offtarget 0.1.0__py3-none-any.whl

offload/_root/data/findings/gcc/gcc-acc-link-in-routine-rejected.f90

@@ -0,0 +1,82 @@
+! # Finding (GCC 16.2): a declare link variable used in an OpenACC routine is rejected (Fortran); the reproducer is the catalog
+!   program acc34-clause-link
+!
+! - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+!   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+! - build: gfortran -fopenacc -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90);
+!   -foffload=nvptx-none -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+! - failure: rejected: with .link. clause used in .routine. function
+! - observed: `Error: 'lv' with 'link' clause used in 'routine' function` on both GPUs.
+! - expected: compiles; `RESULT ran_on_host=0 check=pass`
+! - spec: "The global variable or common block variables may be used in accelerator routines" (OpenACC 3.4 §2.13.3 l.3083-3084), for
+!   a variable in a declare link clause; the device copy exists while a data clause holds it (l.3079-3081), here the parallel loop's
+!   copyin(lv).
+! - GCC 16.2 documentation (libgomp.texi "Enabling OpenACC"): "A complete description of all OpenACC directives accepted may be
+!   found in the OpenACC Application Programming Interface manual, version 2.6"; _OPENACC is 201711 (2.6). No list of OpenACC gaps
+!   is documented.
+! - other families: CCE 19/20/21 compile and run it correctly (MI300A, MI250X); NVHPC 25.11/26.9 compile it and give a wrong answer
+!   (not studied here).
+! - gate and confirmer: 2 cells (1 gcc_g90, 1 gcc_sm80), each confirmed in the gate's four environments (a-d;
+!   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+!   compiler_bug (R11-compile-rejection) 2
+! - catalog programs (1): acc34-clause-link
+! - legality: the catalog programs are legal by construction (offload.engine.gen, c_gen and acc_gen
+!   check every program against the catalog rules and the offload scope as it is built; no
+!   compiler). GCC is a compiler under test only: never a legality witness, here or for another
+!   family's case (catalog/machines.yaml under_test_only).
+! Generated by compliance.gen (OpenACC 3.4, free-form Fortran). Do not edit.
+! id: acc34-clause-link
+! primary atom: acc.clause.link
+! legal: yes
+! observable: yes
+! ran_on_host: inside the compute region of interest (acc_on_device)
+! evidence: construct
+! needs: none
+! std: f2008
+module gen_mod
+  use openacc
+  implicit none
+  integer :: lv(64)
+  !$acc declare link(lv)
+  contains
+    integer function get_lv(k)
+      !$acc routine seq
+      integer, intent(in) :: k
+      get_lv = lv(k)
+    end function get_lv
+end module gen_mod
+
+program gen_prog
+  use openacc, only: acc_device_host, acc_on_device
+  use gen_mod
+  implicit none
+  integer, parameter :: n = 64
+  integer, allocatable, target :: a(:), b(:)
+  integer :: ref(n)
+  integer :: i, host_flag
+  logical :: ok
+  allocate(a(n), b(n))
+  do i = 1, n
+    a(i) = i
+    b(i) = 0
+    ref(i) = 2*i + 1
+  end do
+  host_flag = 1
+  do i = 1, n
+    lv(i) = 2*i + 1
+  end do
+  !$acc parallel loop copyin(a) copyout(b) copy(host_flag) copyin(lv)
+  do i = 1, n
+    if (i == 1) host_flag = merge(1, 0, acc_on_device(acc_device_host))
+    b(i) = get_lv(a(i))
+  end do
+  ok = all(b == ref)
+  print '(a,i0,a,a)', 'RESULT ran_on_host=', host_flag, ' check=', merge('pass', 'fail', ok)
+  deallocate(a, b)
+end program gen_prog
+! site:
+! title: OpenACC: a declare link variable used in a routine is rejected ("with 'link' clause used in 'routine' function")
+! summary: A variable in a declare link clause may be used in an accelerator routine (OpenACC 3.4 §2.13.3). GCC 16.2 rejects a
+! + routine seq procedure that reads such a variable: "'lv' with 'link' clause used in 'routine' function", in Fortran and C. CCE
+! + compiles and runs the Fortran program.
+! end site

offload/_root/data/findings/gcc/gcc-acc-use-device-allocatable-rejected.f90

@@ -0,0 +1,88 @@
+! # Finding (GCC 16.2): an allocatable array in use_device is rejected (Fortran); the reproducer is the catalog program
+!   acc34-clause-use_device
+!
+! - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+!   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+! - build: gfortran -fopenacc -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90);
+!   -foffload=nvptx-none -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+! - failure: rejected: ALLOCATABLE object .* in USE_DEVICE clause
+! - observed: `Error: ALLOCATABLE object 'a' in USE_DEVICE clause at (1)` on both GPUs (the host_data and deviceptr programs pass
+!   their allocatable arrays the same way).
+! - expected: compiles; `RESULT ran_on_host=0 check=pass`
+! - spec: "A var in a use_device clause must be the name of a variable or array" (OpenACC 3.4 §2.8 l.2230); the restriction names no
+!   attribute, and an allocatable array is an array.
+! - GCC 16.2 documentation (libgomp.texi "Enabling OpenACC"): "A complete description of all OpenACC directives accepted may be
+!   found in the OpenACC Application Programming Interface manual, version 2.6"; _OPENACC is 201711 (2.6). No list of OpenACC gaps
+!   is documented.
+! - other families: CCE 19/20/21 (MI300A, MI250X) and NVHPC 25.11/26.9 compile and run all three programs correctly.
+! - gate and confirmer: 6 cells (3 gcc_g90, 3 gcc_sm80), each confirmed in the gate's four environments (a-d;
+!   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+!   compiler_bug (R11-compile-rejection) 6
+! - catalog programs (3): acc34-clause-deviceptr, acc34-clause-use_device, acc34-directive-host_data
+! - legality: the catalog programs are legal by construction (offload.engine.gen, c_gen and acc_gen
+!   check every program against the catalog rules and the offload scope as it is built; no
+!   compiler). GCC is a compiler under test only: never a legality witness, here or for another
+!   family's case (catalog/machines.yaml under_test_only).
+! Generated by compliance.gen (OpenACC 3.4, free-form Fortran). Do not edit.
+! id: acc34-clause-use_device
+! primary atom: acc.clause.use_device
+! legal: yes
+! observable: yes
+! ran_on_host: inside the compute region of interest (acc_on_device)
+! evidence: copy_mode_only
+! needs: none
+! std: f2008
+module gen_mod
+  use openacc
+  implicit none
+  contains
+    subroutine dev_twice(x, y, m, hf)
+      integer, intent(in) :: m
+      integer :: x(m), y(m), hf
+      integer :: k
+      !$acc parallel loop deviceptr(x, y) copy(hf)
+      do k = 1, m
+        if (k == 1) hf = merge(1, 0, acc_on_device(acc_device_host))
+        y(k) = 2*x(k) + 1
+      end do
+    end subroutine dev_twice
+end module gen_mod
+
+program gen_prog
+  use openacc, only: acc_device_host, acc_on_device
+  use gen_mod
+  implicit none
+  integer, parameter :: n = 64
+  integer, allocatable, target :: a(:), b(:)
+  integer :: ref(n)
+  integer :: i, host_flag
+  logical :: ok
+  logical :: pres, shared_mem
+  integer :: zc(1)
+  allocate(a(n), b(n))
+  do i = 1, n
+    a(i) = i
+    b(i) = 0
+    ref(i) = 2*i + 1
+  end do
+  host_flag = 1
+  pres = .true.
+  zc = 0
+  !$acc serial copyin(zc)
+    zc(1) = 1
+  !$acc end serial
+  shared_mem = zc(1) == 1
+  !$acc data copyin(a) copyout(b)
+    !$acc host_data use_device(a, b)
+      call dev_twice(a, b, n, host_flag)
+    !$acc end host_data
+  !$acc end data
+  ok = all(b == ref) .and. pres
+  print '(a,i0,a,a)', 'RESULT ran_on_host=', host_flag, ' check=', merge('pass', 'fail', ok)
+  deallocate(a, b)
+end program gen_prog
+! site:
+! title: gfortran: an allocatable array in host_data use_device is rejected ("ALLOCATABLE object ... in USE_DEVICE clause")
+! summary: GCC 16.2's Fortran front end rejects an allocatable array in the use_device clause of host_data ("ALLOCATABLE object 'a'
+! + in USE_DEVICE clause"). OpenACC asks only that the item be a variable or an array. CCE and NVHPC compile and run the programs.
+! end site

offload/_root/data/findings/gcc/gcc-c-motion-mapper-rejected.c

@@ -0,0 +1,110 @@
+// # Finding (GCC 16.2): the mapper modifier of the to and from clauses is rejected (C); the reproducer is the catalog program
+//   omp52-c-clause-to-mapper
+//
+// - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+//   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+// - build: gcc -fopenmp -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90); -foffload=nvptx-none
+//   -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+// - failure: rejected: clause with modifier other than .iterator. or .present.
+// - observed: `error: 'to' clause with modifier other than 'iterator' or 'present'` (and the same for from), on both GPUs. The
+//   declare mapper directive and map(mapper(mm), tofrom: s) of the C map programs compile and pass at 5.0 and 5.2.
+// - expected: compiles; `RESULT ran_on_host=0 check=pass`
+// - GCC 16.2 documentation (libgomp.texi of the GCC 16.2.0 release tarball, "OpenMP Implementation Status", OpenMP 5.0 table):
+//   "declare mapper directive": P, "Initial support and for C/C++, only"; no row names the motion clauses' mapper modifier.
+// - spec: "mapper" is a modifier of the to clause (5.2 §5.9.1 p.166).
+// - other families: AFAR and CCE 19/20/21 compile and run the program correctly (MI300A, both modes; MI250X); NVHPC rejects it.
+// - gate and confirmer: 12 cells (6 gcc_g90, 6 gcc_sm80), each confirmed in the gate's four environments (a-d;
+//   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+//   compiler_bug (R11-compile-rejection) 8, unsupported (R11a-newer-form-unclaimed) 4
+// - catalog programs (6): omp50-c-clause-from-mapper, omp50-c-clause-to-mapper, omp52-c-clause-from-mapper,
+//   omp52-c-clause-to-mapper, omp60-c-clause-from-mapper, omp60-c-clause-to-mapper
+// - legality: the catalog programs are legal by construction (offload.engine.gen, c_gen and acc_gen
+//   check every program against the catalog rules and the offload scope as it is built; no
+//   compiler). GCC is a compiler under test only: never a legality witness, here or for another
+//   family's case (catalog/machines.yaml under_test_only).
+// Generated by compliance.c_gen (OpenMP 5.2, C11): the C port (offload.engine.f2c) of omp52-clause-to-mapper. Do not edit.
+// id: omp52-c-clause-to-mapper
+// primary atom: omp.clause.to.mapper
+// legal: yes
+// observable: yes
+// ran_on_host: inside the target region of interest
+// evidence: copy_mode_only
+// needs: none
+// std: c11
+#include <stdio.h>
+#include <stdbool.h>
+#include <omp.h>
+
+enum { n = 64 };
+struct dt {
+  int v[8];
+  int w[8];
+};
+#pragma omp declare mapper(mm: struct dt q) map(q.v)
+
+int main(void) {
+  int a[n];
+  int b[n];
+  int ref[n];
+  int host_flag;
+  bool ok;
+  struct dt s;
+  int rv;
+  int rw;
+  bool pres;
+  bool shared_mem;
+  int zc[1];
+  for (int i = 1; i <= n; i++) {
+    a[i - 1] = i;
+    b[i - 1] = 0;
+    ref[i - 1] = 2 * i + 1;
+  }
+  host_flag = 1;
+  for (int k1_ = 0; k1_ < 8; k1_++)
+    s.v[k1_] = 1;
+  for (int k2_ = 0; k2_ < 8; k2_++)
+    s.w[k2_] = 1;
+  rv = 0;
+  rw = 0;
+  pres = true;
+  for (int k3_ = 0; k3_ < 1; k3_++)
+    zc[k3_] = 0;
+  #pragma omp target map(to: zc)
+  {
+    zc[0] = 1;
+  }
+  shared_mem = zc[0] == 1;
+  #pragma omp target map(to: a) map(from: b) map(tofrom: host_flag)
+  {
+    host_flag = (omp_is_initial_device() ? 1 : 0);
+    for (int i = 1; i <= n; i++) {
+      b[i - 1] = 2 * a[i - 1] + 1;
+    }
+  }
+  #pragma omp target enter data map(to: s)
+  for (int k4_ = 0; k4_ < 8; k4_++)
+    s.v[k4_] = 5;
+  for (int k5_ = 0; k5_ < 8; k5_++)
+    s.w[k5_] = 5;
+  #pragma omp target update to(mapper(mm): s)
+  #pragma omp target map(from: rv, rw)
+  {
+    rv = s.v[0];
+    rw = s.w[0];
+  }
+  #pragma omp target exit data map(delete: s)
+  if (!shared_mem) pres = pres && ((rv == 5) && (rw == 1));
+  bool t6_ = true;
+  for (int k7_ = 0; k7_ < n; k7_++)
+    t6_ = t6_ && (b[k7_] == ref[k7_]);
+  ok = t6_ && pres;
+  printf("RESULT ran_on_host=%d check=%s\n", host_flag, ok ? "pass" : "fail");
+  return 0;
+}
+// site:
+// title: target update to(mapper(m): x) and from(mapper(m): x) are rejected: "clause with modifier other than 'iterator' or
+// + 'present'"
+// summary: GCC 16.2 accepts declare mapper and the mapper modifier of map in C, but rejects the same modifier on the to and from
+// + clauses of target update ("'to' clause with modifier other than 'iterator' or 'present'"). AFAR and CCE compile and run the
+// + program correctly.
+// end site

offload/_root/data/findings/gcc/gcc-c-omp60-undocumented-rejected.c

@@ -0,0 +1,76 @@
+// # Finding (GCC 16.2): requires device_safesync and defaultmap(private) are rejected, and GCC's table does not list them (C); the
+//   reproducer is the catalog program omp60-c-clause-device_safesync
+//
+// - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+//   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+// - build: gcc -fopenmp -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90); -foffload=nvptx-none
+//   -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+// - failure: rejected
+// - observed: `error: expected 'unified_address', 'unified_shared_memory', 'self_maps', 'dynamic_allocators', ...` at
+//   device_safesync; `error: expected 'alloc', 'to', 'from', 'tofrom', 'firstprivate', 'none' or 'default' before 'private'` at
+//   defaultmap(private); both GPUs.
+// - expected: compiles; `RESULT ran_on_host=0 check=pass`
+// - GCC 16.2 documentation (libgomp.texi of the GCC 16.2.0 release tarball, "OpenMP Implementation Status", OpenMP 6.0 table): no
+//   row names the device_safesync requirement clause or private as a defaultmap implicit-behavior (the table's defaultmap row is
+//   "self modifier to map and self as defaultmap argument"). Undocumented.
+// - spec: the device_safesync clause is a 6.0 requirement clause (6.0 §5.4 p.173 lists it among the clauses that take a
+//   directive-name modifier); defaultmap's implicit-behavior private is new in 6.0.
+// - other families (catalog matrix): requires device_safesync: CCE 19/20/21 and NVHPC reject it; AFAR amdflang compiles and runs
+//   the Fortran program, AFAR amdclang's C compile does not finish (afar-clang-requires-device-safesync-hang.c).
+//   defaultmap(private): AFAR amdclang and CCE 21 run the C program correctly, CCE 20 the Fortran one; CCE 19 and NVHPC reject it.
+// - gate and confirmer: 8 cells (4 gcc_g90, 4 gcc_sm80), each confirmed in the gate's four environments (a-d;
+//   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+//   unsupported (R11a-newer-form-unclaimed) 8
+// - catalog programs (4): omp60-c-clause-defaultmap-private, omp60-c-clause-device_safesync,
+//   omp60-c-clause-device_safesync-directive_name, omp60-c-clause-device_safesync-required
+// - legality: the catalog programs are legal by construction (offload.engine.gen, c_gen and acc_gen
+//   check every program against the catalog rules and the offload scope as it is built; no
+//   compiler). GCC is a compiler under test only: never a legality witness, here or for another
+//   family's case (catalog/machines.yaml under_test_only).
+// Generated by compliance.c_gen (OpenMP 6.0, C11): the C port (offload.engine.f2c) of omp60-clause-device_safesync. Do not edit.
+// id: omp60-c-clause-device_safesync
+// primary atom: omp.clause.device_safesync
+// legal: yes
+// observable: no (requires only changes semantics: no checked device effect)
+// ran_on_host: inside the target region of interest
+// evidence: construct
+// needs: none
+// std: c11
+#include <stdio.h>
+#include <stdbool.h>
+#include <omp.h>
+
+enum { n = 64 };
+#pragma omp requires device_safesync
+
+int main(void) {
+  int a[n];
+  int b[n];
+  int ref[n];
+  int host_flag;
+  bool ok;
+  for (int i = 1; i <= n; i++) {
+    a[i - 1] = i;
+    b[i - 1] = 0;
+    ref[i - 1] = 2 * i + 1;
+  }
+  host_flag = 1;
+  #pragma omp target teams distribute parallel for map(to: a) map(from: b) \
+      map(tofrom: host_flag) default(none) shared(a, b, host_flag)
+  for (int i = 1; i <= n; i++) {
+    b[i - 1] = 2 * a[i - 1] + 1;
+    if (i == 1) host_flag = (omp_is_initial_device() ? 1 : 0);
+  }
+  bool t1_ = true;
+  for (int k2_ = 0; k2_ < n; k2_++)
+    t1_ = t1_ && (b[k2_] == ref[k2_]);
+  ok = t1_;
+  printf("RESULT ran_on_host=%d check=%s\n", host_flag, ok ? "pass" : "fail");
+  return 0;
+}
+// site:
+// title: OpenMP 6.0 requires device_safesync and defaultmap(private) are rejected (not in GCC's implementation status table)
+// summary: GCC 16.2 rejects two OpenMP 6.0 forms that no row of its implementation status table covers: the device_safesync clause
+// + of requires ("expected 'unified_address', 'unified_shared_memory', ...") and defaultmap(private) ("expected 'alloc', 'to', ...
+// + before 'private'", the 5.2 list). The other 6.0 rejections are in the finding for the features GCC lists as not implemented.
+// end site

offload/_root/data/findings/gcc/gcc-c-omp60-unimplemented-rejected.c

@@ -0,0 +1,119 @@
+// # Finding (GCC 16.2): OpenMP 6.0 forms GCC documents as not implemented are rejected (C); the reproducer is the catalog program
+//   omp60-c-clause-induction
+//
+// - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+//   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+// - build: gcc -fopenmp -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90); -foffload=nvptx-none
+//   -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+// - failure: rejected: expected an OpenMP clause before
+// - observed: a compile-time error for every program of the list (the first: `error: expected an OpenMP clause before 'induction'`;
+//   nogroup on target_data: `'nogroup' is not valid for '#pragma omp target data'`; the ignored declare_mapper: `mapper 'mm' not
+//   found`); the same on both GPUs.
+// - expected: compiles; `RESULT ran_on_host=0 check=pass`
+// - GCC 16.2 documentation (libgomp.texi of the GCC 16.2.0 release tarball, "OpenMP Implementation Status", OpenMP 6.0 table) lists
+//   each of these as not implemented (N): "directive-name-modifier accepted in all clauses", "Support for inductions", "local
+//   clause to declare_target directive", "New taskgraph construct including saved modifier and replayable clause", "New transparent
+//   clause for multi-generational task-dependence graphs", "strict modifier keyword to num_threads", "default clause on the target
+//   directive and accepting variable categories", "New storage map-type modifier", "self modifier to map and self as defaultmap
+//   argument", "ref modifier to the map clause", "Reduction over private variables with reduction clause" (the original modifier),
+//   "Boolean argument to nowait and nogroup may be non constant", "target_data as composite construct" (nogroup on target_data),
+//   "All inarguable clauses take now an optional Boolean argument" (the requires clauses' argument), "For Fortran, the automap
+//   modifier to the enter clause", "workdistribute directive for Fortran" and "Multi-word directives now use underscore by default"
+//   (declare_simd, declare_mapper, target_loop). The two 6.0 forms no row covers (requires device_safesync, defaultmap(private))
+//   are a finding of their own (gcc-c-omp60-undocumented-rejected).
+// - spec: the induction clause (6.0 §7.6.13 p.257), the directive-name-modifier (6.0 §5.4 p.173), the replayable clause (6.0 §14.6
+//   p.440), the transparent clause (6.0 §17.9.6 p.510).
+// - other families: AFAR, CCE and NVHPC reject many of the same 6.0 forms (their own findings and the catalog's unsupported cells).
+// - gate and confirmer: 168 cells (84 gcc_g90, 84 gcc_sm80), each confirmed in the gate's four environments (a-d;
+//   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+//   unsupported (R11a-newer-form-unclaimed) 168
+// - catalog programs (84): omp60-c-clause-aligned-directive_name, omp60-c-clause-allocate-directive_name,
+//   omp60-c-clause-atomic_default_mem_order-directive_name, omp60-c-clause-bind-directive_name,
+//   omp60-c-clause-collapse-directive_name, omp60-c-clause-default-aggregate, omp60-c-clause-default-all,
+//   omp60-c-clause-default-directive_name, omp60-c-clause-default-pointer, omp60-c-clause-default-scalar,
+//   omp60-c-clause-defaultmap-directive_name, omp60-c-clause-defaultmap-self, omp60-c-clause-defaultmap-storage,
+//   omp60-c-clause-depend-directive_name, omp60-c-clause-device-directive_name, omp60-c-clause-device_type-directive_name,
+//   omp60-c-clause-dist_schedule-directive_name, omp60-c-clause-dynamic_allocators-directive_name,
+//   omp60-c-clause-dynamic_allocators-required, omp60-c-clause-enter-directive_name, omp60-c-clause-firstprivate-directive_name,
+//   omp60-c-clause-firstprivate-saved, omp60-c-clause-from-directive_name, omp60-c-clause-has_device_addr-directive_name,
+//   omp60-c-clause-in_reduction-directive_name, omp60-c-clause-indirect-directive_name, omp60-c-clause-induction,
+//   omp60-c-clause-induction-directive_name, omp60-c-clause-induction-plus, omp60-c-clause-induction-relaxed,
+//   omp60-c-clause-induction-step, omp60-c-clause-induction-strict, omp60-c-clause-induction-times,
+//   omp60-c-clause-is_device_ptr-directive_name, omp60-c-clause-lastprivate-directive_name, omp60-c-clause-linear-directive_name,
+//   omp60-c-clause-link-directive_name, omp60-c-clause-local, omp60-c-clause-local-directive_name,
+//   omp60-c-clause-map-directive_name, omp60-c-clause-map-mapper, omp60-c-clause-map-self, omp60-c-clause-map-storage,
+//   omp60-c-clause-message-directive_name, omp60-c-clause-nogroup, omp60-c-clause-nogroup-directive_name,
+//   omp60-c-clause-nogroup-do_not_synchronize, omp60-c-clause-nontemporal-directive_name, omp60-c-clause-nowait-directive_name,
+//   omp60-c-clause-nowait-do_not_synchronize, omp60-c-clause-num_teams-directive_name, omp60-c-clause-num_threads-directive_name,
+//   omp60-c-clause-num_threads-strict, omp60-c-clause-order-directive_name, omp60-c-clause-ordered-directive_name,
+//   omp60-c-clause-private-directive_name, omp60-c-clause-proc_bind-directive_name, omp60-c-clause-reduction-directive_name,
+//   omp60-c-clause-reduction-original, omp60-c-clause-replayable, omp60-c-clause-replayable-directive_name,
+//   omp60-c-clause-replayable-replayable_expression, omp60-c-clause-reverse_offload-directive_name,
+//   omp60-c-clause-reverse_offload-required, omp60-c-clause-safelen-directive_name, omp60-c-clause-schedule-directive_name,
+//   omp60-c-clause-self_maps-directive_name, omp60-c-clause-self_maps-required, omp60-c-clause-severity-directive_name,
+//   omp60-c-clause-shared-directive_name, omp60-c-clause-simdlen-directive_name, omp60-c-clause-thread_limit-directive_name,
+//   omp60-c-clause-to-directive_name, omp60-c-clause-transparent, omp60-c-clause-transparent-directive_name,
+//   omp60-c-clause-transparent-impex_type, omp60-c-clause-unified_address-directive_name, omp60-c-clause-unified_address-required,
+//   omp60-c-clause-unified_shared_memory-directive_name, omp60-c-clause-unified_shared_memory-required,
+//   omp60-c-clause-use_device_addr-directive_name, omp60-c-clause-use_device_ptr-directive_name,
+//   omp60-c-clause-uses_allocators-directive_name, omp60-c-directive-target_loop
+// - legality: the catalog programs are legal by construction (offload.engine.gen, c_gen and acc_gen
+//   check every program against the catalog rules and the offload scope as it is built; no
+//   compiler). GCC is a compiler under test only: never a legality witness, here or for another
+//   family's case (catalog/machines.yaml under_test_only).
+// Generated by compliance.c_gen (OpenMP 6.0, C11): the C port (offload.engine.f2c) of omp60-clause-induction. Do not edit.
+// id: omp60-c-clause-induction
+// primary atom: omp.clause.induction
+// legal: yes
+// observable: yes
+// ran_on_host: inside the target region of interest
+// evidence: construct
+// needs: none
+// std: c11
+#include <stdio.h>
+#include <stdbool.h>
+#include <omp.h>
+
+enum { n = 64 };
+
+int main(void) {
+  int a[n];
+  int b[n];
+  int ref[n];
+  int host_flag;
+  bool ok;
+  int x;
+  for (int i = 1; i <= n; i++) {
+    a[i - 1] = i;
+    b[i - 1] = 0;
+    ref[i - 1] = 2 * i + 1;
+  }
+  host_flag = 1;
+  x = 3;
+  #pragma omp target map(to: a) map(from: b) map(tofrom: x) map(tofrom: host_flag)
+  {
+    host_flag = (omp_is_initial_device() ? 1 : 0);
+    #pragma omp parallel for default(none) shared(a, b) induction(step(1), +: x)
+    for (int i = 1; i <= n; i++) {
+      b[i - 1] = 2 * a[i - 1] + 1 + (x - (i + 2));
+      x = x + 1;
+    }
+  }
+  bool t1_ = true;
+  for (int k2_ = 0; k2_ < n; k2_++)
+    t1_ = t1_ && (b[k2_] == ref[k2_]);
+  ok = t1_ && (x == n + 3);
+  printf("RESULT ran_on_host=%d check=%s\n", host_flag, ok ? "pass" : "fail");
+  return 0;
+}
+// site:
+// title: OpenMP 6.0 features that GCC 16.2 lists as not implemented are rejected (directive-name modifiers, induction, taskgraph
+// + clauses, map and defaultmap self/storage, and more)
+// summary: GCC 16.2 rejects OpenMP 6.0 forms that its implementation status table lists as not implemented: directive-name
+// + modifiers on clauses, the induction, local, replayable and transparent clauses, num_threads(strict:), variable categories on
+// + default, defaultmap(self/storage), map(self:/storage:/ref_*), firstprivate(saved:), reduction(original:), arguments on requires
+// + clauses and nowait, nogroup on target_data, and the underscore spellings of multi-word directives. Most stop with a syntax
+// + error; declare target local stops with "Sorry, ... not implemented", nogroup with "'nogroup' is not valid for '#pragma omp
+// + target data'", and the ignored declare_mapper leaves its mapper "not found". GCC 16.2 claims OpenMP 5.2 (_OPENMP 202111) and
+// + parts of 6.0.
+// end site

offload/_root/data/findings/gcc/gcc-c-target-in-task-mandatory-device-not-found.c

@@ -0,0 +1,87 @@
+// # Finding (GCC 16.2): under OMP_TARGET_OFFLOAD=mandatory a target region in an explicit task stops with "device not found"
+//
+// - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+//   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+// - build: gcc -fopenmp -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90); -foffload=nvptx-none
+//   -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+// - failure: run_error: OMP_TARGET_OFFLOAD is set to MANDATORY, but device not found
+// - observed: `libgomp: OMP_TARGET_OFFLOAD is set to MANDATORY, but device not found`, exit 1, no kernel launch, in every run on
+//   both GPUs (this file: gcc_sm80 5 of 5, gcc_g90 5 of 5; results/gcc/<tc>/probes-findings.jsonl). With OMP_TARGET_OFFLOAD unset
+//   the same build runs the kernel on the device and prints `bad= 0` (ran_on_host=0) in 5 of 5 runs on each GPU.
+// - expected: `bad= 0`, then `RESULT ran_on_host=0 check=pass`
+// - rule: a target construct without a device clause uses the default device: "if no device clause is present, the behavior is as
+//   if the device clause appears without a device-modifier and with a device-description that evaluates to the value of the
+//   default-device-var ICV" (5.2 §13.2 p.276 l.21-23), and an explicit task has the encountering task's value of it: "When a task
+//   construct, a parallel construct or a teams construct is encountered, each generated task inherits the values of the data
+//   environment scoped ICVs from each generating task's ICV values" (5.2 §2.4 p.45 l.10-12). Under mandatory only an unavailable
+//   device is an error: "The mandatory value specifies that the effect of any device construct or device memory routine that uses a
+//   device that is unavailable or not supported by the implementation, or uses a non-conforming device number, is as if the
+//   omp_invalid_device device number was used" (5.2 §21.2.8 p.610 l.28-30). The device is available: the same region in the initial
+//   task runs on it (the control).
+// - same-compiler controls (correct, 5 of 5 runs on each GPU): the same region in the initial task, without the task construct
+//   (controls/gcc-target-in-task-control.{c,f90}). The GCC review's probes (tests/data/gcc_probes/inred_*.c,
+//   results/gcc/<tc>/probes.jsonl, both GPUs) narrow it: the region fails the same way in an explicit task, in a parallel region
+//   (single) and in a taskgroup with task_reduction (with in_reduction or with a plain map); in a plain taskgroup or with no
+//   taskgroup it runs. The catalog's in_reduction programs fail through it: their target sits in a task_reduction taskgroup.
+// - GCC 16.2 documentation (libgomp manual): OMP_TARGET_OFFLOAD "If set to 'MANDATORY', the program terminates with an error if any
+//   device construct or device memory routine uses a device that is unavailable or not supported by the implementation, or uses a
+//   non-conforming device number" (the same condition as 5.2 §21.2.8); the OpenMP 5.0 table lists "in_reduction clause on target
+//   constructs": P, "nowait only stub". The catalog's in_reduction programs use no nowait and pass with OMP_TARGET_OFFLOAD unset on
+//   both GPUs, so their failure is this finding's, not the stub's.
+// - other families: AFAR and CCE 19/20 run the catalog's in_reduction programs on the device and pass (MI300A, both modes; MI250X);
+//   CCE 21 fails to link them (cce21-in-reduction-taskwait-link.f90); NVHPC rejects in_reduction on target. The site's finding
+//   matrix runs this file on every compiler.
+// - gate and confirmer: 64 cells (32 gcc_g90, 32 gcc_sm80), each confirmed in the gate's four environments (a-d;
+//   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+//   compiler_bug (R20-runtime-bug) 22, needs_human (R22-expected-weak) 42
+// - catalog programs (32): omp50-c-clause-in_reduction, omp50-c-clause-in_reduction-and, omp50-c-clause-in_reduction-iand,
+//   omp50-c-clause-in_reduction-ieor, omp50-c-clause-in_reduction-ior, omp50-c-clause-in_reduction-max,
+//   omp50-c-clause-in_reduction-min, omp50-c-clause-in_reduction-minus, omp50-c-clause-in_reduction-or,
+//   omp50-c-clause-in_reduction-plus, omp50-c-clause-in_reduction-times, omp52-c-clause-in_reduction,
+//   omp52-c-clause-in_reduction-and, omp52-c-clause-in_reduction-iand, omp52-c-clause-in_reduction-ieor,
+//   omp52-c-clause-in_reduction-ior, omp52-c-clause-in_reduction-max, omp52-c-clause-in_reduction-min,
+//   omp52-c-clause-in_reduction-minus, omp52-c-clause-in_reduction-or, omp52-c-clause-in_reduction-plus,
+//   omp52-c-clause-in_reduction-times, omp60-c-clause-in_reduction, omp60-c-clause-in_reduction-and,
+//   omp60-c-clause-in_reduction-iand, omp60-c-clause-in_reduction-ieor, omp60-c-clause-in_reduction-ior,
+//   omp60-c-clause-in_reduction-max, omp60-c-clause-in_reduction-min, omp60-c-clause-in_reduction-or,
+//   omp60-c-clause-in_reduction-plus, omp60-c-clause-in_reduction-times
+// - legality: read against the spec (the rule above); the catalog programs of the list are legal by construction (offload.engine.gen,
+//   c_gen and acc_gen; no compiler). GCC is a compiler under test only and never a legality witness.
+#include <stdio.h>
+#include <omp.h>
+
+enum { n = 64 };
+
+int main(void) {
+  int a[n], b[n];
+  int host_flag = 1, bad = 0;
+  for (int i = 0; i < n; i++) {
+    a[i] = i + 1;
+    b[i] = 0;
+  }
+  #pragma omp task shared(a, b, host_flag)
+  {
+    #pragma omp target map(to: a) map(from: b) map(tofrom: host_flag)
+    {
+      host_flag = omp_is_initial_device();
+      for (int i = 0; i < n; i++)
+        b[i] = 2 * a[i] + 1;
+    }
+  }
+  #pragma omp taskwait
+  for (int i = 0; i < n; i++)
+    bad += b[i] != 2 * (i + 1) + 1;
+  printf("bad= %d\n", bad);
+  printf("RESULT ran_on_host=%d check=%s\n", host_flag, bad == 0 ? "pass" : "fail");
+  return 0;
+}
+// site:
+// title: libgomp: with OMP_TARGET_OFFLOAD=mandatory, a target region in an explicit task, a parallel region or a task_reduction
+// + taskgroup stops with "device not found"
+// summary: With OMP_TARGET_OFFLOAD=mandatory, a target region that GCC 16.2's libgomp meets in an explicit task stops the program
+// + with "OMP_TARGET_OFFLOAD is set to MANDATORY, but device not found", although the same region in the initial task runs on the
+// + GPU. The same happens in a parallel region and in a taskgroup with task_reduction, so every target construct with in_reduction
+// + fails (70 catalog programs per GPU, Fortran and C). With OMP_TARGET_OFFLOAD unset the region runs on the device and gives the
+// + right answer. Seen on the MI250X and the A100.
+// control: gcc/controls/gcc-target-in-task-control.c the same region in the initial task (no task construct)
+// end site

offload/_root/data/findings/gcc/gcc-c-uses-allocators-traits-unimplemented.c

@@ -0,0 +1,85 @@
+// # Finding (GCC 16.2): uses_allocators with traits or a memory space is not implemented (C); the reproducer is the catalog program
+//   omp52-c-clause-uses_allocators-traits
+//
+// - toolchains: gcc_g90 (GCC 16.2.0, AMD Instinct MI250X gfx90a, Tioga) and gcc_sm80 (GCC 16.2.0,
+//   NVIDIA A100 sm_80, wingtip); separate memory only; -O2; OMP_TARGET_OFFLOAD=mandatory
+// - build: gcc -fopenmp -foffload=amdgcn-amdhsa -foffload-options=amdgcn-amdhsa=-march=gfx90a -O2 (gcc_g90); -foffload=nvptx-none
+//   -foffload-options=nvptx-none=-march=sm_80 -foffload-options=nvptx-none=-latomic -O2 (gcc_sm80)
+// - failure: rejected: sorry, unimplemented: .uses_allocators. clause
+// - observed: `sorry, unimplemented: 'uses_allocators' clause` on both GPUs.
+// - expected: compiles; `RESULT ran_on_host=0 check=pass`
+// - GCC 16.2 documentation (libgomp.texi of the GCC 16.2.0 release tarball, "OpenMP Implementation Status", OpenMP 5.2 table): "New
+//   memspace and traits modifiers for uses_allocators": P (partial), "Only predefined allocators". The message is GCC's own "sorry,
+//   unimplemented".
+// - spec: the uses_allocators clause and its traits and memspace modifiers (5.2 §6.8 p.181).
+// - other families: AFAR reports uses_allocators as not yet implemented; CCE rejects the 5.2 traits form
+//   (catalog/cce-c-uses-allocators-traits-rejected.c).
+// - gate and confirmer: 10 cells (5 gcc_g90, 5 gcc_sm80), each confirmed in the gate's four environments (a-d;
+//   results/gcc/gate-tioga/ and gate-wingtip/ verdicts.jsonl); confirmer verdicts (results/gcc/confirm/batch.gcc_*.jsonl):
+//   unsupported (R05-unsupported) 10
+// - catalog programs (5): omp50-c-clause-uses_allocators-traits, omp52-c-clause-uses_allocators-memspace,
+//   omp52-c-clause-uses_allocators-traits, omp60-c-clause-uses_allocators-memspace, omp60-c-clause-uses_allocators-traits
+// - legality: the catalog programs are legal by construction (offload.engine.gen, c_gen and acc_gen
+//   check every program against the catalog rules and the offload scope as it is built; no
+//   compiler). GCC is a compiler under test only: never a legality witness, here or for another
+//   family's case (catalog/machines.yaml under_test_only).
+// Generated by compliance.c_gen (OpenMP 5.2, C11): the C port (offload.engine.f2c) of omp52-clause-uses_allocators-traits. Do not edit.
+// id: omp52-c-clause-uses_allocators-traits
+// primary atom: omp.clause.uses_allocators.traits
+// legal: yes
+// observable: no (unverified: the check covers the device kernel, not the atom)
+// ran_on_host: inside the target region of interest
+// evidence: unverified
+// needs: none
+// std: c11
+// unverified (oracle_not_written): re-categorised (was implementation_defined): memory from the allocator 'will be byte
+// aligned to at least the value specified for the alignment trait' (5.2 §6.2 p.173, 6.0 §8.2 p.306), so a check that t's address is
+// a multiple of 64 is a deterministic oracle of the original; it is not written. A mutant without the trait could still be 64-byte
+// aligned (any alignment conforms then), AFAR rejects uses_allocators and CCE 19's mutant without the trait is an internal compiler
+// error (final-review re-audit)
+#include <stdio.h>
+#include <stdbool.h>
+#include <omp.h>
+
+enum { n = 64 };
+
+int main(void) {
+  int a[n];
+  int b[n];
+  int ref[n];
+  int host_flag;
+  bool ok;
+  int t;
+  omp_allocator_handle_t my_alloc;
+  const const omp_alloctrait_t my_traits[1] = {{omp_atk_alignment, 64}};
+  for (int i = 1; i <= n; i++) {
+    a[i - 1] = i;
+    b[i - 1] = 0;
+    ref[i - 1] = 2 * i + 1;
+  }
+  host_flag = 1;
+  #pragma omp target map(to: a) map(from: b) map(tofrom: host_flag) \
+      uses_allocators(traits(my_traits): my_alloc)
+  {
+    host_flag = (omp_is_initial_device() ? 1 : 0);
+    #pragma omp parallel for default(none) shared(a, b, my_alloc) private(t) \
+        allocate(my_alloc: t)
+    for (int i = 1; i <= n; i++) {
+      t = 2 * a[i - 1];
+      b[i - 1] = t + 1;
+    }
+  }
+  bool t1_ = true;
+  for (int k2_ = 0; k2_ < n; k2_++)
+    t1_ = t1_ && (b[k2_] == ref[k2_]);
+  ok = t1_;
+  printf("RESULT ran_on_host=%d check=%s\n", host_flag, ok ? "pass" : "fail");
+  return 0;
+}
+// site:
+// title: uses_allocators with allocator traits or a memory space: "sorry, unimplemented"
+// summary: GCC 16.2 stops with "sorry, unimplemented: 'uses_allocators' clause with traits and memory spaces" (Fortran) or "sorry,
+// + unimplemented: 'uses_allocators' clause" (C) on a target construct whose uses_allocators clause defines an allocator with
+// + traits or a memory space, in the 5.0 and 5.2 forms. Its implementation status table says only predefined allocators are
+// + supported.
+// end site