[flang][cuda] Perform inlined assignment when field is c_devptr (llvm#124322)

When a field in a derived type is `c_devptr`, keep check if we can do a
memcpy instead of falling back to the runtime assignment.

Many internal CUDA Fortran derived type have a `c_devptr` field and this
would lead to stack overflow on the device if the assignment is
performed by the runtime function.

Author: clementval

flang/lib/Optimizer/Builder/FIRBuilder.cpp

@@ -1410,7 +1410,8 @@ static bool recordTypeCanBeMemCopied(fir::RecordType recordType) {
   for (auto [_, fieldType] : recordType.getTypeList()) {
     // Derived type component may have user assignment (so far, we cannot tell
     // in FIR, so assume it is always the case, TODO: get the actual info).
-    if (mlir::isa<fir::RecordType>(fir::unwrapSequenceType(fieldType)))
+    if (mlir::isa<fir::RecordType>(fir::unwrapSequenceType(fieldType)) &&
+        !fir::isa_builtin_c_devptr_type(fir::unwrapSequenceType(fieldType)))
       return false;
     // Allocatable components need deep copy.
     if (auto boxType = mlir::dyn_cast<fir::BaseBoxType>(fieldType))

flang/test/Lower/CUDA/cuda-devptr.cuf

@@ -4,6 +4,12 @@
 
 module cudafct
   use __fortran_builtins, only : c_devptr => __builtin_c_devptr
+  
+  type :: t1
+    type(c_devptr) :: devp
+    integer :: a
+  end type
+
 contains
   function c_devloc(x)
     use iso_c_binding, only: c_loc
@@ -12,6 +18,10 @@ contains
     real, target, device :: x
     c_devloc%cptr = c_loc(x)
   end function
+
+  attributes(device) function get_t1()
+    type(t1) :: get_t1
+  end
 end
 
 subroutine sub1()
@@ -68,3 +78,12 @@ end subroutine
 ! CHECK: %[[P_ADDR_COORD:.*]] = fir.coordinate_of %[[P_CPTR_COORD]], %[[ADDRESS_FIELD]] : (!fir.ref<!fir.type<_QM__fortran_builtinsT__builtin_c_ptr{__address:i64}>>, !fir.field) -> !fir.ref<i64>
 ! CHECK: %[[ADDR:.*]] = fir.load %[[RES_ADDR_COORD]] : !fir.ref<i64>
 ! CHECK: fir.store %[[ADDR]] to %[[P_ADDR_COORD]] : !fir.ref<i64>
+
+attributes(global) subroutine assign_nested_c_devptr(p, a)
+  use cudafct
+  type(t1), device :: p
+  p = get_t1()
+end subroutine
+
+! CHECK-LABEL: func.func @_QPassign_nested_c_devptr
+! CHECK-NOT: fir.call @_FortranAAssign