Up to now the libgomp GCN plugin has been finding the offload variables
by using a symbol lookup, but the AMD runtime requires that the symbols are
global for that to work. This was ensured by mkoffload as a post-procssing
step, but the LLVM 13 assembler no longer accepts this in the case where the
variable was previously declared differently.
This patch switches to locating the symbols directly from the
offload_var_table, which means that only one symbol needs to be forced
global.
This changes breaks the libgomp image compatibility so GOMP_VERSION_GCN has
also been bumped.
gcc/ChangeLog:
* config/gcn/mkoffload.c (process_asm): Process the variable table
completely differently.
(process_obj): Encode the varaible data differently.
include/ChangeLog:
* gomp-constants.h (GOMP_VERSION_GCN): Bump.
libgomp/ChangeLog:
* plugin/plugin-gcn.c (struct gcn_image_desc): Remove global_variables.
(GOMP_OFFLOAD_load_image): Locate the offload variables via the
table, not individual symbols.
Some testcases for libgomp.c++ only works for non-shared address space offloading,
because it exercises the zero-length array section behavior for offloaded
address space, testing for NULL/non-NULL cases.
libgomp/ChangeLog:
* testsuite/libgomp.c++/target-lambda-1.C: Only run under
"target offload_device_nonshared_as"
* testsuite/libgomp.c++/target-this-3.C: Likewise.
* testsuite/libgomp.c++/target-this-4.C: Likewise.
This patch implements three pieces of functionality:
(1) Adjust array section mapping to have standards conforming behavior,
mapping array sections should *NOT* also map the base-pointer:
struct S { int *ptr; ... };
struct S s;
Instead of generating this during gimplify:
map(to:*_1 [len: 400]) map(attach:s.ptr [bias: 0])
Now, adjust to:
(i.e. do not map the base-pointer together. The attach operation is still
generated, and if s.ptr is already mapped prior, attachment will happen)
The correct way of achieving the base-pointer-also-mapped behavior would be to
use:
(A small Fortran front-end patch to trans-openmp.c:gfc_trans_omp_array_section
is also included, which removes generation of a GOMP_MAP_ALWAYS_POINTER for
array types, which appears incorrect and causes a regression in
libgomp.fortranlibgomp.fortran/struct-elem-map-1.f90)
(2) Related to the first item above, are fixes in libgomp/target.c to not
overwrite attached pointers when handling device<->host copies, mainly for the
"always" case.
(3) The third is a set of changes to the C/C++ front-ends to extend the allowed
component access syntax in map clauses. These changes are enabled for both
OpenACC and OpenMP.
gcc/c/ChangeLog:
* c-parser.c (struct omp_dim): New struct type for use inside
c_parser_omp_variable_list.
(c_parser_omp_variable_list): Allow multiple levels of array and
component accesses in array section base-pointer expression.
(c_parser_omp_clause_to): Set 'allow_deref' to true in call to
c_parser_omp_var_list_parens.
(c_parser_omp_clause_from): Likewise.
* c-typeck.c (handle_omp_array_sections_1): Extend allowed range
of base-pointer expressions involving INDIRECT/MEM/ARRAY_REF and
POINTER_PLUS_EXPR.
(c_finish_omp_clauses): Extend allowed ranged of expressions
involving INDIRECT/MEM/ARRAY_REF and POINTER_PLUS_EXPR.
gcc/cp/ChangeLog:
* parser.c (struct omp_dim): New struct type for use inside
cp_parser_omp_var_list_no_open.
(cp_parser_omp_var_list_no_open): Allow multiple levels of array and
component accesses in array section base-pointer expression.
(cp_parser_omp_all_clauses): Set 'allow_deref' to true in call to
cp_parser_omp_var_list for to/from clauses.
* semantics.c (handle_omp_array_sections_1): Extend allowed range
of base-pointer expressions involving INDIRECT/MEM/ARRAY_REF and
POINTER_PLUS_EXPR.
(handle_omp_array_sections): Adjust pointer map generation of
references.
(finish_omp_clauses): Extend allowed ranged of expressions
involving INDIRECT/MEM/ARRAY_REF and POINTER_PLUS_EXPR.
gcc/fortran/ChangeLog:
* trans-openmp.c (gfc_trans_omp_array_section): Do not generate
GOMP_MAP_ALWAYS_POINTER map for main array maps of ARRAY_TYPE type.
gcc/ChangeLog:
* gimplify.c (extract_base_bit_offset): Add 'tree *offsetp' parameter,
accomodate case where 'offset' return of get_inner_reference is
non-NULL.
(is_or_contains_p): Further robustify conditions.
(omp_target_reorder_clauses): In alloc/to/from sorting phase, also
move following GOMP_MAP_ALWAYS_POINTER maps along. Add new sorting
phase where we make sure pointers with an attach/detach map are ordered
correctly.
(gimplify_scan_omp_clauses): Add modifications to avoid creating
GOMP_MAP_STRUCT and associated alloc map for attach/detach maps.
gcc/testsuite/ChangeLog:
* c-c++-common/goacc/deep-copy-arrayofstruct.c: Adjust testcase.
* c-c++-common/gomp/target-enter-data-1.c: New testcase.
* c-c++-common/gomp/target-implicit-map-2.c: New testcase.
libgomp/ChangeLog:
* target.c (gomp_map_vars_existing): Make sure attached pointer is
not overwritten during cross-host/device copying.
(gomp_update): Likewise.
(gomp_exit_data): Likewise.
* testsuite/libgomp.c++/target-11.C: Adjust testcase.
* testsuite/libgomp.c++/target-12.C: Likewise.
* testsuite/libgomp.c++/target-15.C: Likewise.
* testsuite/libgomp.c++/target-16.C: Likewise.
* testsuite/libgomp.c++/target-17.C: Likewise.
* testsuite/libgomp.c++/target-21.C: Likewise.
* testsuite/libgomp.c++/target-23.C: Likewise.
* testsuite/libgomp.c/target-23.c: Likewise.
* testsuite/libgomp.c/target-29.c: Likewise.
* testsuite/libgomp.c-c++-common/target-implicit-map-2.c: New testcase.
This patch implements several C++ specific mapping capabilities introduced for
OpenMP 5.0, including implicit mapping of this[:1] for non-static member
functions, zero-length array section mapping of pointer-typed members,
lambda captured variable access in target regions, and use of lambda objects
inside target regions.
Several adjustments to the C/C++ front-ends to allow more member-access syntax
as valid is also included.
PR middle-end/92120
gcc/cp/ChangeLog:
* cp-tree.h (finish_omp_target): New declaration.
(finish_omp_target_clauses): Likewise.
* parser.c (cp_parser_omp_clause_map): Adjust call to
cp_parser_omp_var_list_no_open to set 'allow_deref' argument to true.
(cp_parser_omp_target): Factor out code, adjust into calls to new
function finish_omp_target.
* pt.c (tsubst_expr): Add call to finish_omp_target_clauses for
OMP_TARGET case.
* semantics.c (handle_omp_array_sections_1): Add handling to create
'this->member' from 'member' FIELD_DECL. Remove case of rejecting
'this' when not in declare simd.
(handle_omp_array_sections): Likewise.
(finish_omp_clauses): Likewise. Adjust to allow 'this[]' in OpenMP
map clauses. Handle 'A->member' case in map clauses. Remove case of
rejecting 'this' when not in declare simd.
(struct omp_target_walk_data): New struct for walking over
target-directive tree body.
(finish_omp_target_clauses_r): New function for tree walk.
(finish_omp_target_clauses): New function.
(finish_omp_target): New function.
gcc/c/ChangeLog:
* c-parser.c (c_parser_omp_clause_map): Set 'allow_deref' argument in
call to c_parser_omp_variable_list to 'true'.
* c-typeck.c (handle_omp_array_sections_1): Add strip of MEM_REF in
array base handling.
(c_finish_omp_clauses): Handle 'A->member' case in map clauses.
gcc/ChangeLog:
* gimplify.c ("tree-hash-traits.h"): Add include.
(gimplify_scan_omp_clauses): Change struct_map_to_clause to type
hash_map<tree_operand, tree> *. Adjust struct map handling to handle
cases of *A and A->B expressions. Under !DECL_P case of
GOMP_CLAUSE_MAP handling, add STRIP_NOPS for indir_p case, add to
struct_deref_set for map(*ptr_to_struct) cases. Add MEM_REF case when
handling component_ref_p case. Add unshare_expr and gimplification
when created GOMP_MAP_STRUCT is not a DECL. Add code to add
firstprivate pointer for *pointer-to-struct case.
(gimplify_adjust_omp_clauses): Move GOMP_MAP_STRUCT removal code for
exit data directives code to earlier position.
* omp-low.c (lower_omp_target):
Handle GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION, and
GOMP_MAP_POINTER_TO_ZERO_LENGTH_ARRAY_SECTION map kinds.
* tree-pretty-print.c (dump_omp_clause): Likewise.
gcc/testsuite/ChangeLog:
* gcc.dg/gomp/target-3.c: New testcase.
* g++.dg/gomp/target-3.C: New testcase.
* g++.dg/gomp/target-lambda-1.C: New testcase.
* g++.dg/gomp/target-lambda-2.C: New testcase.
* g++.dg/gomp/target-this-1.C: New testcase.
* g++.dg/gomp/target-this-2.C: New testcase.
* g++.dg/gomp/target-this-3.C: New testcase.
* g++.dg/gomp/target-this-4.C: New testcase.
* g++.dg/gomp/target-this-5.C: New testcase.
* g++.dg/gomp/this-2.C: Adjust testcase.
include/ChangeLog:
* gomp-constants.h (enum gomp_map_kind):
Add GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION, and
GOMP_MAP_POINTER_TO_ZERO_LENGTH_ARRAY_SECTION map kinds.
(GOMP_MAP_POINTER_P):
Include GOMP_MAP_POINTER_TO_ZERO_LENGTH_ARRAY_SECTION.
libgomp/ChangeLog:
* libgomp.h (gomp_attach_pointer): Add bool parameter.
* oacc-mem.c (acc_attach_async): Update call to gomp_attach_pointer.
(goacc_enter_data_internal): Likewise.
* target.c (gomp_map_vars_existing): Update assert condition to
include GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION.
(gomp_map_pointer): Add 'bool allow_zero_length_array_sections'
parameter, add support for mapping a pointer with NULL target.
(gomp_attach_pointer): Add 'bool allow_zero_length_array_sections'
parameter, add support for attaching a pointer with NULL target.
(gomp_map_vars_internal): Update calls to gomp_map_pointer and
gomp_attach_pointer, add handling for
GOMP_MAP_ATTACH_ZERO_LENGTH_ARRAY_SECTION, and
GOMP_MAP_POINTER_TO_ZERO_LENGTH_ARRAY_SECTION cases.
* testsuite/libgomp.c++/target-23.C: New testcase.
* testsuite/libgomp.c++/target-lambda-1.C: New testcase.
* testsuite/libgomp.c++/target-lambda-2.C: New testcase.
* testsuite/libgomp.c++/target-this-1.C: New testcase.
* testsuite/libgomp.c++/target-this-2.C: New testcase.
* testsuite/libgomp.c++/target-this-3.C: New testcase.
* testsuite/libgomp.c++/target-this-4.C: New testcase.
* testsuite/libgomp.c++/target-this-5.C: New testcase.
Implements moste of OpenMP 5.1 atomic extensions,
except that 'compare' is parsed but rejected during
resolution. (As the trans-openmp.c handling is missing.)
gcc/fortran/ChangeLog:
* dump-parse-tree.c (show_omp_clauses): Handle
weak/compare/fail clause.
* gfortran.h (gfc_omp_clauses): Add weak, compare, fail.
* openmp.c (enum omp_mask1, gfc_match_omp_clauses,
OMP_ATOMIC_CLAUSES): Update for new clauses.
(gfc_match_omp_atomic): Update for 5.1 atomic changes.
(is_conversion): Support widening in one go.
(is_scalar_intrinsic_expr): New.
(resolve_omp_atomic): Update for 5.1 atomic changes.
* parse.c (parse_omp_oacc_atomic): Update for compare.
* resolve.c (gfc_resolve_blocks): Update asserts.
* trans-openmp.c (gfc_trans_omp_atomic): Handle new clauses.
gcc/testsuite/ChangeLog:
* gfortran.dg/gomp/atomic-2.f90: Move now supported code to ...
* gfortran.dg/gomp/atomic.f90: here.
* gfortran.dg/gomp/atomic-10.f90: New test.
* gfortran.dg/gomp/atomic-12.f90: New test.
* gfortran.dg/gomp/atomic-15.f90: New test.
* gfortran.dg/gomp/atomic-16.f90: New test.
* gfortran.dg/gomp/atomic-17.f90: New test.
* gfortran.dg/gomp/atomic-18.f90: New test.
* gfortran.dg/gomp/atomic-19.f90: New test.
* gfortran.dg/gomp/atomic-20.f90: New test.
* gfortran.dg/gomp/atomic-22.f90: New test.
* gfortran.dg/gomp/atomic-24.f90: New test.
* gfortran.dg/gomp/atomic-25.f90: New test.
* gfortran.dg/gomp/atomic-26.f90: New test.
libgomp/ChangeLog
* libgomp.texi (OpenMP 5.1): Update status.
Fix issue with the Fortran front-end when mapping arrays: when creating the
data MEM_REF for the map clause, there was a convention of casting the
referencing pointer to 'c_char *' by
fold_convert (build_pointer_type (char_type_node), ptr).
This causes the alignment passed to the libgomp runtime for array data
hardwared to '1', and causes alignment errors on the offload target.
This patch fixes this by removing the char_type_node pointer converts, and
adding gcc_asserts to ensure POINTER_TYPE_P (TREE_TYPE (ptr)).
PR fortran/90030
gcc/fortran/ChangeLog:
* trans-openmp.c (gfc_omp_finish_clause): Remove fold_convert to pointer
to char_type_node, add gcc_assert of POINTER_TYPE_P.
(gfc_trans_omp_array_section): Likewise.
(gfc_trans_omp_clauses): Likewise.
gcc/testsuite/ChangeLog:
* gfortran.dg/goacc/finalize-1.f: Adjust scan test.
* gfortran.dg/gomp/affinity-clause-1.f90: Likewise.
* gfortran.dg/gomp/affinity-clause-5.f90: Likewise.
* gfortran.dg/gomp/defaultmap-4.f90: Likewise.
* gfortran.dg/gomp/defaultmap-5.f90: Likewise.
* gfortran.dg/gomp/defaultmap-6.f90: Likewise.
* gfortran.dg/gomp/map-3.f90: Likewise.
* gfortran.dg/gomp/pr78260-2.f90: Likewise.
* gfortran.dg/gomp/pr78260-3.f90: Likewise.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-fortran/pr90030.f90: New test.
* testsuite/libgomp.fortran/pr90030.f90: New test.
This patch promotes all OpenACC gang reductions on orphan loops as
errors. Accord to the spec, orphan loops are those which are not
lexically nested inside an OpenACC parallel or kernels regions. I.e.,
acc loops inside acc routines.
At first I thought this could be a warning because the gang reduction
finalizer uses an atomic update. However, because there is no
synchronization between gangs, there is way to guarantee that reduction
will have completed once a single gang entity returns from the acc
routine call.
gcc/c/
* c-typeck.c (c_finish_omp_clauses): Emit an error on orphan
OpenACC gang reductions.
gcc/cp/
* semantics.c (finish_omp_clauses): Emit an error on orphan
OpenACC gang reductions.
gcc/fortran/
* openmp.c (oacc_is_parallel, oacc_is_kernels): New 'static'
functions.
(resolve_oacc_loop_blocks): Emit an error on orphan OpenACC gang
reductions.
gcc/
* omp-general.h (enum oacc_loop_flags): Add OLF_REDUCTION enum.
* omp-low.c (lower_oacc_head_mark): Use it to mark OpenACC
reductions.
* omp-offload.c (oacc_loop_auto_partitions): Don't assign gang
level parallelism to orphan reductions.
gcc/testsuite/
* c-c++-common/goacc/nested-reductions-1-routine.c: Adjust.
* c-c++-common/goacc/nested-reductions-2-routine.c: Likewise.
* gcc.dg/goacc/loop-processing-1.c: Likewise.
* gfortran.dg/goacc/nested-reductions-1-routine.f90: Likewise.
* gfortran.dg/goacc/nested-reductions-2-routine.f90: Likewise.
* c-c++-common/goacc/orphan-reductions-1.c: New test.
* c-c++-common/goacc/orphan-reductions-2.c: New test.
* gfortran.dg/goacc/orphan-reductions-1.f90: New test.
* gfortran.dg/goacc/orphan-reductions-2.f90: New test.
libgomp/
* testsuite/libgomp.oacc-fortran/parallel-dims.f90: Temporarily
skip.
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
This removes unreachable return statements as diagnosed by
the -Wunreachable-code patch. Some cases are more obviously
an improvement than others - in fact some may get you the idea
to replace them with gcc_unreachable () instead, leading to
cases of the 'Remove unreachable gcc_unreachable () at the end
of functions' patch.
2021-11-25 Richard Biener <rguenther@suse.de>
* vec.c (qsort_chk): Do not return the void return value
from the noreturn qsort_chk_error.
* ccmp.c (expand_ccmp_expr_1): Remove unreachable return.
* df-scan.c (df_ref_equal_p): Likewise.
* dwarf2out.c (is_base_type): Likewise.
(add_const_value_attribute): Likewise.
* fixed-value.c (fixed_arithmetic): Likewise.
* gimple-fold.c (gimple_fold_builtin_fputs): Likewise.
* gimple-ssa-strength-reduction.c (stmt_cost): Likewise.
* graphite-isl-ast-to-gimple.c
(gcc_expression_from_isl_expr_op): Likewise.
(gcc_expression_from_isl_expression): Likewise.
* ipa-fnsummary.c (will_be_nonconstant_expr_predicate):
Likewise.
* lto-streamer-in.c (lto_input_mode_table): Likewise.
gcc/c-family/
* c-opts.c (c_common_post_options): Remove unreachable return.
* c-pragma.c (handle_pragma_target): Likewise.
(handle_pragma_optimize): Likewise.
gcc/c/
* c-typeck.c (c_tree_equal): Remove unreachable return.
* c-parser.c (get_matching_symbol): Likewise.
libgomp/
* oacc-plugin.c (GOMP_PLUGIN_acc_default_dim): Remove unreachable
return.
As the testcase shows, we weren't handling kind(host) and kind(nohost) properly
in the ACCEL_COMPILERs, the code written in there is valid for the host
compiler only, where if we are maybe offloaded, we defer resolution after IPA,
otherwise return 0 for kind(nohost) and accept it for kind(host). Note,
omp_maybe_offloaded is false after IPA. If ACCEL_COMPILER is defined, it is
the other way around, but also we know we are after IPA.
2021-11-24 Jakub Jelinek <jakub@redhat.com>
PR middle-end/103384
gcc/
* omp-general.c (omp_context_selector_matches): For ACCEL_COMPILER,
return 0 for kind(host) and continue for kind(nohost).
libgomp/
* testsuite/libgomp.c/declare-variant-2.c: New test.
struct gomp_team has struct gomp_work_share array inside of it.
If that latter structure has 64-byte aligned member in the middle,
the whole struct gomp_team needs to be 64-byte aligned, but we weren't
allocating it using gomp_aligned_alloc.
This patch fixes that, except that on gcn team_malloc is special, so
I've instead decided at least for now to avoid using aligned member
and use the padding instead on gcn.
2021-11-18 Jakub Jelinek <jakub@redhat.com>
PR libgomp/102838
* libgomp.h (GOMP_USE_ALIGNED_WORK_SHARES): Define if
GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC is defined and __AMDGCN__ is not.
(struct gomp_work_share): Use GOMP_USE_ALIGNED_WORK_SHARES instead of
GOMP_HAVE_EFFICIENT_ALIGNED_ALLOC.
* work.c (alloc_work_share, gomp_work_share_start): Likewise.
* team.c (gomp_new_team): If GOMP_USE_ALIGNED_WORK_SHARES, use
gomp_aligned_alloc instead of team_malloc.
C says that aligned_alloc size must be an integral multiple of alignment.
While glibc doesn't care about it, apparently Solaris does.
So, this patch decreases the priority of aligned_alloc among the other
variants because it needs more work and can waste more memory and rounds
up the size to multiple of alignment.
2021-11-18 Jakub Jelinek <jakub@redhat.com>
PR libgomp/102838
* alloc.c (gomp_aligned_alloc): Prefer _aligned_alloc over
memalign over posix_memalign over aligned_alloc over fallback
with malloc instead of aligned_alloc over _aligned_alloc over
posix_memalign over memalign over fallback with malloc. For
aligned_alloc, round up size up to multiple of al.
After the Fortran changes we can mark it as implemented...
2021-11-16 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (OpenMP 5.1): Mark thread_limit clause to target
construct as implemented.
OpenMP 5.1 says that thread_limit clause can also appear on target,
and similarly to teams should affect the thread-limit-var ICV.
On combined target teams, the clause goes to both.
We actually passed thread_limit internally on target already before,
but only used it for gcn/ptx offloading to hint how many threads should be
created and for ptx didn't set thread_limit_var in that case.
Similarly for host fallback.
Also, I found that we weren't copying the args array that contains encoded
thread_limit and num_teams clause for target (etc.) for async target.
2021-11-15 Jakub Jelinek <jakub@redhat.com>
gcc/
* gimplify.c (optimize_target_teams): Only add OMP_CLAUSE_THREAD_LIMIT
to OMP_TARGET_CLAUSES if it isn't there already.
gcc/c-family/
* c-omp.c (c_omp_split_clauses) <case OMP_CLAUSE_THREAD_LIMIT>:
Duplicate to both OMP_TARGET and OMP_TEAMS.
gcc/c/
* c-parser.c (OMP_TARGET_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_THREAD_LIMIT.
gcc/cp/
* parser.c (OMP_TARGET_CLAUSE_MASK): Add
PRAGMA_OMP_CLAUSE_THREAD_LIMIT.
libgomp/
* task.c (gomp_create_target_task): Copy args array as well.
* target.c (gomp_target_fallback): Add args argument.
Set gomp_icv (true)->thread_limit_var if thread_limit is present.
(GOMP_target): Adjust gomp_target_fallback caller.
(GOMP_target_ext): Likewise.
(gomp_target_task_fn): Likewise.
* config/nvptx/team.c (gomp_nvptx_main): Set
gomp_global_icv.thread_limit_var.
* testsuite/libgomp.c-c++-common/thread-limit-1.c: New test.
Here is a PTX implementation of what I was talking about, that for
num_teams_upper 0 or whenever num_teams_lower <= num_blocks, the current
implementation is fine but if the user explicitly asks for more
teams than we can provide in hardware, we need to stop assuming that
omp_get_team_num () is equal to the hw team id, but instead need to use some
team specific memory (it is .shared for PTX), or if none is
provided, array indexed by the hw team id and run some teams serially within
the same hw thread.
2021-11-15 Jakub Jelinek <jakub@redhat.com>
* config/nvptx/team.c (__gomp_team_num): Define as
__attribute__((shared)) var.
(gomp_nvptx_main): Initialize __gomp_team_num to 0.
* config/nvptx/target.c (__gomp_team_num): Declare as
extern __attribute__((shared)) var.
(GOMP_teams4): Use __gomp_team_num as the team number instead of
%ctaid.x. If first, initialize it to %ctaid.x. If num_teams_lower
is bigger than num_blocks, use num_teams_lower teams and arrange for
bumping of __gomp_team_num if !first and returning false once we run
out of teams.
* config/nvptx/teams.c (__gomp_team_num): Declare as
extern __attribute__((shared)) var.
(omp_get_team_num): Return __gomp_team_num value instead of %ctaid.x.
This is https://github.com/OpenMP/spec/issues/3183
There is an agreement that we should return 1 team inside of target,
even if that target is inside of host teams. We were doing that
when offloading and not during host fallback, r12-5151 should fix that
even for host fallback.
2021-11-15 Jakub Jelinek <jakub@redhat.com>
* testsuite/libgomp.c/teams-5.c: New test.
My recent libgomp change apparently broke libgomp build for gcn offloading.
The problem is that gcn, unlike nvptx, doesn't override teams.c source file
and the patch I've committed assumed all the non-LIBGOMP_USE_PTHREADS targets
do not use it. My understanding is that gcn included omp_get_num_teams
and omp_get_team_num definitions in both icv-device.o and teams.o,
with the definitions only in the former working correctly.
This patch brings gcn into sync with how nvptx does it, that teams.c
is overridden, provides a dummy GOMP_teams_reg and omp_get_{num_teams,team_num}
definitions and icv-device.c doesn't provide those.
2021-11-12 Jakub Jelinek <jakub@redhat.com>
PR target/103201
* config/gcn/icv-device.c (omp_get_num_teams, omp_get_team_num): Move
to ...
* config/gcn/teams.c: ... here. New file.
This patch implements relaxing the requirements when a map with the implicit
attribute encounters an overlapping existing map. As the OpenMP 5.0 spec
describes on page 320, lines 18-27 (and 5.1 spec, page 352, lines 13-22):
"If a single contiguous part of the original storage of a list item with an
implicit data-mapping attribute has corresponding storage in the device data
environment prior to a task encountering the construct that is associated with
the map clause, only that part of the original storage will have corresponding
storage in the device data environment as a result of the map clause."
2021-11-12 Chung-Lin Tang <cltang@codesourcery.com>
include/ChangeLog:
* gomp-constants.h (GOMP_MAP_FLAG_SPECIAL_3): Define special bit macro.
(GOMP_MAP_IMPLICIT): New special map kind bits value.
(GOMP_MAP_FLAG_SPECIAL_BITS): Define helper mask for whole set of
special map kind bits.
(GOMP_MAP_IMPLICIT_P): New predicate macro for implicit map kinds.
gcc/ChangeLog:
* tree.h (OMP_CLAUSE_MAP_RUNTIME_IMPLICIT_P): New access macro for
'implicit' bit, using 'base.deprecated_flag' field of tree_node.
* tree-pretty-print.c (dump_omp_clause): Add support for printing
implicit attribute in tree dumping.
* gimplify.c (gimplify_adjust_omp_clauses_1):
Set OMP_CLAUSE_MAP_RUNTIME_IMPLICIT_P to 1 if map clause is implicitly
created.
(gimplify_adjust_omp_clauses): Adjust place of adding implicitly created
clauses, from simple append, to starting of list, after non-map clauses.
* omp-low.c (lower_omp_target): Add GOMP_MAP_IMPLICIT bits into kind
values passed to libgomp for implicit maps.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/target-implicit-map-1.c: New test.
* c-c++-common/goacc/combined-reduction.c: Adjust scan test pattern.
* c-c++-common/goacc/firstprivate-mappings-1.c: Likewise.
* c-c++-common/goacc/mdc-1.c: Likewise.
* g++.dg/goacc/firstprivate-mappings-1.C: Likewise.
libgomp/ChangeLog:
* target.c (gomp_map_vars_existing): Add 'bool implicit' parameter, add
implicit map handling to allow a "superset" existing map as valid case.
(get_kind): Adjust to filter out GOMP_MAP_IMPLICIT bits in return value.
(get_implicit): New function to extract implicit status.
(gomp_map_fields_existing): Adjust arguments in calls to
gomp_map_vars_existing, and add uses of get_implicit.
(gomp_map_vars_internal): Likewise.
* testsuite/libgomp.c-c++-common/target-implicit-map-1.c: New test.
The following patch implements what I've been talking about earlier,
honor that for explicit num_teams clause we create at least the
lower-bound (if not specified, upper-bound) teams in the league.
For host fallback, it still means we only have one thread doing all the
teams, sequentially one after another.
For PTX and GCN, I think the new teams-2.c test and maybe teams-4.c too
will or might fail.
For these offloads, I think it is ok to remove symbols no longer used
from libgomp.a.
If num_teams_lower is bigger than the provided num_blocks or num_workgroups,
we should arrange for gomp_num_teams_var to be num_teams_lower - 1,
stop using the %ctaid.x or __builtin_gcn_dim_pos (0) for omp_get_team_num ()
and instead use for it some .shared var that GOMP_teams4 initializes to
%ctaid.x or __builtin_gcn_dim_pos (0) when first and for !first
increment that by num_blocks or num_workgroups each time and only
return false when we are above num_teams_lower.
Any help with actually implementing this for the 2 architectures highly
appreciated.
2021-11-12 Jakub Jelinek <jakub@redhat.com>
gcc/
* omp-builtins.def (BUILT_IN_GOMP_TEAMS): Remove.
(BUILT_IN_GOMP_TEAMS4): New.
* builtin-types.def (BT_FN_VOID_UINT_UINT): Remove.
(BT_FN_BOOL_UINT_UINT_UINT_BOOL): New.
* omp-low.c (lower_omp_teams): Use GOMP_teams4 instead of
GOMP_teams, pass to it also num_teams lower-bound expression
or a dup of upper-bound if it is missing and a flag whether
it is the first call or not.
gcc/fortran/
* types.def (BT_FN_VOID_UINT_UINT): Remove.
(BT_FN_BOOL_UINT_UINT_UINT_BOOL): New.
libgomp/
* libgomp_g.h (GOMP_teams4): Declare.
* libgomp.map (GOMP_5.1): Export GOMP_teams4.
* target.c (GOMP_teams4): New function.
* config/nvptx/target.c (GOMP_teams): Remove.
(GOMP_teams4): New function.
* config/gcn/target.c (GOMP_teams): Remove.
(GOMP_teams4): New function.
* testsuite/libgomp.c/teams-4.c (main): Expect exactly 2
teams instead of <= 2.
* testsuite/libgomp.c-c++-common/teams-2.c: New test.
When thinking about GOMP_teams3, I've realized that using global variables
for the values returned by omp_get_num_teams()/omp_get_team_num() calls
is incorrect even with our right now dumb way of implementing host teams.
The problems are two, one is if host teams is used from multiple pthread_create
created threads - the spec says that host teams can't be nested inside of
explicit parallel or other teams constructs, but with pthread_create the
standard says obviously nothing about it. Another more important thing
is host fallback, right now we don't do anything for omp_get_num_teams()
or omp_get_team_num() which was fine before host teams was introduced and
the 5.1 requirement that num_teams clause specifies minimum of teams, but
with the global vars it means inside of target teams num_teams (2) we happily
return omp_get_num_teams() == 4 if the target teams is inside of host teams
with num_teams(4). With target fallback being invoked from parallel
regions global vars simply can't work right on the host.
So, this patch moves them to struct gomp_thread and propagates those for
parallel to child threads. For host fallback, the implicit zeroing of
*thr results in us returning omp_get_num_teams () == 1 and
omp_get_team_num () == 0 which is fine for target teams without num_teams
clause, for target teams with num_teams clause something to work on and
for target without teams nested in it I've asked on omp-lang what should
be done.
2021-11-11 Jakub Jelinek <jakub@redhat.com>
* libgomp.h (struct gomp_thread): Add num_teams and team_num members.
* team.c (struct gomp_thread_start_data): Likewise.
(gomp_thread_start): Initialize thr->num_teams and thr->team_num.
(gomp_team_start): Initialize start_data->num_teams and
start_data->team_num. Update nthr->num_teams and nthr->team_num.
* teams.c (gomp_num_teams, gomp_team_num): Remove.
(GOMP_teams_reg): Set and restore thr->num_teams and thr->team_num
instead of gomp_num_teams and gomp_team_num.
(omp_get_num_teams): Use thr->num_teams + 1 instead of gomp_num_teams.
(omp_get_team_num): Use thr->team_num instead of gomp_team_num.
* testsuite/libgomp.c/teams-4.c: New test.
In OpenMP 5.1, num_teams clause can accept either one expression as before,
but it in that case changed meaning, rather than create <= expression
teams it is now create == expression teams. Or it accepts two expressions
separated by :, with the meaning that the first is low bound and second upper
bound on how many teams should be created. The other ways to set number of
teams are upper bounds with lower bound of 1.
The following patch does parsing of this for C/C++. For host teams, we
actually don't need to do anything further right now, we always create
(pretend to create) exactly the requested number of teams, so we can just
evaluate and throw away the lower bound for now.
For teams nested in target, we don't guarantee that though and further
work will be needed.
In particular, omplower now turns the teams part of:
struct S { S (); S (const S &); ~S (); int s; };
void bar (S &, S &);
int baz ();
_Pragma ("omp declare target to (baz)");
void
foo (void)
{
S a, b;
#pragma omp target private (a) map (b)
{
#pragma omp teams firstprivate (b) num_teams (baz ())
{
bar (a, b);
}
}
}
into:
retval.0 = baz ();
retval.1 = retval.0;
{
unsigned int retval.3;
struct S * D.2549;
struct S b;
retval.3 = (unsigned int) retval.1;
D.2549 = .omp_data_i->b;
S::S (&b, D.2549);
#pragma omp teams num_teams(retval.1) firstprivate(b) shared(a)
__builtin_GOMP_teams (retval.3, 0);
{
bar (&a, &b);
}
S::~S (&b);
#pragma omp return(nowait)
}
IMHO we want a new API, say GOMP_teams3 which will take 3 arguments
instead of 2 (the lower and upper bounds from num_teams and thread_limit)
and will return a bool whether it should do the teams body or not.
And, we should add right before outermost {} above
while (__builtin_GOMP_teams3 ((unsigned) retval.1, (unsigned) retval.1, 0))
and remove the __builtin_GOMP_teams call. The current function performs
exit equivalent (at least on NVPTX) which seems bad because that means
the destructors of e.g. private variables on target aren't invoked, and
at the current placement neither destructors of the already constructed
privatized variables in teams.
I'll do this next on the compiler side, but I'm afraid I'll need help
with the nvptx and amdgcn implementations. E.g. for nvptx, we won't be
able to use %ctaid.x . I think ideal would be to use a .shared
integer variable for the omp_get_team_num value, but I don't have any
experience with that, are .shared variables zero initialized by default,
or do they have random value at start? PTX docs say they aren't initializable.
2021-11-11 Jakub Jelinek <jakub@redhat.com>
gcc/
* tree.h (OMP_CLAUSE_NUM_TEAMS_EXPR): Rename to ...
(OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR): ... this.
(OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR): Define.
* tree.c (omp_clause_num_ops): Increase num ops for
OMP_CLAUSE_NUM_TEAMS to 2.
* tree-pretty-print.c (dump_omp_clause): Print optional lower bound
for OMP_CLAUSE_NUM_TEAMS.
* gimplify.c (gimplify_scan_omp_clauses): Gimplify
OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR if non-NULL.
(optimize_target_teams): Use OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR instead
of OMP_CLAUSE_NUM_TEAMS_EXPR. Handle OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR.
* omp-low.c (lower_omp_teams): Use OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR
instead of OMP_CLAUSE_NUM_TEAMS_EXPR.
* omp-expand.c (expand_teams_call, get_target_arguments): Likewise.
gcc/c/
* c-parser.c (c_parser_omp_clause_num_teams): Parse optional
lower-bound and store it into OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR.
Use OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR instead of
OMP_CLAUSE_NUM_TEAMS_EXPR.
(c_parser_omp_target): For OMP_CLAUSE_NUM_TEAMS evaluate before
combined target teams even lower-bound expression.
gcc/cp/
* parser.c (cp_parser_omp_clause_num_teams): Parse optional
lower-bound and store it into OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR.
Use OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR instead of
OMP_CLAUSE_NUM_TEAMS_EXPR.
(cp_parser_omp_target): For OMP_CLAUSE_NUM_TEAMS evaluate before
combined target teams even lower-bound expression.
* semantics.c (finish_omp_clauses): Handle
OMP_CLAUSE_NUM_TEAMS_LOWER_EXPR of OMP_CLAUSE_NUM_TEAMS clause.
* pt.c (tsubst_omp_clauses): Likewise.
(tsubst_expr): For OMP_CLAUSE_NUM_TEAMS evaluate before
combined target teams even lower-bound expression.
gcc/fortran/
* trans-openmp.c (gfc_trans_omp_clauses): Use
OMP_CLAUSE_NUM_TEAMS_UPPER_EXPR instead of OMP_CLAUSE_NUM_TEAMS_EXPR.
gcc/testsuite/
* c-c++-common/gomp/clauses-1.c (bar): Supply lower-bound expression
to half of the num_teams clauses.
* c-c++-common/gomp/num-teams-1.c: New test.
* c-c++-common/gomp/num-teams-2.c: New test.
* g++.dg/gomp/attrs-1.C (bar): Supply lower-bound expression
to half of the num_teams clauses.
* g++.dg/gomp/attrs-2.C (bar): Likewise.
* g++.dg/gomp/num-teams-1.C: New test.
* g++.dg/gomp/num-teams-2.C: New test.
libgomp/
* testsuite/libgomp.c-c++-common/teams-1.c: New test.
... that got broken by recent commit c057ed9c52
"openmp: Fix up strtoul and strtoull uses in libgomp", resulting in spurious
FAILs for tests specifying 'dg-set-target-env-var "GOMP_OPENACC_DIM" "[...]"'.
libgomp/
* env.c (parse_gomp_openacc_dim): Restore parsing.
The teams construct only permits omp_get_num_teams and omp_get_team_num
as API call in strictly nested regions - check for it.
Additionally, for Fortran, using DECL_NAME does not show the mangled
name, hence, DECL_ASSEMBLER_NAME had to be used to.
Finally, 'target device(ancestor:1)' wrongly rejected non-API calls
as well.
PR middle-end/102972
gcc/ChangeLog:
* omp-low.c (omp_runtime_api_call): Use DECL_ASSEMBLER_NAME to get
internal Fortran name; new permit_num_teams arg to permit
omp_get_num_teams and omp_get_team_num.
(scan_omp_1_stmt): Update call to it, add missing call for
reverse offload, and check for strictly nested API calls in teams.
gcc/testsuite/ChangeLog:
* c-c++-common/gomp/target-device-ancestor-3.c: Add non-API
routine test.
* gfortran.dg/gomp/order-6.f90: Add missing bind(C).
* c-c++-common/gomp/teams-3.c: New test.
* gfortran.dg/gomp/teams-3.f90: New test.
* gfortran.dg/gomp/teams-4.f90: New test.
libgomp/ChangeLog:
* testsuite/libgomp.c-c++-common/icv-3.c: Nest API calls inside
parallel construct.
* testsuite/libgomp.c-c++-common/icv-4.c: Likewise.
* testsuite/libgomp.c/target-3.c: Likewise.
* testsuite/libgomp.c/target-5.c: Likewise.
* testsuite/libgomp.c/target-6.c: Likewise.
* testsuite/libgomp.c/target-teams-1.c: Likewise.
* testsuite/libgomp.c/teams-1.c: Likewise.
* testsuite/libgomp.c/thread-limit-2.c: Likewise.
* testsuite/libgomp.c/thread-limit-3.c: Likewise.
* testsuite/libgomp.c/thread-limit-4.c: Likewise.
* testsuite/libgomp.c/thread-limit-5.c: Likewise.
* testsuite/libgomp.fortran/icv-3.f90: Likewise.
* testsuite/libgomp.fortran/icv-4.f90: Likewise.
* testsuite/libgomp.fortran/teams1.f90: Likewise.
This patch upgrades the pre-VRP threading passes to fully resolving
backward threaders, and removes the post-VRP threading passes altogether.
With it, we reduce the number of threaders in our pipeline from 9 to 7.
This will leave DOM as the only forward threader client. When the ranger
can handle floats, we should be able to upgrade the pre-DOM threaders to
fully resolving threaders and kill the embedded DOM threader.
The numbers are as follows:
prev: # threads in backward + vrp-threaders = 92624
now: # threads in backward threaders = 94275
Gain: +1.78%
prev: # total threads: 189495
now: # total threads: 193714
Gain: +2.22%
The numbers are not as great as my initial proposal, but I've
recently pushed all the work that got us to this point ;-).
And... the compilation improves by 1.32%!
There's a regression on uninit-pred-7_a.c that I've yet to look at. I
want to make sure it's not a missing thread. If it is, I'll create a PR
and own it.
Also, the tree-ssa/phi_on_compare-*.c tests have all regressed. This
seems to be some special case the forward threader handles that the
backward threader does not (edge_forwards_cmp_to_conditional_jump*).
I haven't dug deep to see if this is solveable within our
infrastructure, but a cursory look shows that even though the VRP
threader threads this, the *.optimized dump ends with more conditional
jumps than without the optimization. I'd like to punt on this for
now, because DOM actually catches this through its lone use of the
forward threader (I've adjusted the tests). However, we will need to
address this sooner or later, if indeed it's still improving the final
assembly.
gcc/ChangeLog:
* passes.def: Replace the pass_thread_jumps before VRP* with
pass_thread_jumps_full. Remove all pass_vrp_threader instances.
* tree-ssa-threadbackward.c (pass_data_thread_jumps_full):
Remove hyphen from "thread-full" name.
libgomp/ChangeLog:
* testsuite/libgomp.graphite/force-parallel-4.c: Adjust for threading changes.
* testsuite/libgomp.graphite/force-parallel-8.c: Same.
gcc/testsuite/ChangeLog:
* gcc.dg/loop-unswitch-2.c: Adjust for threading changes.
* gcc.dg/old-style-asm-1.c: Same.
* gcc.dg/tree-ssa/phi_on_compare-1.c: Same.
* gcc.dg/tree-ssa/phi_on_compare-2.c: Same.
* gcc.dg/tree-ssa/phi_on_compare-3.c: Same.
* gcc.dg/tree-ssa/phi_on_compare-4.c: Same.
* gcc.dg/tree-ssa/pr20701.c: Same.
* gcc.dg/tree-ssa/pr21001.c: Same.
* gcc.dg/tree-ssa/pr21294.c: Same.
* gcc.dg/tree-ssa/pr21417.c: Same.
* gcc.dg/tree-ssa/pr21559.c: Same.
* gcc.dg/tree-ssa/pr21563.c: Same.
* gcc.dg/tree-ssa/pr49039.c: Same.
* gcc.dg/tree-ssa/pr59597.c: Same.
* gcc.dg/tree-ssa/pr61839_1.c: Same.
* gcc.dg/tree-ssa/pr61839_3.c: Same.
* gcc.dg/tree-ssa/pr66752-3.c: Same.
* gcc.dg/tree-ssa/pr68198.c: Same.
* gcc.dg/tree-ssa/pr77445-2.c: Same.
* gcc.dg/tree-ssa/pr77445.c: Same.
* gcc.dg/tree-ssa/ranger-threader-1.c: Same.
* gcc.dg/tree-ssa/ranger-threader-2.c: Same.
* gcc.dg/tree-ssa/ranger-threader-4.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-1.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-11.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-12.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-14.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-16.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-2b.c: Same.
* gcc.dg/tree-ssa/ssa-dom-thread-7.c: Same.
* gcc.dg/tree-ssa/ssa-thread-14.c: Same.
* gcc.dg/tree-ssa/ssa-thread-backedge.c: Same.
* gcc.dg/tree-ssa/ssa-vrp-thread-1.c: Same.
* gcc.dg/tree-ssa/vrp02.c: Same.
* gcc.dg/tree-ssa/vrp03.c: Same.
* gcc.dg/tree-ssa/vrp05.c: Same.
* gcc.dg/tree-ssa/vrp06.c: Same.
* gcc.dg/tree-ssa/vrp07.c: Same.
* gcc.dg/tree-ssa/vrp08.c: Same.
* gcc.dg/tree-ssa/vrp09.c: Same.
* gcc.dg/tree-ssa/vrp33.c: Same.
* gcc.dg/uninit-pred-9_b.c: Same.
* gcc.dg/uninit-pred-7_a.c: xfail.
I've found we claim to support non-rectangular loops, but don't actually
support those in Fortran, as can be seen on:
integer i, j
!$omp parallel do collapse(2)
do i = 0, 10
do j = 0, i
end do
end do
end
To support this, the Fortran FE needs to allow the valid forms of
non-rectangular loops and disallow others, so mainly it needs its
updated version of c-omp.c c_omp_check_loop_iv etc., plus for non-rectangular
lb or ub expressions emit a TREE_VEC instead of normal expression as the C/C++ FE
do, plus testsuite coverage.
2021-10-27 Jakub Jelinek <jakub@redhat.com>
* libgomp.texi (OpenMP 5.0): Mention that Non-rectangular loop nests
aren't implemented for Fortran yet.
