ChangeLog:
PR other/67300
* testsuite/libgomp.c-c++-common/reduction-16.c: Replace
-foffload=nvptx-none= by -foffload-options=nvptx-none= to
avoid disabling other offload targets.
* testsuite/libgomp.c-c++-common/reduction-5.c: Likewise.
* testsuite/libgomp.c-c++-common/reduction-6.c: Likewise.
* testsuite/libgomp.c/target-44.c: Likewise.
Disable some more parts of the test as firstprivate does not work yet
due to PR fortran/90742.
libgomp/
* testsuite/libgomp.fortran/defaultmap-8.f90 (bar): Determine whether
target has shared memory and disable some scalar pointer/allocatable
checks if not as firstprivate does not work.
The dg-shouldfail testcase libgomp.c-c++-common/struct-elem-5.c does not
properly fail for non-shared address space offloading. Adjust testcase
to limit testing only for "target offload_device_nonshared_as".
libgomp/ChangeLog:
PR testsuite/101114
* testsuite/libgomp.c-c++-common/struct-elem-5.c:
Add "target offload_device_nonshared_as" condition for enabling test.
This patch adds support for in_reduction clause on target construct, though
for now only for synchronous targets (without nowait clause).
The encountering thread in that case runs the target task and blocks until
the target region ends, so it is implemented by remapping it before entering
the target, initializing the private copy if not yet initialized for the
current thread and then using the remapped addresses for the mapping
addresses.
For nowait combined with in_reduction the patch contains a hack where the
nowait clause is ignored. To implement it correctly, I think we would need
to create a new private variable for the in_reduction and initialize it before
doing the async target and adjust the map addresses to that private variable
and then pass a function pointer to the library routine with code where the callback
would remap the address to the current threads private variable and use in_reduction
combiner to combine the private variable we've created into the thread's copy.
The library would then need to make sure that the routine is called in some thread
participating in the parallel (and not in an unshackeled thread).
2021-06-24 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.h (OMP_CLAUSE_MAP_IN_REDUCTION): Document meaning for OpenMP.
* gimplify.c (gimplify_scan_omp_clauses): For OpenMP map clauses
with OMP_CLAUSE_MAP_IN_REDUCTION flag partially defer gimplification
of non-decl OMP_CLAUSE_DECL. For OMP_CLAUSE_IN_REDUCTION on
OMP_TARGET user outer_ctx instead of ctx for placeholders and
initializer/combiner gimplification.
* omp-low.c (scan_sharing_clauses): Handle OMP_CLAUSE_MAP_IN_REDUCTION
on target constructs.
(lower_rec_input_clauses): Likewise.
(lower_omp_target): Likewise.
* omp-expand.c (expand_omp_target): Temporarily ignore nowait clause
on target if in_reduction is present.
gcc/c-family/
* c-common.h (enum c_omp_region_type): Add C_ORT_TARGET and
C_ORT_OMP_TARGET.
* c-omp.c (c_omp_split_clauses): For OMP_CLAUSE_IN_REDUCTION on
combined target constructs also add map (always, tofrom:) clause.
gcc/c/
* c-parser.c (omp_split_clauses): Pass C_ORT_OMP_TARGET instead of
C_ORT_OMP for clauses on target construct.
(OMP_TARGET_CLAUSE_MASK): Add in_reduction clause.
(c_parser_omp_target): For non-combined target add
map (always, tofrom:) clauses for OMP_CLAUSE_IN_REDUCTION. Pass
C_ORT_OMP_TARGET to c_finish_omp_clauses.
* c-typeck.c (handle_omp_array_sections): Adjust ort handling
for addition of C_ORT_OMP_TARGET and simplify, mapping clauses are
never present on C_ORT_*DECLARE_SIMD.
(c_finish_omp_clauses): Likewise. Handle OMP_CLAUSE_IN_REDUCTION
on C_ORT_OMP_TARGET, set OMP_CLAUSE_MAP_IN_REDUCTION on
corresponding map clauses.
gcc/cp/
* parser.c (cp_omp_split_clauses): Pass C_ORT_OMP_TARGET instead of
C_ORT_OMP for clauses on target construct.
(OMP_TARGET_CLAUSE_MASK): Add in_reduction clause.
(cp_parser_omp_target): For non-combined target add
map (always, tofrom:) clauses for OMP_CLAUSE_IN_REDUCTION. Pass
C_ORT_OMP_TARGET to finish_omp_clauses.
* semantics.c (handle_omp_array_sections_1): Adjust ort handling
for addition of C_ORT_OMP_TARGET and simplify, mapping clauses are
never present on C_ORT_*DECLARE_SIMD.
(handle_omp_array_sections): Likewise.
(finish_omp_clauses): Likewise. Handle OMP_CLAUSE_IN_REDUCTION
on C_ORT_OMP_TARGET, set OMP_CLAUSE_MAP_IN_REDUCTION on
corresponding map clauses.
* pt.c (tsubst_expr): Pass C_ORT_OMP_TARGET instead of C_ORT_OMP for
clauses on target construct.
gcc/testsuite/
* c-c++-common/gomp/target-in-reduction-1.c: New test.
* c-c++-common/gomp/clauses-1.c: Add in_reduction clauses on
target or combined target constructs.
libgomp/
* testsuite/libgomp.c-c++-common/target-in-reduction-1.c: New test.
* testsuite/libgomp.c-c++-common/target-in-reduction-2.c: New test.
* testsuite/libgomp.c++/target-in-reduction-1.C: New test.
* testsuite/libgomp.c++/target-in-reduction-2.C: New test.
The following testcase FAILs, because the UDR combiner is invoked incorrectly.
lower_omp_rec_clauses expects that when it sets
DECL_VALUE_EXPR/DECL_HAS_VALUE_EXPR_P
for both the placeholder and the var that everything will be properly
regimplified, but as the variable in question is a PARM_DECL rather than
VAR_DECL, lower_omp_regimplify_p doesn't say that it should be regimplified
and so it is not.
2021-06-23 Jakub Jelinek <jakub@redhat.com>
PR middle-end/101167
* omp-low.c (lower_omp_regimplify_p): Regimplify also PARM_DECLs
and RESULT_DECLs that have DECL_HAS_VALUE_EXPR_P set.
* testsuite/libgomp.c-c++-common/task-reduction-15.c: New test.
This patch implement OpenMP 5.0 requirements of incrementing/decrementing
the reference count of a mapped structure at most once (across all elements)
on a construct.
This is implemented by pulling in libgomp/hashtab.h and using htab_t as a
pointer set. Structure element list siblings also have pointers-to-refcounts
linked together, to naturally achieve uniform increment/decrement without
repeating.
There are still some questions on whether using such a htab_t based set is
faster/slower than using a sorted pointer array based implementation. This
is to be researched on later.
libgomp/ChangeLog:
* hashtab.h (htab_clear): New function with initialization code
factored out from...
(htab_create): ...here, adjust to use htab_clear function.
* libgomp.h (REFCOUNT_SPECIAL): New symbol to denote range of
special refcount values, add comments.
(REFCOUNT_INFINITY): Adjust definition to use REFCOUNT_SPECIAL.
(REFCOUNT_LINK): Likewise.
(REFCOUNT_STRUCTELEM): New special refcount range for structure
element siblings.
(REFCOUNT_STRUCTELEM_P): Macro for testing for structure element
sibling maps.
(REFCOUNT_STRUCTELEM_FLAG_FIRST): Flag to indicate first sibling.
(REFCOUNT_STRUCTELEM_FLAG_LAST): Flag to indicate last sibling.
(REFCOUNT_STRUCTELEM_FIRST_P): Macro to test _FIRST flag.
(REFCOUNT_STRUCTELEM_LAST_P): Macro to test _LAST flag.
(struct splay_tree_key_s): Add structelem_refcount and
structelem_refcount_ptr fields into a union with dynamic_refcount.
Add comments.
(gomp_map_vars): Delete declaration.
(gomp_map_vars_async): Likewise.
(gomp_unmap_vars): Likewise.
(gomp_unmap_vars_async): Likewise.
(goacc_map_vars): New declaration.
(goacc_unmap_vars): Likewise.
* oacc-mem.c (acc_map_data): Adjust to use goacc_map_vars.
(goacc_enter_datum): Likewise.
(goacc_enter_data_internal): Likewise.
* oacc-parallel.c (GOACC_parallel_keyed): Adjust to use goacc_map_vars
and goacc_unmap_vars.
(GOACC_data_start): Adjust to use goacc_map_vars.
(GOACC_data_end): Adjust to use goacc_unmap_vars.
* target.c (hash_entry_type): New typedef.
(htab_alloc): New function hook for hashtab.h.
(htab_free): Likewise.
(htab_hash): Likewise.
(htab_eq): Likewise.
(hashtab.h): Add file include.
(gomp_increment_refcount): New function.
(gomp_decrement_refcount): Likewise.
(gomp_map_vars_existing): Add refcount_set parameter, adjust to use
gomp_increment_refcount.
(gomp_map_fields_existing): Add refcount_set parameter, adjust calls
to gomp_map_vars_existing.
(gomp_map_vars_internal): Add refcount_set parameter, add local openmp_p
variable to guard OpenMP specific paths, adjust calls to
gomp_map_vars_existing, add structure element sibling splay_tree_key
sequence creation code, adjust Fortran map case to avoid increment
under OpenMP.
(gomp_map_vars): Adjust to static, add refcount_set parameter, manage
local refcount_set if caller passed in NULL, adjust call to
gomp_map_vars_internal.
(gomp_map_vars_async): Adjust and rename into...
(goacc_map_vars): ...this new function, adjust call to
gomp_map_vars_internal.
(gomp_remove_splay_tree_key): New function with code factored out from
gomp_remove_var_internal.
(gomp_remove_var_internal): Add code to handle removing multiple
splay_tree_key sequence for structure elements, adjust code to use
gomp_remove_splay_tree_key for splay-tree key removal.
(gomp_unmap_vars_internal): Add refcount_set parameter, adjust to use
gomp_decrement_refcount.
(gomp_unmap_vars): Adjust to static, add refcount_set parameter, manage
local refcount_set if caller passed in NULL, adjust call to
gomp_unmap_vars_internal.
(gomp_unmap_vars_async): Adjust and rename into...
(goacc_unmap_vars): ...this new function, adjust call to
gomp_unmap_vars_internal.
(GOMP_target): Manage refcount_set and adjust calls to gomp_map_vars and
gomp_unmap_vars.
(GOMP_target_ext): Likewise.
(gomp_target_data_fallback): Adjust call to gomp_map_vars.
(GOMP_target_data): Likewise.
(GOMP_target_data_ext): Likewise.
(GOMP_target_end_data): Adjust call to gomp_unmap_vars.
(gomp_exit_data): Add refcount_set parameter, adjust to use
gomp_decrement_refcount, adjust to queue splay-tree keys for removal
after main loop.
(GOMP_target_enter_exit_data): Manage refcount_set and adjust calls to
gomp_map_vars and gomp_exit_data.
(gomp_target_task_fn): Likewise.
* testsuite/libgomp.c-c++-common/refcount-1.c: New testcase.
* testsuite/libgomp.c-c++-common/struct-elem-1.c: New testcase.
* testsuite/libgomp.c-c++-common/struct-elem-2.c: New testcase.
* testsuite/libgomp.c-c++-common/struct-elem-3.c: New testcase.
* testsuite/libgomp.c-c++-common/struct-elem-4.c: New testcase.
* testsuite/libgomp.c-c++-common/struct-elem-5.c: New testcase.
Move the OpenACC enter and exit data directives from using a single builtin to
having one each. For most purposes it was easy to tell which was which, from
the clauses given, but it's overhead we can easily avoid, and there may be
future uses where that isn't possible.
gcc/
* omp-builtins.def (BUILT_IN_GOACC_ENTER_EXIT_DATA): Split into...
(BUILT_IN_GOACC_ENTER_DATA, BUILT_IN_GOACC_EXIT_DATA): ... these.
* gimple.h (enum gf_mask): Split
'GF_OMP_TARGET_KIND_OACC_ENTER_EXIT_DATA' into
'GF_OMP_TARGET_KIND_OACC_ENTER_DATA' and
'GF_OMP_TARGET_KIND_OACC_EXIT_DATA'.
(is_gimple_omp_oacc): Update.
* gimple-pretty-print.c (dump_gimple_omp_target): Likewise.
* gimplify.c (gimplify_omp_target_update): Likewise.
* omp-expand.c (expand_omp_target, build_omp_regions_1)
(omp_make_gimple_edges): Likewise.
* omp-low.c (check_omp_nesting_restrictions, lower_omp_target):
Likewise.
gcc/testsuite/
* c-c++-common/goacc-gomp/nesting-fail-1.c: Adjust patterns.
* c-c++-common/goacc/finalize-1.c: Likewise.
* c-c++-common/goacc/mdc-1.c: Likewise.
* c-c++-common/goacc/nesting-fail-1.c: Likewise.
* c-c++-common/goacc/struct-enter-exit-data-1.c: Likewise.
* gfortran.dg/goacc/attach-descriptor.f90: Likewise.
* gfortran.dg/goacc/finalize-1.f: Likewise.
* gfortran.dg/goacc/mapping-tests-3.f90: Likewise.
libgomp/
* libgomp.map (GOACC_2.0.2): New symbol version.
* libgomp_g.h (GOACC_enter_data, GOACC_exit_data) New prototypes.
* oacc-mem.c (GOACC_enter_data, GOACC_exit_data) New functions.
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
libgomp/
* oacc-mem.c (goacc_enter_exit_data_internal): New function,
extracted from...
(GOACC_enter_exit_data): ... here.
(GOACC_declare): Use it.
Co-Authored-By: Andrew Stubbs <ams@codesourcery.com>
This deals with data management, after all.
Small fix-up for r230275 (commit 6e232ba424)
"[OpenACC] declare directive".
libgomp/
* oacc-parallel.c (GOACC_declare): Move...
* oacc-mem.c: ... here.
* libgomp_g.h: Adjust.
Given that we 'continue' for 'GOMP_MAP_POINTER', we cannot possibly encounter
it afterwards.
Small fix-up for r230275 (commit 6e232ba424)
"[OpenACC] declare directive".
libgomp/
* oacc-parallel.c (GOACC_declare): Clean up 'GOMP_MAP_POINTER'
handling.
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
The dsdotr and dsdoti variables uninitialized and the testcase fails e.g.
on i686-linux. Fixed by zero initialization.
2021-06-10 Jakub Jelinek <jakub@redhat.com>
PR tree-optimization/100981
* testsuite/libgomp.fortran/pr100981-2.f90 (cdcdot): Initialize
dsdotr and dsdoti to 0.
Don't add -march=i486 if atomic compare-and-swap is supported on 'int'.
This fixes libgomp tests with "-march=x86-64 -m32 -fcf-protection".
* testsuite/lib/libgomp.exp (libgomp_init): Don't add -march=i486
if atomic compare-and-swap is supported on 'int'.
The following fixes the SLP FMA patterns to preserve reduction
info and the reduction vectorization to consider internal function
call defs for the reduction stmt.
2021-06-09 Richard Biener <rguenther@suse.de>
PR tree-optimization/100981
gcc/
* tree-vect-loop.c (vect_create_epilog_for_reduction): Use
gimple_get_lhs to also handle calls.
* tree-vect-slp-patterns.c (complex_pattern::build): Transfer
reduction info.
gcc/testsuite/
* gfortran.dg/vect/pr100981-1.f90: New testcase.
libgomp/
* testsuite/libgomp.fortran/pr100981-2.f90: New testcase.
... which currently has *not* been forced to 'num_workers (1)'.
In addition to the testcases modified here, this also fixes:
FAIL: libgomp.oacc-c/../libgomp.oacc-c-c++-common/mode-transitions.c -DACC_DEVICE_TYPE_radeon=1 -DACC_MEM_SHARED=0 -foffload=amdgcn-amdhsa -O0 execution test
[Etc.]
mode-transitions.exe: [...]/libgomp.oacc-c-c++-common/mode-transitions.c:702: t17: Assertion `arr_b[i] == (i ^ 31) * 8' failed.
libgomp/
* plugin/plugin-gcn.c (gcn_exec): Force 'num_workers (1)'
unconditionally.
* testsuite/libgomp.oacc-c-c++-common/acc_prof-kernels-1.c:
Update.
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/routine-wv-2.c: Likewise.
..., by simplifying 'libgomp.oacc-c-c++-common/parallel-dims.c', and updating
the former correspondingly. '__builtin_goacc_parlevel_id' does the right thing
for all 'acc_device_*'.
Follow-up to commit 09e0ad6253 "Update OpenACC
tests for amdgcn".
libgomp/
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c: Simplify.
* testsuite/libgomp.oacc-fortran/parallel-dims-aux.c: Update.
... on top of r279378 (commit 26b74ed022)
"Update OpenACC tests for amdgcn".
libgomp/
* testsuite/libgomp.oacc-c-c++-common/acc_prof-kernels-1.c: Fix
for 'acc_device_radeon'.
That is, re-enable it for host-fallback, and enable it for GCN offloading.
Fix-up for r279378 (commit 26b74ed022)
"Update OpenACC tests for amdgcn".
libgomp/
* testsuite/libgomp.oacc-c-c++-common/async_queue-1.c: Don't
require 'openacc_nvidia_accel_selected'. Fix up for
'ACC_DEVICE_TYPE_radeon'.
The GCN support that got added in r278935 (commit
83caa34e2a) "Enable OpenACC GCN testing" was
forked before my r269107 (commit ee332b4a9a)
"[libgomp] Clarify difference between offload target, offload plugin, and
OpenACC device type", and didn't later pick up these changes.
No functional change.
libgomp/
* testsuite/lib/libgomp.exp
(check_effective_target_openacc_radeon_accel_selected):
Streamline.
This problem has been fixed long ago, in r267934 (commit
d41d952c9b) "[nvptx] Handle assignment to
gang-level reduction variable".
libgomp/
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c: Revert
PR80547 workaround.
Small fix-up for r267889 (commit 2b9d9e3937)
"[nvptx] Enable large vectors":
> * testsuite/libgomp.oacc-c-c++-common/parallel-dims.c: Expect vector
> length 2097152 to be reduced to 1024 instead of 32.
libgomp/
* testsuite/libgomp.oacc-c-c++-common/parallel-dims.c
<acc_device_nvidia>: Update comment.
When gcc is configured for nvptx offloading with --without-cuda-driver
and full CUDA isn't installed, many libgomp.oacc-*/* tests fail,
some of them because cuda.h header can't be found, others because
the tests can't be linked against -lcuda, -lcudart or -lcublas.
I usually only have akmod-nvidia and xorg-x11-drv-nvidia-cuda rpms
installed, so libcuda.so.1 can be dlopened and the offloading works,
but linking against those libraries isn't possible nor are the
headers around (for the plugin itself there is the fallback
libgomp/plugin/cuda/cuda.h).
The following patch adds 3 new effective targets and uses them in tests that
needs those.
2021-05-27 Jakub Jelinek <jakub@redhat.com>
* testsuite/lib/libgomp.exp (check_effective_target_openacc_cuda,
check_effective_target_openacc_cublas,
check_effective_target_openacc_cudart): New.
* testsuite/libgomp.oacc-fortran/host_data-4.f90: Require effective
target openacc_cublas.
* testsuite/libgomp.oacc-fortran/host_data-2.f90: Likewise.
* testsuite/libgomp.oacc-fortran/host_data-3.f: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-91.c: Require effective
target openacc_cuda.
* testsuite/libgomp.oacc-c-c++-common/lib-70.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-90.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-75.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-69.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-74.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-81.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-72.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-85.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/pr87835.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-82.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-73.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-83.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-78.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-76.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-84.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/lib-79.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/host_data-1.c: Require effective
targets openacc_cublas and openacc_cudart.
* testsuite/libgomp.oacc-c-c++-common/context-1.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/context-2.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/context-3.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/context-4.c: Likewise.
* testsuite/libgomp.oacc-c-c++-common/acc_get_property-nvptx.c:
Require effective target openacc_cudart.
* testsuite/libgomp.oacc-c-c++-common/asyncwait-1.c: Add -DUSE_CUDA_H
for effective target openacc_cuda and add && defined USE_CUDA_H to
preprocessor conditionals. Guard -lcuda also on openacc_cuda
effective target.
OpenMP Nesting of Regions restrictions say:
- If a target update, target data, target enter data, or target exit data
construct is encountered during execution of a target region, the behavior is unspecified.
- If a target construct is encountered during execution of a target region and a device
clause in which the ancestor device-modifier appears is not present on the construct, the
behavior is unspecified.
That wording is about the dynamic (runtime) behavior, not about lexical nesting,
so while it is UB if omp target * is encountered in the target region, we need to make
it compile and link (for lexical nesting of target * inside of target we actually
emit a warning).
To make this work, I had to do multiple changes.
One was to mark .omp_data_{sizes,kinds}.* variables when static as "omp declare target".
Another one was to add stub GOMP_target* entrypoints to nvptx and gcn libgomp.a.
The entrypoint functions shouldn't be called or passed in the offload regions,
otherwise
libgomp: cuLaunchKernel error: too many resources requested for launch
was reported; fixed by changing those arguments of calls to GOMP_target_ext
to NULL.
And we didn't mark the entrypoints "omp target entrypoint" when the caller
has been "omp declare target".
2021-05-26 Jakub Jelinek <jakub@redhat.com>
PR libgomp/100573
gcc/
* omp-low.c: Include omp-offload.h.
(create_omp_child_function): If current_function_decl has
"omp declare target" attribute and is_gimple_omp_offloaded,
remove that attribute from the copy of attribute list and
add "omp target entrypoint" attribute instead.
(lower_omp_target): Mark .omp_data_sizes.* and .omp_data_kinds.*
variables for offloading if in omp_maybe_offloaded_ctx.
* omp-offload.c (pass_omp_target_link::execute): Nullify second
argument to GOMP_target_data_ext in offloaded code.
libgomp/
* config/nvptx/target.c (GOMP_target_ext, GOMP_target_data_ext,
GOMP_target_end_data, GOMP_target_update_ext,
GOMP_target_enter_exit_data): New dummy entrypoints.
* config/gcn/target.c (GOMP_target_ext, GOMP_target_data_ext,
GOMP_target_end_data, GOMP_target_update_ext,
GOMP_target_enter_exit_data): Likewise.
* testsuite/libgomp.c-c++-common/for-3.c (DO_PRAGMA, OMPTEAMS,
OMPFROM, OMPTO): Define.
(main): Remove #pragma omp target teams around all the tests.
* testsuite/libgomp.c-c++-common/target-41.c: New test.
* testsuite/libgomp.c-c++-common/target-42.c: New test.
When a directive isn't combined with worksharing-loop, it takes much
simpler clause splitting path for reduction, and that one was missing
handling of teams when combined with simd.
2021-05-25 Jakub Jelinek <jakub@redhat.com>
PR middle-end/99928
gcc/c-family/
* c-omp.c (c_omp_split_clauses): Copy reduction to teams when teams is
combined with simd and not with taskloop or for.
gcc/testsuite/
* c-c++-common/gomp/pr99928-8.c: Remove xfails from omp teams r21 and
r28 checks.
* c-c++-common/gomp/pr99928-9.c: Likewise.
* c-c++-common/gomp/pr99928-10.c: Likewise.
libgomp/
* testsuite/libgomp.c-c++-common/reduction-17.c: New test.
This patch implements a method to track the "private-ness" of
OpenACC variables declared in offload regions in gang-partitioned,
worker-partitioned or vector-partitioned modes. Variables declared
implicitly in scoped blocks and those declared "private" on enclosing
directives (e.g. "acc parallel") are both handled. Variables that are
e.g. gang-private can then be adjusted so they reside in GPU shared
memory.
The reason for doing this is twofold: correct implementation of OpenACC
semantics, and optimisation, since shared memory might be faster than
the main memory on a GPU. Handling of private variables is intimately
tied to the execution model for gangs/workers/vectors implemented by
a particular target: for current targets, we use (or on mainline, will
soon use) a broadcasting/neutering scheme.
That is sufficient for code that e.g. sets a variable in worker-single
mode and expects to use the value in worker-partitioned mode. The
difficulty (semantics-wise) comes when the user wants to do something like
an atomic operation in worker-partitioned mode and expects a worker-single
(gang private) variable to be shared across each partitioned worker.
Forcing use of shared memory for such variables makes that work properly.
In terms of implementation, the parallelism level of a given loop is
not fixed until the oaccdevlow pass in the offload compiler, so the
patch delays fixing the parallelism level of variables declared on or
within such loops until the same point. This is done by adding a new
internal UNIQUE function (OACC_PRIVATE) that lists (the address of) each
private variable as an argument, and other arguments set so as to be able
to determine the correct parallelism level to use for the listed
variables. This new internal function fits into the existing scheme for
demarcating OpenACC loops, as described in comments in the patch.
Two new target hooks are introduced: TARGET_GOACC_ADJUST_PRIVATE_DECL and
TARGET_GOACC_EXPAND_VAR_DECL. The first can tweak a variable declaration
at oaccdevlow time, and the second at expand time. The first or both
of these target hooks can be used by a given offload target, depending
on its strategy for implementing private variables.
This patch updates the TARGET_GOACC_ADJUST_PRIVATE_DECL target hook in
the AMD GCN backend to the current name and prototype. (An earlier
version of the hook was already present, but dormant.)
gcc/
PR middle-end/90115
* doc/tm.texi.in (TARGET_GOACC_EXPAND_VAR_DECL)
(TARGET_GOACC_ADJUST_PRIVATE_DECL): Add documentation hooks.
* doc/tm.texi: Regenerate.
* expr.c (expand_expr_real_1): Expand decls using the
expand_var_decl OpenACC hook if defined.
* internal-fn.c (expand_UNIQUE): Handle IFN_UNIQUE_OACC_PRIVATE.
* internal-fn.h (IFN_UNIQUE_CODES): Add OACC_PRIVATE.
* omp-low.c (omp_context): Add oacc_privatization_candidates
field.
(lower_oacc_reductions): Add PRIVATE_MARKER parameter. Insert
before fork.
(lower_oacc_head_tail): Add PRIVATE_MARKER parameter. Modify
private marker's gimple call arguments, and pass it to
lower_oacc_reductions.
(oacc_privatization_scan_clause_chain)
(oacc_privatization_scan_decl_chain, lower_oacc_private_marker):
New functions.
(lower_omp_for, lower_omp_target, lower_omp_1): Use these.
* omp-offload.c (convert.h): Include.
(oacc_loop_xform_head_tail): Treat private-variable markers like
fork/join when transforming head/tail sequences.
(struct var_decl_rewrite_info): Add struct.
(oacc_rewrite_var_decl, is_sync_builtin_call): New functions.
(execute_oacc_device_lower): Support rewriting gang-private
variables using target hook, and fix up addr_expr and var_decl
nodes afterwards.
* target.def (adjust_private_decl, expand_var_decl): New hooks.
* config/gcn/gcn-protos.h (gcn_goacc_adjust_gangprivate_decl):
Rename to...
(gcn_goacc_adjust_private_decl): ...this.
* config/gcn/gcn-tree.c (gcn_goacc_adjust_gangprivate_decl):
Rename to...
(gcn_goacc_adjust_private_decl): ...this. Add LEVEL parameter.
* config/gcn/gcn.c (TARGET_GOACC_ADJUST_GANGPRIVATE_DECL): Rename
definition using gcn_goacc_adjust_gangprivate_decl...
(TARGET_GOACC_ADJUST_PRIVATE_DECL): ...to this, using
gcn_goacc_adjust_private_decl.
* config/nvptx/nvptx.c (tree-pretty-print.h): Include.
(gang_private_shared_size): New global variable.
(gang_private_shared_align): Likewise.
(gang_private_shared_sym): Likewise.
(gang_private_shared_hmap): Likewise.
(nvptx_option_override): Initialize these.
(nvptx_file_end): Output gang_private_shared_sym.
(nvptx_goacc_adjust_private_decl, nvptx_goacc_expand_var_decl):
New functions.
(nvptx_set_current_function): Clear gang_private_shared_hmap.
(TARGET_GOACC_ADJUST_PRIVATE_DECL): Define hook.
(TARGET_GOACC_EXPAND_VAR_DECL): Likewise.
libgomp/
PR middle-end/90115
* testsuite/libgomp.oacc-c-c++-common/private-atomic-1-gang.c: New
test.
* testsuite/libgomp.oacc-fortran/private-atomic-1-gang.f90:
Likewise.
* testsuite/libgomp.oacc-fortran/private-atomic-1-worker.f90:
Likewise.
Co-Authored-By: Chung-Lin Tang <cltang@codesourcery.com>
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
... to at least document/test/XFAIL nvptx offloading: PR83812 "operation not
supported on global/shared address space".
libgomp/
PR target/83812
* testsuite/libgomp.oacc-c-c++-common/private-atomic-1.c: New.
See local 'offload_targets' variable in
'libgomp/testsuite/lib/libgomp.exp:libgomp_check_effective_target_offload_target'
vs. global 'libgomp/testsuite/libgomp-test-support.exp.in:offload_targets'
variable.
libgomp/
* testsuite/lib/libgomp.exp
(check_effective_target_offload_target_nvptx): Don't shadow global
'offload_targets' variable.
Fix-up for recent commit 33b647956c
"OpenMP: Fix SIMT for complex/float reduction with && and ||"; see
commit d42088e453 "Avoid -latomic for amdgcn
offloading".
libgomp/
* testsuite/libgomp.c-c++-common/reduction-5.c: Restrict
'-latomic' to nvptx offloading compilation.
* testsuite/libgomp.c-c++-common/reduction-6.c: Likewise.
The team barrier should be notified of any new tasks that become runnable
as the result of a completing task, otherwise the barrier threads might
not resume processing available tasks, resulting in a hang.
2021-05-17 Kwok Cheung Yeung <kcy@codesourcery.com>
libgomp/
* task.c (omp_fulfill_event): Call gomp_team_barrier_set_task_pending
if new tasks generated.
* testsuite/libgomp.c-c++-common/task-detach-13.c: New.
gcc/ChangeLog:
* omp-low.c (finish_taskreg_scan): Use the proper detach decl.
libgomp/ChangeLog:
* testsuite/libgomp.c-c++-common/task-detach-12.c: New test.
* testsuite/libgomp.fortran/task-detach-12.f90: New test.
When a taskloop doesn't have any iterations, GOMP_taskloop* takes an early
return, doesn't create any tasks and more importantly, doesn't create
a taskgroup and doesn't register task reductions. But, the code emitted
in the callers assumes task reductions have been registered and performs
the reduction handling and task reduction unregistration. The pointer
to the task reduction private variables is reused, on input it is the alignment
and only on output it is the pointer, so in the case taskloop with no iterations
the caller attempts to dereference the alignment value as if it was a pointer
and crashes. We could in the early returns register the task reductions
only to have them looped over and unregistered in the caller, but I think
it is better to tell the caller there is nothing to task reduce and bypass
all that.
2021-05-11 Jakub Jelinek <jakub@redhat.com>
PR middle-end/100471
* omp-low.c (lower_omp_task_reductions): For OMP_TASKLOOP, if data
is 0, bypass the reduction loop including
GOMP_taskgroup_reduction_unregister call.
* taskloop.c (GOMP_taskloop): If GOMP_TASK_FLAG_REDUCTION and not
GOMP_TASK_FLAG_NOGROUP, when doing early return clear the task
reduction pointer.
* testsuite/libgomp.c/task-reduction-4.c: New test.
2021-05-07 Tobias Burnus <tobias@codesourcery.com>
Tom de Vries <tdevries@suse.de>
gcc/ChangeLog:
* omp-low.c (lower_rec_simd_input_clauses): Set max_vf = 1 if
a truth_value_p reduction variable is nonintegral.
libgomp/ChangeLog:
* testsuite/libgomp.c-c++-common/reduction-5.c: New test, testing
complex/floating-point || + && reduction with 'omp target'.
* testsuite/libgomp.c-c++-common/reduction-6.c: Likewise.
C/C++ permit logical AND and logical OR also with floating-point or complex
arguments by doing an unequal zero comparison; the result is an 'int' with
value one or zero. Hence, those are also permitted as reduction variable,
even though it is not the most sensible thing to do.
gcc/c/ChangeLog:
* c-typeck.c (c_finish_omp_clauses): Accept float + complex
for || and && reductions.
gcc/cp/ChangeLog:
* semantics.c (finish_omp_reduction_clause): Accept float + complex
for || and && reductions.
gcc/ChangeLog:
* omp-low.c (lower_rec_input_clauses, lower_reduction_clauses): Handle
&& and || with floating-point and complex arguments.
gcc/testsuite/ChangeLog:
* gcc.dg/gomp/clause-1.c: Use 'reduction(&:..)' instead of '...(&&:..)'.
libgomp/ChangeLog:
* testsuite/libgomp.c-c++-common/reduction-1.c: New test.
* testsuite/libgomp.c-c++-common/reduction-2.c: New test.
* testsuite/libgomp.c-c++-common/reduction-3.c: New test.
The test-case included in this patch contains this target region:
...
for (int i0 = 0 ; i0 < N0 ; i0++ )
counter_N0.i += 1;
...
When running with nvptx accelerator, the counter variable is expected to
be N0 after the region, but instead is N0 / 32. The problem is that rather
than getting the result for all warp lanes, we get it for just one lane.
This is caused by the implementation of SIMT being incomplete. It handles
regular reductions, but appearantly not user-defined reductions.
For now, handle this by disabling SIMT in this case, specifically by setting
sctx->max_vf to 1.
Tested libgomp on x86_64-linux with nvptx accelerator.
gcc/ChangeLog:
2021-05-03 Tom de Vries <tdevries@suse.de>
PR target/100321
* omp-low.c (lower_rec_input_clauses): Disable SIMT for user-defined
reduction.
libgomp/ChangeLog:
2021-05-03 Tom de Vries <tdevries@suse.de>
PR target/100321
* testsuite/libgomp.c/target-44.c: New test.
PR84878 fix adds an assertion which can fail, e.g. when stack pointer
is adjusted inside the loop. We have to prevent it and search earlier
for any 'strange' instruction. The solution is to skip the whole loop
if using 'note_stores' we found that one of hard registers is in
'df->regular_block_artificial_uses' set.
Also patch properly prohibit not single-set instruction in loop body.
gcc/ChangeLog:
PR rtl-optimization/100225
PR rtl-optimization/84878
* modulo-sched.c (sms_schedule): Use note_stores to skip loops
where we have an instruction which touches (writes) any hard
register from df->regular_block_artificial_uses set.
Allow not-single-set instruction only right before basic block
tail.
gcc/testsuite/ChangeLog:
PR rtl-optimization/100225
PR rtl-optimization/84878
* gcc.dg/pr100225.c: New test.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-c-c++-common/atomic_capture-3.c: New test.
Consider the test-case libgomp.c/pr81778.c added in this commit, with
this core loop (note: CANARY_SIZE set to 0 for simplicity):
...
int s = 1;
#pragma omp target simd
for (int i = N - 1; i > -1; i -= s)
a[i] = 1;
...
which, given that N is 32, sets a[0..31] to 1.
After omp-expand, this looks like:
...
<bb 5> :
simduid.7 = .GOMP_SIMT_ENTER (simduid.7);
.omp_simt.8 = .GOMP_SIMT_ENTER_ALLOC (simduid.7);
D.3193 = -s;
s.9 = s;
D.3204 = .GOMP_SIMT_LANE ();
D.3205 = -s.9;
D.3206 = (int) D.3204;
D.3207 = D.3205 * D.3206;
i = D.3207 + 31;
D.3209 = 0;
D.3210 = -s.9;
D.3211 = D.3210 - i;
D.3210 = -s.9;
D.3212 = D.3211 / D.3210;
D.3213 = (unsigned int) D.3212;
D.3213 = i >= 0 ? D.3213 : 0;
<bb 19> :
if (D.3209 < D.3213)
goto <bb 6>; [87.50%]
else
goto <bb 7>; [12.50%]
<bb 6> :
a[i] = 1;
D.3215 = -s.9;
D.3219 = .GOMP_SIMT_VF ();
D.3216 = (int) D.3219;
D.3220 = D.3215 * D.3216;
i = D.3220 + i;
D.3209 = D.3209 + 1;
goto <bb 19>; [100.00%]
...
On nvptx, the first time bb6 is executed, i is in the 0..31 range (depending
on the lane that is executing) at bb entry.
So we have the following sequence:
- a[0..31] is set to 1
- i is updated to -32..-1
- D.3209 is updated to 1 (being 0 initially)
- bb19 is executed, and if condition (D.3209 < D.3213) == (1 < 32) evaluates
to true
- bb6 is once more executed, which should not happen because all the elements
that needed to be handled were already handled.
- consequently, elements that should not be written are written
- with CANARY_SIZE == 0, we may run into a libgomp error:
...
libgomp: cuCtxSynchronize error: an illegal memory access was encountered
...
and with CANARY_SIZE unmodified, we run into:
...
Expected 0, got 1 at base[-961]
Aborted (core dumped)
...
The cause of this is as follows:
- because the step s is a variable rather than a constant, an alternative
IV (D.3209 in our example) is generated in expand_omp_simd, and the
loop condition is tested in terms of the alternative IV rather than
the original IV (i in our example).
- the SIMT code in expand_omp_simd works by modifying step and initial value.
- The initial value fd->loop.n1 is loaded into a variable n1, which is
modified by the SIMT code and then used there-after.
- The step fd->loop.step is loaded into a variable step, which is modified
by the SIMT code, but afterwards there are uses of both step and
fd->loop.step.
- There are uses of fd->loop.step in the alternative IV handling code,
which should use step instead.
Fix this by introducing an additional variable orig_step, which is not
modified by the SIMT code and replacing all remaining uses of fd->loop.step
by either step or orig_step.
Build on x86_64-linux with nvptx accelerator, tested libgomp.
This fixes for-5.c and for-6.c FAILs I'm currently seeing on a quadro m1200
with driver 450.66.
gcc/ChangeLog:
2020-10-02 Tom de Vries <tdevries@suse.de>
* omp-expand.c (expand_omp_simd): Add step_orig, and replace uses of
fd->loop.step by either step or orig_step.
libgomp/ChangeLog:
2020-10-02 Tom de Vries <tdevries@suse.de>
* testsuite/libgomp.c/pr81778.c: New test.
When running the test-case included in this patch using an
nvptx accelerator, it fails in execution.
The problem is that the expansion of GOMP_SIMT_XCHG_BFLY is optimized away
during pass_jump as "trivially dead insns".
This is caused by this code in expand_GOMP_SIMT_XCHG_BFLY:
...
class expand_operand ops[3];
create_output_operand (&ops[0], target, mode);
...
expand_insn (targetm.code_for_omp_simt_xchg_bfly, 3, ops);
...
which doesn't guarantee that target is assigned to by the expanded insn.
F.i., if target is:
...
(gdb) call debug_rtx ( target )
(subreg/s/u:QI (reg:SI 40 [ _61 ]) 0)
...
then after expand_insn, we have:
...
(gdb) call debug_rtx ( ops[0].value )
(reg:QI 57)
...
See commit 3af3bec2e4 "internal-fn: Avoid dropping the lhs of some
calls [PR94941]" for a similar problem.
Fix this in the same way, by adding:
...
if (!rtx_equal_p (target, ops[0].value))
emit_move_insn (target, ops[0].value);
...
where applicable in the expand_GOMP_SIMT_* functions.
Tested libgomp on x86_64 with nvptx accelerator.
gcc/ChangeLog:
2021-04-28 Tom de Vries <tdevries@suse.de>
PR target/100232
* internal-fn.c (expand_GOMP_SIMT_ENTER_ALLOC)
(expand_GOMP_SIMT_LAST_LANE, expand_GOMP_SIMT_ORDERED_PRED)
(expand_GOMP_SIMT_VOTE_ANY, expand_GOMP_SIMT_XCHG_BFLY)
(expand_GOMP_SIMT_XCHG_IDX): Ensure target is assigned to.
If configured with --enable-offload-defaulted, configured but not installed
offload compilers and libgomp plugins are silently ignored. Useful for
distribution compilers where those are in separate optional packages.
2021-04-28 Jakub Jelinek <jakub@redhat.com>
Tobias Burnus <tobias@codesourcery.com>
ChangeLog:
* configure.ac (--enable-offload-defaulted): New.
* configure: Regenerate.
gcc/ChangeLog:
* configure.ac (OFFLOAD_DEFAULTED): AC_DEFINE if offload-defaulted.
* gcc.c (process_command): New variable.
(driver::maybe_putenv_OFFLOAD_TARGETS): If OFFLOAD_DEFAULTED,
set it if -foffload is defaulted.
* lto-wrapper.c (OFFLOAD_TARGET_DEFAULT_ENV): Define.
(compile_offload_image): If OFFLOAD_DEFAULTED and
OFFLOAD_TARGET_DEFAULT is in the environment, don't fail
if corresponding mkoffload can't be found.
(compile_images_for_offload_targets): Likewise. Free and clear
offload_names if no valid offload is found.
* config.in: Regenerate.
* configure: Regenerate.
libgomp/ChangeLog:
* configure.ac (OFFLOAD_DEFAULTED): AC_DEFINE if offload-defaulted.
* target.c (gomp_load_plugin_for_device): If set and if a plugin
can't be dlopened, silently assume it has no devices.
* Makefile.in: Regenerate.
* config.h.in: Regenerate.
* configure: Regenerate.
It turned out that a compiler built without offloading support
and one with can produce slightly different diagnostic.
Offloading support implies ENABLE_OFFLOAD which implies that
g->have_offload is set when offloading is actually needed.
In cgraphunit.c, the latter causes flag_generate_offload = 1,
which in turn affects tree.c's free_lang_data.
The result is that the front-end specific diagnostic gets reset
('tree_diagnostics_defaults (global_dc)'), which affects in this
case 'Warning' vs. 'warning' via the Fortran frontend.
Result: 'Warning:' vs. 'warning:'.
Side note: Other FE also override the diagnostic, leading to
similar differences, e.g. the C++ FE outputs mangled function
names differently, cf. patch thread.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-fortran/par-reduction-2-1.f:
Use [Ww]arning in dg-bogus as FE diagnostic and default
diagnostic differ and the result depends on ENABLE_OFFLOAD.
* testsuite/libgomp.oacc-fortran/par-reduction-2-2.f: Likewise.
* testsuite/libgomp.oacc-fortran/parallel-dims.f90: Likewise.
* testsuite/libgomp.oacc-fortran/parallel-reduction.f90: Likewise.
gcc/testsuite/ChangeLog:
* gfortran.dg/goacc/classify-serial.f95:
Use [Ww]arning in dg-bogus as FE diagnostic and default
diagnostic differ and the result depends on ENABLE_OFFLOAD.
* gfortran.dg/goacc/kernels-decompose-2.f95: Likewise.
* gfortran.dg/goacc/routine-module-mod-1.f90: Likewise.
