Mention that the HSA_XNACK variable is automatically set by the toolchain.
libgomp/ChangeLog:
* libgomp.texi (AMD GCN): Mention HSA_XNACK is set automatically.
This patch implements "managed" memory for AMD GCN GPUs in OpenMP. It
builds on the support added to the NVPTX libgomp for CUDA Managed Memory, a
week or two ago.
These features were first posted here a few years ago, as part of a larger
Unified Shared Memory patch series, and then in a slightly changed version just
over a year ago. Hopefully this time the controversial bits have been removed.
Since we do not use HIP we cannot use hipMallocManaged, so this patch attempts
to replicate the same effect by setting the appropriate attributes. This works
on more devices than support proper USM, but still I cannot be sure that the
settings are correct for every device out there (I have tested on gfx900,
gfx906, gfx908, gfx90a, and gfx1100).
The HSA header file update uses the most recent files relicensed for us by AMD,
at the time of the first patch posting. Those files have certainly moved on in
the upstream sources, but I did not ask to get those relicensed.
include/ChangeLog:
* hsa.h: Import newer version.
* hsa_ext_amd.h: Likewise.
* hsa_ext_image.h: Likewise.
libgomp/ChangeLog:
* Makefile.in: Regenerate.
* libgomp-plugin.h (gomp_simple_alloc_init_context): New prototype.
(gomp_simple_alloc_register_memory): New prototype.
(gomp_simple_alloc): New prototype.
(gomp_simple_free): New prototype.
(gomp_simple_realloc): New prototype.
* libgomp.h (gomp_simple_alloc_init_context): Move to libgomp-plugin.h.
(gomp_simple_alloc_register_memory): Likewise.
(gomp_simple_alloc): Likewise.
(gomp_simple_free): Likewise.
(gomp_simple_realloc): Likewise.
* libgomp.texi: Update AMD managed memory description.
* plugin/Makefrag.am (libgomp_plugin_gcn_la_SOURCES): Add
simple-allocator.c and plugin/mutex.c.
* plugin/plugin-gcn.c: Include sys/mman.h and unistd.h.
(struct hsa_runtime_fn_info): Add hsa_amd_svm_attributes_set_fn.
(dump_hsa_system_info): Add HSA_AMD_SYSTEM_INFO_SVM_SUPPORTED and
HSA_AMD_SYSTEM_INFO_SVM_ACCESSIBLE_BY_DEFAULT to the GCN_DEBUG output.
(init_hsa_runtime_functions): Add hsa_amd_svm_attributes_set.
(isa_matches_agent): Add a new error message for the case where the
ISA doesn't match but the name does.
(managed_ctx): New variable.
(managed_heap_create): New function.
(GOMP_OFFLOAD_managed_alloc): Likewise.
(GOMP_OFFLOAD_managed_free): Likewise.
* simple-allocator.c (gomp_fatal): New macro.
* testsuite/lib/libgomp.exp (check_effective_target_omp_managedmem):
Add amdgcn support checker.
(check_effective_target_offload_target_amdgcn_with_xnack): New.
* testsuite/libgomp.c-c++-common/requires-4.c: Ignore xnack warning.
* testsuite/libgomp.c-c++-common/requires-4a.c: Ignore xnack warning.
* testsuite/libgomp.c-c++-common/requires-5.c: Ignore xnack warning.
* testsuite/libgomp.c++/alloc-managed-1.C: Add -mxnack=on, if needed.
* testsuite/libgomp.c/alloc-managed-1.c: Likewise.
* testsuite/libgomp.c/alloc-managed-2.c: Likewise.
* testsuite/libgomp.c/alloc-managed-3.c: Likewise.
* testsuite/libgomp.c/alloc-managed-4.c: Likewise.
* testsuite/libgomp.fortran/alloc-managed-1.f90: Likewise.
* plugin/mutex.c: New file.
Consider the following source code, assuming tiles is allocatable:
```
!$omp target enter data map(var%tiles(1)%den1, var%tiles(1)%den2) ! (1)
[...]
!$omp target ! implicitly maps var, which triggers deep mapping of tiles (2)
```
Each omp directive causes a run-time error in libgomp:
(1) libgomp: Mapped array elements must be the same (0x14d729c0 vs 0x14d72a18)
(2) libgomp: Trying to map into device [0x3704ca50..0x3704cb00) object when
[0x3704ca50..0x3704caa8) is already mapped
Regarding (1), the OpenMP spec has the following restriction: "If multiple list
items are explicitly mapped on the same construct and have the same containing
array or have base pointers that share original storage, and if any of the list
items do not have corresponding list items that are present in the device data
environment prior to a task encountering the construct, then the list items must
refer to *the same array elements* of either the containing array or the
implicit array of the base pointers."
Because tiles is allocatable, we cannot prove at compile time that array
elements are the same, so the check is deferred to libgomp. But there the
condition enforcing that all addresses are the same is too strict, so this patch
relaxes it to only check that addresses are sorted in increasing order.
The OpenMP spec allows (2) as long as it is implicit, without extending the
original mapping. So this patch sets the GOMP_MAP_IMPLICIT flag appropriately
on deep maps at compile time to let libgomp know that it is fine.
This patch ensures that such user code is accepted by:
(1) Setting the GOMP_MAP_IMPLICIT flag appropriately on deep maps;
(2) Relaxing the restriction on struct mapping from different containing arrays,
so that the element index need not be the same, instead addresses must be sorted
in increasing order.
This fixes the two errors currently seen when running SPEC HPC clvleaf
benchmark. However, further mapping issues prevent the benchmark from running to
completion.
PR fortran/120505
gcc/ChangeLog:
* omp-low.cc (lower_omp_target): Set GOMP_MAP_IMPLICIT flag.
libgomp/ChangeLog:
* target.c (gomp_map_vars_internal): Allow struct mapping from different
containing array elements as long as adresses are in increasing order.
* testsuite/libgomp.c-c++-common/map-arrayofstruct-2.c: Adjust
dg-output.
* testsuite/libgomp.c-c++-common/map-arrayofstruct-3.c: Likewise.
* testsuite/libgomp.fortran/map-subarray-5.f90: Likewise.
* testsuite/libgomp.fortran/map-subarray-10.f90: New test.
* testsuite/libgomp.fortran/map-subarray-9.f90: New test.
On Mon, Nov 03, 2025 at 01:34:28PM -0500, Marek Polacek via Gcc wrote:
> I would like us to declare that C++20 is no longer experimental and
> change the default dialect to gnu++20. Last time we changed the default
> was over 5 years ago in GCC 11:
> <https://gcc.gnu.org/git/?p=gcc.git;a=commitdiff;h=0801f419440c14f6772b28f763ad7d40f7f7a580>
> and before that in 2015 in GCC 6.1, so this happens roughly every 5 years.
>
> I had been hoping to move to C++20 in GCC 15 (see bug 113920), but at that time
> libstdc++ still had incomplete C++20 support and the compiler had issues to iron
> out (mangling of concepts, modules work, etc.). Are we ready now? Is anyone
> aware of any blockers? Presumably we still wouldn't enable Modules by default.
>
> I'm willing to do the work if we decide that it's time to switch the default
> C++ dialect (that includes updating cxx-status.html and adding a new caveat to
> changes.html).
I haven't seen a patch posted for this, so just that something is posted
during stage1 if we decide to do it, here is a patch.
The patch makes -std=gnu++20 the default C++ dialect and documents that
-fmodules is still not implied by that or -std=c++20 and modules support
is still experimental.
2025-11-26 Jakub Jelinek <jakub@redhat.com>
gcc/
* doc/invoke.texi (gnu++17): Remove comment about the default.
(c++20): Remove note about experimental support, except add a note
that modules are still experimental and need to be enabled separately.
(gnu++20): Likewise. Move here comment about the default.
(fcoroutines): Mention it is enabled by default for C++20 and later.
* doc/standards.texi: Document that the default for C++ is
-std=gnu++20.
gcc/c-family/
* c-opts.cc (c_common_init_options): Call set_std_cxx20 rather than
set_std_cxx17.
* c.opt (std=c++2a): Change description to deprecated option wording.
(std=c++20): Remove experimental support part.
(std=c++2b): Change description to deprecated option wording.
(std=gnu++2a): Likewise.
(std=gnu++20): Remove experimental support part.
(std=gnu++2b): Change description to deprecated option wording.
gcc/testsuite/
* lib/target-supports.exp: Set cxx_default to c++20 rather than
c++17.
* lib/g++-dg.exp (g++-std-flags): Reorder list to put 20 first
and 17 after 26.
* g++.dg/debug/pr80461.C (bar): Use v = v + 1; instead of ++v;.
* g++.dg/debug/pr94459.C: Add -std=gnu++17 to dg-options.
* g++.dg/diagnostic/virtual-constexpr.C: Remove dg-skip-if,
instead use { c++11 && c++17_down } effective target instead of
c++11.
* g++.dg/guality/pr67192.C: Add -std=gnu++17.
* g++.dg/torture/pr84961-1.C: Likewise.
* g++.dg/torture/pr84961-2.C: Likewise.
* g++.dg/torture/pr51482.C (anim_track_bez_wvect::tangent): Cast
key_class to int before multiplying it by float.
* g++.dg/torture/stackalign/unwind-4.C (foo): Use g_a = g_a + 1;
instead of g_a++;.
* g++.dg/tree-prof/partition1.C (bar): Use l = l + 1; return l;
instead of return ++l;.
* obj-c++.dg/exceptions-3.mm: Add -std=gnu++17.
* obj-c++.dg/exceptions-5.mm: Likewise.
libgomp/
* testsuite/libgomp.c++/atomic-12.C (main): Add ()s around array
reference index.
* testsuite/libgomp.c++/atomic-13.C: Likewise.
* testsuite/libgomp.c++/atomic-8.C: Likewise.
* testsuite/libgomp.c++/atomic-9.C: Likewise.
* testsuite/libgomp.c++/loop-6.C: Use count = count + 1;
return count > 0; instead of return ++count > 0;.
* testsuite/libgomp.c++/pr38650.C: Add -std=gnu++17.
* testsuite/libgomp.c++/target-lambda-1.C (merge_data_func):
Use [=,this] instead of just [=] in lambda captures.
* testsuite/libgomp.c-c++-common/target-40.c (f1): Use v += 1;
instead of v++;.
* testsuite/libgomp.c-c++-common/depend-iterator-2.c: Use v = v + 1;
instead of v++.
The toolchain autobuilds for ia64 failed ([1]) yesterday with:
```
libtool: compile: /usr/src/t2-src/src.gcc.ia64-toolchain.251121.040147.278918/gcc-16-20251116/objs/gcc/xgcc-wrapper /usr/src/t2-src/src.gcc.ia64-toolchain.251121.040147.278918/gcc-16-20251116/objs/./gcc/xgcc -B/usr/src/t2-src/src.gcc.ia64-toolchain.251121.040147.278918/gcc-16-20251116/objs/./gcc/ -B/usr/src/t2-src/build/ia64-toolchain-24-svn-generic-ia64-itanium2-cross-linux/TOOLCHAIN/cross/usr/ia64-t2-linux-gnu/bin/ -B/usr/src/t2-src/build/ia64-toolchain-24-svn-generic-ia64-itanium2-cross-linux/TOOLCHAIN/cross/usr/ia64-t2-linux-gnu/lib/ -isystem /usr/src/t2-src/build/ia64-toolchain-24-svn-generic-ia64-itanium2-cross-linux/TOOLCHAIN/cross/usr/ia64-t2-linux-gnu/include -isystem /usr/src/t2-src/build/ia64-toolchain-24-svn-generic-ia64-itanium2-cross-linux/TOOLCHAIN/cross/usr/ia64-t2-linux-gnu/sys-include --sysroot=/usr/src/t2-src/build/ia64-toolchain-24-svn-generic-ia64-itanium2-cross-linux -DHAVE_CONFIG_H -I. -I../../../libgomp -I../../../libgomp/config/linux/ia64 -I../../../libgomp/config/linux -I../../../libgomp/config/posix -I../../../libgomp -I../../../libgomp/../include -Wall -Werror -ftls-model=initial-exec -pthread -DUSING_INITIAL_EXEC_TLS -g -O2 -MT oacc-cuda.lo -MD -MP -MF .deps/oacc-cuda.Tpo -c ../../../libgomp/oacc-cuda.c -o oacc-cuda.o >/dev/null 2>&1
../../../libgomp/affinity-fmt.c: In function 'gomp_display_affinity':
../../../libgomp/affinity-fmt.c:330:25: error: initialization discards 'const' qualifier from pointer target type [-Werror=discarded-qualifiers]
330 | char *q = strchr (p + 1, '}');
| ^~~~~~
```
[1]: https://github.com/johnny-mnemonic/toolchain-autobuilds/actions/runs/19559235881
This is not ia64-specific but due to the changes in the recent glibc
commit "Implement C23 const-preserving standard library macros" (i.e.
[2]) now requiring "char *q" to be a pointer to a const char to compile
w/o error because of the return value of strchr() .
[2]: https://sourceware.org/git/?p=glibc.git;a=commit;h=cd748a63ab1a7ae846175c532a3daab341c62690
Also see the related discussion at [3] for details.
[3]: https://sourceware.org/pipermail/libc-alpha/2025-November/172809.html
The GCC build is fixed by the attached patch, see [4] for a successful
build with the then latest snapshots of binutils, glibc and GCC.
[4]: https://github.com/johnny-mnemonic/toolchain-autobuilds/actions/runs/19585045571
Idea from Tomas, attached patch from me.
Cheers,
Frank
0001-libgomp-Fix-GCC-build-after-glibc-cd748a6.patch
From 80af9c233c694904174b54a59404d311378f41f8 Mon Sep 17 00:00:00 2001
From: Frank Scheiner <frank.scheiner@web.de>
Date: Sat, 22 Nov 2025 14:58:10 +0100
Subject: [PATCH] libgomp: Fix GCC build after glibc@cd748a6
char *q needs to be a pointer to a const char for the return value of
strchr() with glibc after "Implement C23 const-preserving standard library
macros".
[glibc@cd748a6]: https://sourceware.org/git/?p=glibc.git;a=commit;h=cd748a63ab1a7ae846175c532a3daab341c62690
libgomp/ChangeLog:
* affinity-fmt.c: Make char *q a pointer to a const char.
In r11-3059-g8183ebcdc1c843, Julian fixed a few issues with
atomic_capture-2.c relying on iteration order guarantees that do not
exist under OpenACC parallelized loops and, notably, do not happen even
by accident on AMDGCN.
The atomic_capture-3.c testcase was made by copying it from
atomic_capture-2.c and adding additional options in commit
r12-310-g4cf3b10f27b199, but from an older version of
atomic_capture-2.c, which lacked these ordering fixes fixes, so they
resurfaced in this test.
This patch ports those fixes from atomic_capture-2.c into
atomic_capture-3.c.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-c-c++-common/atomic_capture-3.c: Copy
changes in r11-3059-g8183ebcdc1c843 from atomic_capture-2.c.
This patch implements C++ support for the "begin declare variant"
construct. The OpenMP specification is hazy on interaction of this
feature with C++ language features. Variant functions in classes are
supported but must be defined as members in the class definition,
using an unqualified name for the base function which also must be
present in that class. Similarly variant functions in a namespace can
only be defined in that namespace using an unqualified name for a base
function already declared in that namespace. Variants for template
functions or inside template classes seem to (mostly) work.
gcc/c-family/ChangeLog
* c-omp.cc (c_omp_directives): Uncomment "begin declare variant"
and "end declare variant".
gcc/cp/ChangeLog
* cp-tree.h (struct cp_omp_declare_variant_attr): New.
(struct saved_scope): Add omp_declare_variant_attribute field.
* decl.cc (omp_declare_variant_finalize_one): Add logic to inject
"this" parameter for method calls.
* parser.cc (cp_parser_skip_to_pragma_omp_end_declare_variant): New.
(cp_parser_translation_unit): Handle leftover "begin declare variant"
functions.
(omp_start_variant_function): New.
(omp_finish_variant_function): New.
(omp_maybe_record_variant_base): New.
(cp_parser_init_declarator): Handle variant functions.
(cp_parser_class_specifier): Handle deferred lookup of base functions
when the entire class has been seen.
(cp_parser_member_declaration): Handle variant functions.
(cp_finish_omp_declare_variant): Merge context selectors if in
a "begin declare variant" block.
(cp_parser_omp_begin): Match "omp begin declare variant". Adjust
error messages.
(cp_parser_omp_end): Match "omp end declare variant".
* parser.h (struct omp_begin_declare_variant_map_entry): New.
(struct cp_parser): Add omp_begin_declare_variant_map field.
* semantics.cc (finish_translation_unit): Detect unmatched
"omp begin declare variant".
gcc/testsuite/ChangeLog
* g++.dg/gomp/delim-declare-variant-1.C: New.
* g++.dg/gomp/delim-declare-variant-2.C: New.
* g++.dg/gomp/delim-declare-variant-3.C: New.
* g++.dg/gomp/delim-declare-variant-4.C: New.
* g++.dg/gomp/delim-declare-variant-5.C: New.
* g++.dg/gomp/delim-declare-variant-6.C: New.
* g++.dg/gomp/delim-declare-variant-7.C: New.
* g++.dg/gomp/delim-declare-variant-40.C: New.
* g++.dg/gomp/delim-declare-variant-41.C: New.
* g++.dg/gomp/delim-declare-variant-50.C: New.
* g++.dg/gomp/delim-declare-variant-51.C: New.
* g++.dg/gomp/delim-declare-variant-52.C: New.
* g++.dg/gomp/delim-declare-variant-70.C: New.
* g++.dg/gomp/delim-declare-variant-71.C: New.
libgomp/
* testsuite/libgomp.c++/bdv_module1.C: New.
* testsuite/libgomp.c++/bdv_module1_main.C: New.
* testsuite/libgomp.c++/bdv_module2.C: New.
* testsuite/libgomp.c++/bdv_module2_impl.C: New.
* testsuite/libgomp.c++/bdv_module2_main.C: New.
* testsuite/libgomp.c++/bdv_module3.C: New.
* testsuite/libgomp.c++/bdv_module3_impl.C: New.
* testsuite/libgomp.c++/bdv_module3_main.C: New.
* testsuite/libgomp.c++/delim-declare-variant-1.C: New.
* testsuite/libgomp.c++/delim-declare-variant-2.C: New.
* testsuite/libgomp.c++/delim-declare-variant-7.C: New.
Co-Authored-By: Julian Brown <julian@codesourcery.com>
Co-Authored-By: waffl3x <waffl3x@baylibre.com>
In the testcases, the kernels scheduled on queues 11, 12, 13, 14 have
data dependencies on, respectively, 'b', 'c', 'd', and 'e', as they
write to them.
However, they also have a data dependency on 'a' and 'N', as they read
those.
Previously, the testcases exited 'a' on queue 10 and 'N' on queue 15,
meaning that it was possible for the aforementioned kernels to execute
and to have 'a' and 'N' pulled under their feet.
This patch adds waits for each of the kernels onto queue 10 before
freeing 'a', guaranteeing that 'a' outlives the kernels, and the same on
'N'.
libgomp/ChangeLog:
* testsuite/libgomp.oacc-c-c++-common/data-2-lib.c (explanatory
header): Fix typo.
(main): Insert waits on kernels reading 'a' into queue 10 before
exiting 'a', and waits on kernels reading 'N' into queue 15
before exiting 'N'.
* testsuite/libgomp.oacc-c-c++-common/data-2.c: Ditto.
Due to the if statement in ipa_compute_jump_functions_for_bb, callback
edges were never constructed for builtin functions unless LTO was
enabled. This patch corrects this behavior, allowing GCC to optimize
callbacks more broadly. It also extends our testing capabilities.
gcc/ChangeLog:
* attr-callback.cc (callback_edge_callee_has_attr): New
function.
* attr-callback.h (callback_edge_callee_has_attr): New function
decl.
* ipa-prop.cc (ipa_compute_jump_functions_for_bb): Don't skip
callback carriers when calculating jump functions.
libgomp/ChangeLog:
* testsuite/libgomp.c/ipcp-cb-spec1.c: Remove LTO requirement.
* testsuite/libgomp.c/ipcp-cb-spec2.c: Likewise.
* testsuite/libgomp.c/ipcp-cb1.c: Likewise.
Signed-off-by: Josef Melcr <josef.melcr@suse.com>
This adds support for using Cuda Managed Memory with omp_alloc. AMD support
will be added in a future patch.
There is one new predefined allocator, "ompx_gnu_managed_mem_alloc", plus a
corresponding memory space, which can be used to allocate memory in the
"managed" space.
The nvptx plugin is modified to make the necessary Cuda calls, via two new
(optional) plugin interfaces.
gcc/fortran/ChangeLog:
* openmp.cc (is_predefined_allocator): Use GOMP_OMP_PREDEF_ALLOC_MAX
and GOMP_OMPX_PREDEF_ALLOC_MIN/MAX instead of hardcoded values in the
comment.
include/ChangeLog:
* cuda/cuda.h (cuMemAllocManaged): Add declaration and related
CU_MEM_ATTACH_GLOBAL flag.
* gomp-constants.h (GOMP_OMPX_PREDEF_ALLOC_MAX): Update to 201.
(GOMP_OMP_PREDEF_MEMSPACE_MAX): New constant.
(GOMP_OMPX_PREDEF_MEMSPACE_MIN): New constant.
(GOMP_OMPX_PREDEF_MEMSPACE_MAX): New constant.
libgomp/ChangeLog:
* allocator.c (ompx_gnu_max_predefined_alloc): Update to
ompx_gnu_managed_mem_alloc.
(_Static_assert): Fix assertion messages for allocators and add
new assertions for memspace constants.
(omp_max_predefined_mem_space): New define.
(ompx_gnu_min_predefined_mem_space): New define.
(ompx_gnu_max_predefined_mem_space): New define.
(MEMSPACE_ALLOC): Add check for non-standard memspaces.
(MEMSPACE_CALLOC): Likewise.
(MEMSPACE_REALLOC): Likewise.
(MEMSPACE_VALIDATE): Likewise.
(predefined_ompx_gnu_alloc_mapping): Add ompx_gnu_managed_mem_space.
(omp_init_allocator): Add ompx_gnu_managed_mem_space validation.
* config/gcn/allocator.c (gcn_memspace_alloc): Add check for
non-standard memspaces.
(gcn_memspace_calloc): Likewise.
(gcn_memspace_realloc): Likewise.
(gcn_memspace_validate): Update to validate standard vs non-standard
memspaces.
* config/linux/allocator.c (linux_memspace_alloc): Add managed
memory space handling.
(linux_memspace_calloc): Likewise.
(linux_memspace_free): Likewise.
(linux_memspace_realloc): Likewise (returns NULL for fallback).
* config/nvptx/allocator.c (nvptx_memspace_alloc): Add check for
non-standard memspaces.
(nvptx_memspace_calloc): Likewise.
(nvptx_memspace_realloc): Likewise.
(nvptx_memspace_validate): Update to validate standard vs non-standard
memspaces.
* env.c (parse_allocator): Add ompx_gnu_managed_mem_alloc,
ompx_gnu_managed_mem_space, and some static asserts so I don't forget
them again.
* libgomp-plugin.h (GOMP_OFFLOAD_managed_alloc): New declaration.
(GOMP_OFFLOAD_managed_free): New declaration.
* libgomp.h (gomp_managed_alloc): New declaration.
(gomp_managed_free): New declaration.
(struct gomp_device_descr): Add managed_alloc_func and
managed_free_func fields.
* libgomp.texi: Document ompx_gnu_managed_mem_alloc and
ompx_gnu_managed_mem_space, add C++ template documentation, and
describe NVPTX and AMD support.
* omp.h.in: Add ompx_gnu_managed_mem_space and
ompx_gnu_managed_mem_alloc enumerators, and gnu_managed_mem C++
allocator template.
* omp_lib.f90.in: Add Fortran bindings for new allocator and
memory space.
* omp_lib.h.in: Likewise.
* plugin/cuda-lib.def: Add cuMemAllocManaged.
* plugin/plugin-nvptx.c (nvptx_alloc): Add managed parameter to
support cuMemAllocManaged.
(GOMP_OFFLOAD_alloc): Move contents to ...
(cleanup_and_alloc): ... this new function, and add managed support.
(GOMP_OFFLOAD_managed_alloc): New function.
(GOMP_OFFLOAD_managed_free): New function.
* target.c (gomp_managed_alloc): New function.
(gomp_managed_free): New function.
(gomp_load_plugin_for_device): Load optional managed_alloc
and managed_free plugin APIs.
* testsuite/lib/libgomp.exp: Add check_effective_target_omp_managedmem.
* testsuite/libgomp.c++/alloc-managed-1.C: New test.
* testsuite/libgomp.c/alloc-managed-1.c: New test.
* testsuite/libgomp.c/alloc-managed-2.c: New test.
* testsuite/libgomp.c/alloc-managed-3.c: New test.
* testsuite/libgomp.c/alloc-managed-4.c: New test.
* testsuite/libgomp.fortran/alloc-managed-1.f90: New test.
Co-authored-by: Kwok Cheung Yeung <kcyeung@baylibre.com>
Co-authored-by: Thomas Schwinge <tschwinge@baylibre.com>
Commit r16-5188-g5da963d988e8ea added omp_default_device such that -5
became a conforming device number, but the tests used them to test
for as non-conforming number; now -6 is used.
libgomp/ChangeLog:
PR libgomp/119677
* testsuite/libgomp.c-c++-common/target-is-accessible-1.c: Modify
test as -5 is now a conforming device number.
* testsuite/libgomp.fortran/target-is-accessible-1.f90: Likewise.
OpenMP TR 14 (OpenMP 6.1) adds omp_default_device < -1 as
named constant alongside omp_initial_device and omp_default_device.
GCC supports it already internally via GOMP_DEVICE_DEFAULT_OMP_61,
but this patch now adds the omp_default_device enum/PARAMETER to
omp.h / omp_lib.
Note that PR119677 requests some cleanups, which still have to be
done.
PR libgomp/119677
gcc/fortran/ChangeLog:
* intrinsic.texi (OpenMP Modules): Add omp_default_device.
* openmp.cc (gfc_resolve_omp_context_selector): Accept
omp_default_device as conforming device number.
libgomp/ChangeLog:
* omp.h.in (omp_default_device): New enum value.
* omp_lib.f90.in: New parameter.
* omp_lib.h.in: Likewise
* target.c (gomp_get_default_device): New. Split off from ...
(resolve_device): ... here; call it.
(omp_target_alloc, omp_target_free, omp_target_is_present,
omp_target_memcpy_check, omp_target_memset, omp_target_memset_async,
omp_target_associate_ptr, omp_get_mapped_ptr,
omp_target_is_accessible, omp_pause_resource,
omp_get_uid_from_device): Handle omp_default_device.
* testsuite/libgomp.c/device_uid.c: Likewise.
* testsuite/libgomp.fortran/device_uid.f90: Likewise.
* testsuite/libgomp.c-c++-common/omp-default-device.c: New test.
* testsuite/libgomp.fortran/omp-default-device.f90: New test.
Fix the unified-shared memory test,
libgomp.c++/target-std__multimap-concurrent-usm.C
added in commit r16-1010-g83ca283853f195
libgomp: Add testcases for concurrent access to standard C++ containers
on offload targets, a number of USM variants
This tests includes the actual code of target-std__multimap-concurrent.C.
The issue is that multimap.insert allocates memory – which is freed by
the destructor. However, if the memory is allocated on a device
('insert'), it also needs to be freed there ('clear') as in general
freeing device-allocated memory is not possible on the host.
libgomp/ChangeLog:
* testsuite/libgomp.c++/target-std__multimap-concurrent.C: Fix memory
freeing of device allocated memory with USM.
In OpenMP, pointers are implicitly mapped - which means for Fortran that
their pointer target is also mapped. However, for uninitialized memory,
this means that some random pointee with some random amount of memory is
copied - in the good case, size == 0, but if not, odd things can happen.
Solution: Use 'fptr => null()' before the target mapping or - as done here -
declare the pointer inside the region.
libgomp/ChangeLog:
PR libgomp/122543
* testsuite/libgomp.fortran/omp_target_memset.f90: Move fptr inside
the target to avoid implicit mapping of its uninit pointee.
* testsuite/libgomp.fortran/omp_target_memset-2.f90: Likewise.
... where it currently runs into:
libgomp: cuCtxSynchronize error: an illegal memory access was encountered
... for nvptx, or similarly for GCN:
Memory access fault by GPU node-1 (Agent handle: 0x34d77290) on address 0x7fff3c553000. Reason: Page not present or supervisor privilege.
Fix-up for commit r16-4961-ge2cbcd1b27c0da92bdcd96664064d3d0c1d44e6f
"Fix gimple_copy for OpenMP atomic load/store [PR122281, PR105001]".
PR libgomp/122281
libgomp/
* testsuite/libgomp.c/pr122281.c: Fix for non-USM offloading
execution.
Minor fix-up for commit r13-2545-g9f2fca56593a2b87026b399d26adcdca90705685
"OpenMP, libgomp: Environment variable syntax extension". Supplying an array
of three elements is misleading: 'parse_stacksize' only looks at 'params[0]'.
Just let the compiler figure out the array size (..., as done elsewhere, too).
libgomp/
* env.c (initialize_env): Simplify 'parse_stacksize' call.
This patch introduces a new custom memory allocator for use with pinned
memory (in the case where the Cuda allocator isn't available). In future,
this allocator will also be used for Managed Memory. Both memories are
incompatible with the system malloc because allocated memory cannot share a
page with memory allocated for other purposes.
This means that small allocations will no longer consume an entire page of
pinned memory. Unfortunately, it also means that pinned memory pages will
never be unmapped (although they may be reused). This isn't a technical
limitation; the "free" algorithm could be extended in future, if needed.
The implementation is not perfect; there are various corner cases (especially
related to extending onto new pages) where allocations and reallocations may
be sub-optimal, but it should still be a step forward in support for small
allocations.
I have considered using libmemkind's "fixed" memory but rejected it for three
reasons: 1) libmemkind may not always be present at runtime, 2) there's no
currently documented means to extend a "fixed" kind one page at a time
(although the code appears to have an undocumented function that may do the
job, and/or extending libmemkind to support the MAP_LOCKED mmap flag with its
regular kinds would be straight-forward), 3) Managed Memory benefits from
having the metadata located in different memory and using an external
implementation makes it hard to guarantee this.
libgomp/ChangeLog:
* Makefile.am (libgomp_la_SOURCES): Add simple-allocator.c.
* Makefile.in: Regenerate.
* basic-allocator.c: Mention simple-allocator in the comment.
* config/linux/allocator.c: Include unistd.h.
(pin_ctx): New variable.
(ctxlock): New variable.
(linux_init_pin_ctx): New function.
(linux_memspace_alloc): Use simple-allocator for pinned memory.
(linux_memspace_free): Likewise.
(linux_memspace_realloc): Likewise.
* libgomp.h (gomp_simple_alloc_init_context): New prototype.
(gomp_simple_alloc_register_memory): New prototype.
(gomp_simple_alloc): New prototype.
(gomp_simple_free): New prototype.
(gomp_simple_realloc): New prototype.
* libgomp.texi: Update pinned memory trait documentation.
* testsuite/libgomp.c/alloc-pinned-8.c: New test.
* simple-allocator.c: New file.
Use Cuda to pin memory, instead of Linux mlock, when available.
There are two advantages: firstly, this gives a significant speed boost for
NVPTX offloading, and secondly, it side-steps the usual OS ulimit/rlimit
setting.
The design adds a device independent plugin API for allocating pinned memory,
and then implements it for NVPTX. At present, the other supported devices do
not have equivalent capabilities (or requirements).
libgomp/ChangeLog:
* config/linux/allocator.c: Include assert.h.
(using_device_for_page_locked): New variable.
(linux_memspace_alloc): Add init0 parameter. Support device pinning.
(linux_memspace_calloc): Set init0 to true.
(linux_memspace_free): Support device pinning.
(linux_memspace_realloc): Support device pinning.
(MEMSPACE_ALLOC): Set init0 to false.
* libgomp-plugin.h
(GOMP_OFFLOAD_page_locked_host_alloc): New prototype.
(GOMP_OFFLOAD_page_locked_host_free): Likewise.
* libgomp.h (gomp_page_locked_host_alloc): Likewise.
(gomp_page_locked_host_free): Likewise.
(struct gomp_device_descr): Add page_locked_host_alloc_func and
page_locked_host_free_func.
* libgomp.texi: Adjust the docs for the pinned trait.
* plugin/plugin-nvptx.c
(GOMP_OFFLOAD_page_locked_host_alloc): New function.
(GOMP_OFFLOAD_page_locked_host_free): Likewise.
* target.c (device_for_page_locked): New variable.
(get_device_for_page_locked): New function.
(gomp_page_locked_host_alloc): Likewise.
(gomp_page_locked_host_free): Likewise.
(gomp_load_plugin_for_device): Add page_locked_host_alloc and
page_locked_host_free.
* testsuite/libgomp.c/alloc-pinned-1.c: Change expectations for NVPTX
devices.
* testsuite/libgomp.c/alloc-pinned-2.c: Likewise.
* testsuite/libgomp.c/alloc-pinned-3.c: Likewise.
* testsuite/libgomp.c/alloc-pinned-4.c: Likewise.
* testsuite/libgomp.c/alloc-pinned-5.c: Likewise.
* testsuite/libgomp.c/alloc-pinned-6.c: Likewise.
Co-Authored-By: Thomas Schwinge <thomas@codesourcery.com>
This patch addresses the incorrectly placed tests, which fail if the
testsuite is ran and gcc has not been installed yet, as discussed
here:
https://gcc.gnu.org/pipermail/gcc-patches/2025-October/698095.html.
gcc/testsuite/ChangeLog:
* gcc.dg/ipa/ipcp-cb-spec1.c: Moved to libgomp/testsuite/libgomp.c/.
* gcc.dg/ipa/ipcp-cb-spec2.c: Likewise.
* gcc.dg/ipa/ipcp-cb1.c: Likewise.
libgomp/ChangeLog:
* testsuite/libgomp.c/ipcp-cb-spec1.c: Moved from
gcc/testsuite/gcc.dg/ipa/.
* testsuite/libgomp.c/ipcp-cb-spec2.c: Likewise.
* testsuite/libgomp.c/ipcp-cb1.c: Likewise.
Signed-off-by: Josef Melcr <jmelcr02@gmail.com>
With commit r16-4212-gf256a13f8aed833fe964a2ba541b7b30ad9b4a76
"c++, gimplify: Implement C++26 P2795R5 - Erroneous behavior for uninitialized reads [PR114457]",
we acquired:
{+FAIL: libgomp.c++/target-flex-101.C (internal compiler error: in assign_temp, at function.cc:990)+}
[-PASS:-]{+FAIL:+} libgomp.c++/target-flex-101.C (test for excess errors)
[-PASS:-]{+UNRESOLVED:+} libgomp.c++/target-flex-101.C [-execution test-]{+compilation failed to produce executable+}
... for GCN, nvptx offloading compilation, and on the other hand:
[-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_map-concurrent.C (internal compiler error[-: in assign_temp, at function.cc:990)-]
[-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_map-concurrent.C (test for excess errors)
[-UNRESOLVED:-]{+PASS:+} libgomp.c++/target-std__flat_map-concurrent.C [-compilation failed to produce executable-]{+execution test+}
[-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_multimap-concurrent.C (internal compiler error[-: in assign_temp, at function.cc:990)-]
[-XFAIL:-]{+XPASS:+} libgomp.c++/target-std__flat_multimap-concurrent.C (test for excess errors)
[-UNRESOLVED:-]{+PASS:+} libgomp.c++/target-std__flat_multimap-concurrent.C [-compilation failed to produce executable-]{+execution test+}
... for GCN offloading compilation (already PASSed for nvptx).
Note that these test cases explicitly use '-std=c++23', so don't undergo the
new C++26 P2795R5 functionality. Yet, comparing before vs. after that commit,
in the 'gimple' dumps (that is, early host compilation), there are a lot of
changes where 'gimple_assign <constructor, [...], {CLOBBER(bob)}, NULL, NULL>'s
and relatedly 'gimple_bind's newly appear/no longer appear elsewhere. This
leads to correspondingly different code at the beginning of offloading
compilation. Why/how that now ('libgomp.c++/target-flex-101.C') vs. before
('libgomp.c++/{target-std__flat_map-concurrent.C,target-std__flat_multimap-concurrent.C}')
translates into 'expand' ICEs, I can't tell.
PR c++/114457
PR c++/122268
PR c++/120450
libgomp/
* testsuite/libgomp.c++/target-flex-101.C: XFAIL GCN, nvptx
offloading compilation.
* testsuite/libgomp.c++/target-std__flat_map-concurrent.C:
Un-XFAIL GCN offloading compilation.
* testsuite/libgomp.c++/target-std__flat_multimap-concurrent.C:
Likewise.
Add missing tests for gfx* context selectors; mark all but the
default-arch declare-variant-4.c with 'dg-excess-errors' to
silence libgomp not-found errors (still passing the
scan-offload-tree-dump check) - or at least causing just
UNRESOLVED errors if the error is
"built without library support ... consider compiling for
the associated generic architecture".
In case the multilib is configured, the result will be
an XPASS.
libgomp/ChangeLog:
* testsuite/libgomp.c/declare-variant-4-gfx10-3-generic.c: Add
dg-excess-errors to handle possible missing libgomp multi lib.
* testsuite/libgomp.c/declare-variant-4-gfx1030.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1036.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx11-generic.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1100.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1103.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx9-4-generic.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx9-generic.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx900.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx906.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx908.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx90a.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx90c.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx942.c: Likewise.
* testsuite/libgomp.c/declare-variant-4-gfx1031.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1032.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1033.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1034.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1035.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1101.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1102.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1150.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1151.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1152.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx1153.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx902.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx904.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx909.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx950.c: New test.
The definition for gfx942 and gfx950 missed the GFX9_4_GENERIC
family flag.
For OpenMP context selectors: The t-omp-device file missed the
generic selectors.
Additionally, there is now a note in the OpenMP documentation that
there is a one-to-one match for ISA names, ignoring any compatibility.
For instance, for Nvidia GPUs 'isa("sm_70")' is only true when compiling
for 'sm_70', even though sm < 7.0 code also runs on sm_70 hardware.
And, for AMD GPUs, gfx9-4-generic neither matches 'gfx942'
(even though such generic code runs on gfx942) - nor the reverse
(although all gfx9-4-generic code runs on gfx942).
gcc/ChangeLog:
* config/gcn/gcn-devices.def (gfx942, gfx950): Set generic name
to GFX9_4_GENERIC.
* config/gcn/t-omp-device: Include generic names for OpenMP's
ISA trait.
libgomp/ChangeLog:
* libgomp.texi (OpenMP Context Selectors): Add note that there is
currently an exact match between ISA and compilation, ignoring
compatibilities in both ways.
* testsuite/libgomp.c/declare-variant-4.h: Add missing variant
functions for specific and generic AMD GPUs.
* testsuite/libgomp.c/declare-variant-4-gfx10-3-generic.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx11-generic.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx9-4-generic.c: New test.
* testsuite/libgomp.c/declare-variant-4-gfx9-generic.c: New test.
The added function is currently '#if 0' but is planned to be used to enable
self mapping automatically. Prerequisite for auto self maps is still mapping
'declare target' variables (if any, in libgomp) or converting all
'declare target' variables to 'declare target link' in the compiler
(as required for 'omp requires self_maps').
include/ChangeLog:
* hsa_ext_amd.h (enum hsa_amd_agent_info_s): Add
HSA_AMD_AGENT_INFO_MEMORY_PROPERTIES.
(enum): Add HSA_AMD_MEMORY_PROPERTY_AGENT_IS_APU.
libgomp/ChangeLog:
* plugin/plugin-gcn.c (is_integrated_apu): New; currently '#if 0'.
* plugin/plugin-nvptx.c (is_integrated_apu): Likewise.
