[MLIR][OpenMP] Add OpenMPToLLVMIRTranslation support for is_device_ptr

flang/lib/Lower/OpenMP/ClauseProcessor.cpp

@@ -1145,8 +1145,11 @@ bool ClauseProcessor::processIsDevicePtr(
  [&](const omp::clause::IsDevicePtr &clause,
  const parser::CharBlock &source) {
  mlir::Location location = converter.genLocation(source);
+ // Force a map so the descriptor is materialized on the device with the
+ // device address inside.
  mlir::omp::ClauseMapFlags mapTypeBits =
- mlir::omp::ClauseMapFlags::is_device_ptr;
+ mlir::omp::ClauseMapFlags::is_device_ptr |
+ mlir::omp::ClauseMapFlags::to;
  processMapObjects(stmtCtx, location, clause.v, mapTypeBits,
  parentMemberIndices, result.isDevicePtrVars,
  isDeviceSyms);

flang/test/Integration/OpenMP/map-types-and-sizes.f90

@@ -33,6 +33,15 @@ subroutine mapType_array
  !$omp end target
 end subroutine mapType_array
 
+!CHECK: @.offload_sizes{{.*}} = private unnamed_addr constant [1 x i64] [i64 8]
+!CHECK: @.offload_maptypes{{.*}} = private unnamed_addr constant [1 x i64] [i64 33]
+subroutine mapType_is_device_ptr
+ use iso_c_binding, only : c_ptr
+ type(c_ptr) :: p
+ !$omp target is_device_ptr(p)
+ !$omp end target
+end subroutine mapType_is_device_ptr
+
 !CHECK: @.offload_sizes{{.*}} = private unnamed_addr constant [4 x i64] [i64 0, i64 24, i64 8, i64 0]
 !CHECK: @.offload_maptypes{{.*}} = private unnamed_addr constant [4 x i64] [i64 32, i64 281474976711169, i64 281474976711171, i64 281474976711187]
 subroutine mapType_ptr

mlir/lib/Target/LLVMIR/Dialect/OpenMP/OpenMPToLLVMIRTranslation.cpp

@@ -332,10 +332,6 @@ static LogicalResult checkImplementationStatus(Operation &op) {
  op.getInReductionSyms())
  result = todo("in_reduction");
  };
- auto checkIsDevicePtr = [&todo](auto op, LogicalResult &result) {
- if (!op.getIsDevicePtrVars().empty())
- result = todo("is_device_ptr");
- };
  auto checkLinear = [&todo](auto op, LogicalResult &result) {
  if (!op.getLinearVars().empty() || !op.getLinearStepVars().empty())
  result = todo("linear");
@@ -444,7 +440,6 @@ static LogicalResult checkImplementationStatus(Operation &op) {
  checkBare(op, result);
  checkDevice(op, result);
  checkInReduction(op, result);
- checkIsDevicePtr(op, result);
  })
  .Default([](Operation &) {
  // Assume all clauses for an operation can be translated unless they are
@@ -3835,6 +3830,9 @@ convertClauseMapFlags(omp::ClauseMapFlags mlirFlags) {
  auto mapTypeToBool = [&mlirFlags](omp::ClauseMapFlags flag) {
  return (mlirFlags & flag) == flag;
  };
+ const bool hasExplicitMap =
+ (mlirFlags & ~omp::ClauseMapFlags::is_device_ptr) !=
+ omp::ClauseMapFlags::none;
 
  llvm::omp::OpenMPOffloadMappingFlags mapType =
  llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_NONE;
@@ -3875,6 +3873,12 @@ convertClauseMapFlags(omp::ClauseMapFlags mlirFlags) {
  if (mapTypeToBool(omp::ClauseMapFlags::attach))
  mapType |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_ATTACH;
 
+ if (mapTypeToBool(omp::ClauseMapFlags::is_device_ptr)) {
+ mapType |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_TARGET_PARAM;
+ if (!hasExplicitMap)
+ mapType |= llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_LITERAL;
+ }
+
  return mapType;
 }
 
@@ -3996,6 +4000,9 @@ static void collectMapDataFromMapOperands(
  llvm::Value *origValue = moduleTranslation.lookupValue(offloadPtr);
  auto mapType = convertClauseMapFlags(mapOp.getMapType());
  auto mapTypeAlways = llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_ALWAYS;
+ bool isDevicePtr =
+ (mapOp.getMapType() & omp::ClauseMapFlags::is_device_ptr) !=
+ omp::ClauseMapFlags::none;
 
  mapData.OriginalValue.push_back(origValue);
  mapData.BasePointers.push_back(origValue);
@@ -4022,14 +4029,18 @@ static void collectMapDataFromMapOperands(
  mapData.Mappers.push_back(nullptr);
  }
  } else {
+ // For is_device_ptr we need the map type to propagate so the runtime
+ // can materialize the device-side copy of the pointer container.
  mapData.Types.push_back(
- llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_LITERAL);
+ isDevicePtr ? mapType
+ : llvm::omp::OpenMPOffloadMappingFlags::OMP_MAP_LITERAL);
  mapData.Mappers.push_back(nullptr);
  }
  mapData.Names.push_back(LLVM::createMappingInformation(
  mapOp.getLoc(), *moduleTranslation.getOpenMPBuilder()));
  mapData.DevicePointers.push_back(
- llvm::OpenMPIRBuilder::DeviceInfoTy::Address);
+ isDevicePtr ? llvm::OpenMPIRBuilder::DeviceInfoTy::Pointer
+ : llvm::OpenMPIRBuilder::DeviceInfoTy::Address);
  mapData.IsAMapping.push_back(false);
  mapData.IsAMember.push_back(checkIsAMember(hasDevAddrOperands, mapOp));
  }

mlir/test/Target/LLVMIR/omptarget-llvm.mlir

@@ -622,3 +622,20 @@ module attributes {omp.target_triples = ["amdgcn-amd-amdhsa"]} {
 // CHECK: br label %[[VAL_40]]
 // CHECK: omp.done: ; preds = %[[VAL_68]], %[[VAL_63]], %[[VAL_32]]
 // CHECK: ret void
+
+// -----
+
+module attributes {omp.target_triples = ["amdgcn-amd-amdhsa"]} {
+ llvm.func @_QPomp_target_is_device_ptr(%arg0 : !llvm.ptr) {
+ %map = omp.map.info var_ptr(%arg0 : !llvm.ptr, !llvm.ptr)
+ map_clauses(is_device_ptr) capture(ByRef) -> !llvm.ptr {name = ""}
+ omp.target map_entries(%map -> %ptr_arg : !llvm.ptr) {
+ omp.terminator
+ }
+ llvm.return
+ }
+}
+
+// CHECK: @.offload_sizes = private unnamed_addr constant [1 x i64] [i64 8]
+// CHECK: @.offload_maptypes = private unnamed_addr constant [1 x i64] [i64 288]
+// CHECK-LABEL: define void @_QPomp_target_is_device_ptr

mlir/test/Target/LLVMIR/openmp-todo.mlir

@@ -238,17 +238,6 @@ llvm.func @target_in_reduction(%x : !llvm.ptr) {
 
 // -----
 
-llvm.func @target_is_device_ptr(%x : !llvm.ptr) {
- // expected-error@below {{not yet implemented: Unhandled clause is_device_ptr in omp.target operation}}
- // expected-error@below {{LLVM Translation failed for operation: omp.target}}
- omp.target is_device_ptr(%x : !llvm.ptr) {
- omp.terminator
- }
- llvm.return
-}
-
-// -----
-
 llvm.func @target_enter_data_depend(%x: !llvm.ptr) {
  // expected-error@below {{not yet implemented: Unhandled clause depend in omp.target_enter_data operation}}
  // expected-error@below {{LLVM Translation failed for operation: omp.target_enter_data}}

offload/test/offloading/fortran/target-is-device-ptr.f90

@@ -0,0 +1,60 @@
+! Validate that a device pointer obtained via omp_get_mapped_ptr can be used
+! inside a TARGET region with the is_device_ptr clause.
+! REQUIRES: flang, amdgcn-amd-amdhsa
+
+! RUN: %libomptarget-compile-fortran-run-and-check-generic
+
+module mod
+ implicit none
+ integer, parameter :: n = 4
+contains
+ subroutine kernel(dptr)
+ use iso_c_binding, only : c_ptr, c_f_pointer
+ implicit none
+
+ type(c_ptr) :: dptr
+ integer, dimension(:), pointer :: b
+ integer :: i
+
+ b => null()
+
+ !$omp target is_device_ptr(dptr)
+ call c_f_pointer(dptr, b, [n])
+ do i = 1, n
+ b(i) = b(i) + 1
+ end do
+ !$omp end target
+ end subroutine kernel
+end module mod
+
+program is_device_ptr_target
+ use iso_c_binding, only : c_ptr, c_loc, c_f_pointer
+ use omp_lib, only: omp_get_default_device, omp_get_mapped_ptr
+ use mod, only: kernel, n
+ implicit none
+
+ integer, dimension(n), target :: a
+ integer :: dev
+ type(c_ptr) :: dptr
+
+ a = [2, 4, 6, 8]
+ print '("BEFORE:", I3)', a
+
+ dev = omp_get_default_device()
+
+ !$omp target data map(tofrom: a)
+ dptr = omp_get_mapped_ptr(c_loc(a), dev)
+ call kernel(dptr)
+ !$omp end target data
+
+ print '("AFTER: ", I3)', a
+
+ if (all(a == [3, 5, 7, 9])) then
+ print '("PASS")'
+ else
+ print '("FAIL ", I3)', a
+ end if
+
+end program is_device_ptr_target
+
+!CHECK: PASS