Always ensure consistency of new MPI datatypes

Also:

* skip test of custom primitive type of size 80 on 32-bit systems in Julia
  v1.12+, as the MPI extent doesn't match Julia's aligned size
* add a test for a new primitive datatype larger than 64 bits which should work
  on 32-bit systems

src/datatypes.jl

@@ -149,7 +149,7 @@ end
 
 function Datatype(::Type{T}) where {T}
     global created_datatypes
-    get!(created_datatypes, T) do
+    datatype = get!(created_datatypes, T) do
         datatype = Datatype()
         # lazily initialize so that it can be safely precompiled
         function init()
@@ -162,6 +162,15 @@ function Datatype(::Type{T}) where {T}
         init()
         datatype
     end
+
+    # Make sure the "aligned" size of the type matches the MPI "extent".
+    sz = sizeof(T)
+    al = Base.datatype_alignment(T)
+    mpi_extent = MPI.Types.extent(datatype)
+    aligned_size = (0, cld(sz,al)*al)
+    @assert mpi_extent == aligned_size "The MPI extent of type $(T) ($(mpi_extent[2])) does not match the size expected by Julia ($(aligned_size[2]))"
+
+    return datatype
 end
 
 function Base.show(io::IO, datatype::Datatype)

test/test_datatype.jl

@@ -87,17 +87,19 @@ end
 primitive type Primitive16 16 end
 primitive type Primitive24 24 end
 primitive type Primitive80 80 end
+primitive type Primitive104 104 end
+
+@testset for PrimitiveType in (Primitive16, Primitive24, Primitive80, Primitive104)
+    if PrimitiveType == Primitive80 && VERSION >= v"1.12-" && Sys.WORD_SIZE == 32
+        # LLVM alignment on 32-bit systems doesn't necessarily agree with MPI
+        # extent for all types larger than 64 bits.
+        continue
+    end
 
-@testset for PrimitiveType in (Primitive16, Primitive24, Primitive80)
     sz = sizeof(PrimitiveType)
     al = Base.datatype_alignment(PrimitiveType)
     @test MPI.Types.extent(MPI.Datatype(PrimitiveType)) == (0, cld(sz,al)*al)
 
-    if VERSION < v"1.3" && PrimitiveType == Primitive80
-        # alignment is broken on earlier Julia versions
-        continue
-    end
-
     arr = [Core.Intrinsics.trunc_int(PrimitiveType, UInt128(comm_rank + i)) for i = 1:4]
     arr_recv = Array{PrimitiveType}(undef,4)