[flang][cuda] Add UNIFIED data attribute (#88171)

Latest version of the specification introduced the `UNIFIED` attribute
for data.


https://docs.nvidia.com/hpc-sdk/compilers/cuda-fortran-prog-guide/#cfref-var-attr-unified-data

This patch adds the attribute to parsing, semantic and lowering. 

The matching rules for dummy/actual arguments is not part of this patch.

Author: clementval

flang/include/flang/Common/Fortran.h

@@ -85,7 +85,8 @@ static constexpr int maxRank{15};
 ENUM_CLASS(CUDASubprogramAttrs, Host, Device, HostDevice, Global, Grid_Global)
 
 // CUDA data attributes; mutually exclusive
-ENUM_CLASS(CUDADataAttr, Constant, Device, Managed, Pinned, Shared, Texture)
+ENUM_CLASS(
+    CUDADataAttr, Constant, Device, Managed, Pinned, Shared, Texture, Unified)
 
 // OpenACC device types
 ENUM_CLASS(

flang/include/flang/Optimizer/Support/Utils.h

@@ -158,6 +158,9 @@ getCUDADataAttribute(mlir::MLIRContext *mlirContext,
     case Fortran::common::CUDADataAttr::Texture:
       // Obsolete attribute
       return {};
+    case Fortran::common::CUDADataAttr::Unified:
+      attr = fir::CUDADataAttribute::Unified;
+      break;
     }
     return fir::CUDADataAttributeAttr::get(mlirContext, attr);
   }

flang/include/flang/Parser/parse-tree.h

@@ -991,7 +991,7 @@ struct ComponentArraySpec {
 //        access-spec | ALLOCATABLE |
 //        CODIMENSION lbracket coarray-spec rbracket |
 //        CONTIGUOUS | DIMENSION ( component-array-spec ) | POINTER |
-// (CUDA) CONSTANT | DEVICE | MANAGED | PINNED | SHARED | TEXTURE
+// (CUDA) CONSTANT | DEVICE | MANAGED | PINNED | SHARED | TEXTURE | UNIFIED
 EMPTY_CLASS(Allocatable);
 EMPTY_CLASS(Pointer);
 EMPTY_CLASS(Contiguous);

flang/lib/Parser/Fortran-parsers.cpp

@@ -703,13 +703,15 @@ TYPE_PARSER(construct<AttrSpec>(accessSpec) ||
     extension<LanguageFeature::CUDA>(
         construct<AttrSpec>(Parser<common::CUDADataAttr>{})))
 
-// CUDA-data-attr -> CONSTANT | DEVICE | MANAGED | PINNED | SHARED | TEXTURE
+// CUDA-data-attr ->
+//     CONSTANT | DEVICE | MANAGED | PINNED | SHARED | TEXTURE | UNIFIED
 TYPE_PARSER("CONSTANT" >> pure(common::CUDADataAttr::Constant) ||
     "DEVICE" >> pure(common::CUDADataAttr::Device) ||
     "MANAGED" >> pure(common::CUDADataAttr::Managed) ||
     "PINNED" >> pure(common::CUDADataAttr::Pinned) ||
     "SHARED" >> pure(common::CUDADataAttr::Shared) ||
-    "TEXTURE" >> pure(common::CUDADataAttr::Texture))
+    "TEXTURE" >> pure(common::CUDADataAttr::Texture) ||
+    "UNIFIED" >> pure(common::CUDADataAttr::Unified))
 
 // R804 object-name -> name
 constexpr auto objectName{name};

flang/lib/Semantics/check-declarations.cpp

@@ -988,6 +988,13 @@ void CheckHelper::CheckObjectEntity(
             symbol.name());
       }
       break;
+    case common::CUDADataAttr::Unified:
+      if ((!subpDetails || inDeviceSubprogram) && !isComponent) {
+        messages_.Say(
+            "Object '%s' with ATTRIBUTES(UNIFIED) must be declared in a host subprogram"_err_en_US,
+            symbol.name());
+      }
+      break;
     case common::CUDADataAttr::Texture:
       messages_.Say(
           "ATTRIBUTES(TEXTURE) is obsolete and no longer supported"_err_en_US);

flang/test/Lower/CUDA/cuda-data-attribute.cuf

@@ -19,16 +19,20 @@ subroutine local_var_attrs
   real, device :: rd
   real, allocatable, managed :: rm
   real, allocatable, pinned :: rp
+  real, unified :: ru
 end subroutine
 
 ! CHECK-LABEL: func.func @_QMcuda_varPlocal_var_attrs()
 ! CHECK: %{{.*}}:2 = hlfir.declare %{{.*}} {cuda_attr = #fir.cuda<device>, uniq_name = "_QMcuda_varFlocal_var_attrsErd"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)
 ! CHECK: %{{.*}}:2 = hlfir.declare %{{.*}} {cuda_attr = #fir.cuda<managed>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QMcuda_varFlocal_var_attrsErm"} : (!fir.ref<!fir.box<!fir.heap<f32>>>) -> (!fir.ref<!fir.box<!fir.heap<f32>>>, !fir.ref<!fir.box<!fir.heap<f32>>>)
 ! CHECK: %{{.*}}:2 = hlfir.declare %{{.*}} {cuda_attr = #fir.cuda<pinned>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QMcuda_varFlocal_var_attrsErp"} : (!fir.ref<!fir.box<!fir.heap<f32>>>) -> (!fir.ref<!fir.box<!fir.heap<f32>>>, !fir.ref<!fir.box<!fir.heap<f32>>>)
+! CHECK: %{{.*}}:2 = hlfir.declare %{{.*}} {cuda_attr = #fir.cuda<unified>, uniq_name = "_QMcuda_varFlocal_var_attrsEru"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)
+
 
 ! FIR: %{{.*}} = fir.declare %{{.*}} {cuda_attr = #fir.cuda<device>, uniq_name = "_QMcuda_varFlocal_var_attrsErd"} : (!fir.ref<f32>) -> !fir.ref<f32>
 ! FIR: %{{.*}} = fir.declare %{{.*}} {cuda_attr = #fir.cuda<managed>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QMcuda_varFlocal_var_attrsErm"} : (!fir.ref<!fir.box<!fir.heap<f32>>>) -> !fir.ref<!fir.box<!fir.heap<f32>>>
 ! FIR: %{{.*}} = fir.declare %{{.*}} {cuda_attr = #fir.cuda<pinned>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QMcuda_varFlocal_var_attrsErp"} : (!fir.ref<!fir.box<!fir.heap<f32>>>) -> !fir.ref<!fir.box<!fir.heap<f32>>>
+! FIR: %{{.*}} = fir.declare %{{.*}} {cuda_attr = #fir.cuda<unified>, uniq_name = "_QMcuda_varFlocal_var_attrsEru"} : (!fir.ref<f32>) -> !fir.ref<f32>
 
 subroutine dummy_arg_device(dd)
   real, device :: dd
@@ -51,4 +55,11 @@ end subroutine
 ! CHECK-SAME: %[[ARG0:.*]]: !fir.ref<!fir.box<!fir.heap<f32>>> {fir.bindc_name = "dp", fir.cuda_attr = #fir.cuda<pinned>}) {
 ! CHECK: %{{.*}}:2 = hlfir.declare %[[ARG0]] {cuda_attr = #fir.cuda<pinned>, fortran_attrs = #fir.var_attrs<allocatable>, uniq_name = "_QMcuda_varFdummy_arg_pinnedEdp"} : (!fir.ref<!fir.box<!fir.heap<f32>>>) -> (!fir.ref<!fir.box<!fir.heap<f32>>>, !fir.ref<!fir.box<!fir.heap<f32>>>)
 
+subroutine dummy_arg_unified(du)
+  real, unified :: du
+end subroutine
+! CHECK-LABEL: func.func @_QMcuda_varPdummy_arg_unified(
+! CHECK-SAME: %[[ARG0:.*]]: !fir.ref<f32> {fir.bindc_name = "du", fir.cuda_attr = #fir.cuda<unified>})
+! CHECK: %{{.*}}:2 = hlfir.declare %[[ARG0]] {cuda_attr = #fir.cuda<unified>, uniq_name = "_QMcuda_varFdummy_arg_unifiedEdu"} : (!fir.ref<f32>) -> (!fir.ref<f32>, !fir.ref<f32>)
+
 end module

flang/test/Semantics/cuf03.cuf

@@ -1,6 +1,12 @@
 ! RUN: %python %S/test_errors.py %s %flang_fc1
 ! Exercise CUDA data attribute checks
 module m
+  type :: t1
+    integer :: i
+  end type
+  type :: t2
+    real, unified :: r(10) ! ok
+  end type
   real, constant :: mc ! ok
   real, constant :: mci = 1. ! ok
   !ERROR: Object 'mcl' with ATTRIBUTES(CONSTANT) may not be allocatable, pointer, or target
@@ -48,6 +54,9 @@ module m
   real, texture, pointer :: mt
   !ERROR: 'bigint' has intrinsic type 'INTEGER(16)' that is not available on the device
   integer(16), device :: bigint
+  !ERROR: Object 'um' with ATTRIBUTES(UNIFIED) must be declared in a host subprogram
+  real, unified :: um
+
  contains
   attributes(device) subroutine devsubr(n,da)
     integer, intent(in) :: n
@@ -57,6 +66,8 @@ module m
     !WARNING: Pointer 'dp' may not be associated in a device subprogram
     real, device, pointer :: dp
     real, constant :: rc ! ok
+    !ERROR: Object 'u' with ATTRIBUTES(UNIFIED) must be declared in a host subprogram
+    real, unified :: u
   end subroutine
 
   subroutine host()
@@ -70,4 +81,10 @@ module m
     rs = 1 ! ok
   end subroutine
 
+  subroutine host2()
+    real, unified :: ru ! ok
+    type(t1), unified :: tu ! ok
+    type(t2) :: t ! ok
+  end subroutine
+
 end module