diff --git a/src/device/cuda.jl b/src/device/cuda.jl
index 4e2772ad..758999a0 100644
--- a/src/device/cuda.jl
+++ b/src/device/cuda.jl
@@ -11,7 +11,9 @@ include("cuda/assertion.jl")
 include("cuda/memory_dynamic.jl")
 include("cuda/atomics.jl")
 include("cuda/misc.jl")
+if VERSION >= v"1.4.1"
 include("cuda/wmma.jl")
+end
 
 # functionality from libdevice
 #
diff --git a/src/device/cuda/memory_shared.jl b/src/device/cuda/memory_shared.jl
index c19d5692..99ed6e03 100644
--- a/src/device/cuda/memory_shared.jl
+++ b/src/device/cuda/memory_shared.jl
@@ -61,7 +61,6 @@ end
     end
 
     T_ptr = convert(LLVMType, DevicePtr{T,AS.Shared})
-    T_actual_ptr = LLVM.PointerType(eltyp)
 
     # create a function
     llvm_f, _ = create_function(T_ptr)
@@ -92,10 +91,9 @@ end
         entry = BasicBlock(llvm_f, "entry", JuliaContext())
         position!(builder, entry)
 
-        ptr_with_as = gep!(builder, gv, [ConstantInt(0, JuliaContext()),
-                                         ConstantInt(0, JuliaContext())])
+        ptr = gep!(builder, gv, [ConstantInt(0, JuliaContext()),
+                                 ConstantInt(0, JuliaContext())])
 
-        ptr = addrspacecast!(builder, ptr_with_as, T_actual_ptr)
         val = ptrtoint!(builder, ptr, T_ptr)
         ret!(builder, val)
     end
diff --git a/src/device/cuda/wmma.jl b/src/device/cuda/wmma.jl
index b8c75571..088166c8 100644
--- a/src/device/cuda/wmma.jl
+++ b/src/device/cuda/wmma.jl
@@ -7,9 +7,6 @@ using CUDAnative: AS, DevicePtr
 # CONSTANTS
 ################################################################################
 
-# Determines whether or not to Core.AddrSpacePtr is available
-const addrspaceptr_available = (VERSION >= v"1.5.0-DEV.324")
-
 # Maps PTX types to Julia array types
 const map_ptx_to_jl_array = Dict(
                                  "f16" => Float16,
@@ -52,24 +49,14 @@ get_frag_info(matrix, ptx_el_type) = (
 
 get_addrspace_info(addr_space) = convert(Int, map_ptx_as_to_as_ty[addr_space])
 
-if addrspaceptr_available
 @generated function Base.cconvert(::Type{Core.AddrSpacePtr{T, as}}, x::DevicePtr{T, AS}) where {T, as, AS}
-    # Addrspacecast from i8* to i8* is invalid in LLVM
-    if as == 0
-        return quote
-            return Base.bitcast(Core.AddrSpacePtr{T, as}, x)
-        end
-    else
-        ir = "%p = inttoptr i64 %0 to i8*
-            %ptr = addrspacecast i8* %p to i8 addrspace($as)*
-            ret i8 addrspace($as)* %ptr"
+    ir = "%ptr = inttoptr i64 %0 to i8 addrspace($as)*
+          ret i8 addrspace($as)* %ptr"
 
-        return quote
-            return Base.llvmcall($ir, Core.AddrSpacePtr{T, as}, Tuple{Int64}, Base.bitcast(Int64, x))
-        end
+    return quote
+        return Base.llvmcall($ir, Core.AddrSpacePtr{T, as}, Tuple{Int64}, Base.bitcast(Int64, x))
     end
 end
-end
 
 # Fix for https://github.com/JuliaGPU/CUDAnative.jl/issues/587.
 # Instead of ccall'ing the intrinsics with NTuple{N, T} (which gets lowered to
@@ -141,7 +128,7 @@ for mat in ["a", "b", "c"],
 
     ccall_name = "extern $llvm_intr"
 
-    ptr_ty = addrspaceptr_available ? Core.AddrSpacePtr{arr_ty, addr_space_int} : Ref{arr_ty}
+    ptr_ty = Core.AddrSpacePtr{arr_ty, addr_space_int}
     struct_ty = Symbol("LLVMStruct$sz")
 
     @eval $func_name(src_addr, stride) = convert(NTuple{$sz, $frag_ty}, ccall($ccall_name, llvmcall, $struct_ty{$frag_ty}, ($ptr_ty, Int32), src_addr, stride))
@@ -196,7 +183,7 @@ for mat in ["d"],
     frag_types = ntuple(i -> frag_ty, sz)
     frag_vars = ntuple(i -> :(data[$i]), sz)
 
-    ptr_ty = addrspaceptr_available ? Core.AddrSpacePtr{arr_ty, addr_space_int} : Ref{arr_ty}
+    ptr_ty = Core.AddrSpacePtr{arr_ty, addr_space_int}
 
     @eval $func_name(dst_addr, data, stride) = ccall($ccall_name, llvmcall, Nothing, ($ptr_ty, $(frag_types...), Int32), dst_addr, $(frag_vars...), stride)
     @eval export $func_name
diff --git a/src/device/pointer.jl b/src/device/pointer.jl
index 14b942a9..7ea9cd4a 100644
--- a/src/device/pointer.jl
+++ b/src/device/pointer.jl
@@ -118,7 +118,7 @@ Base.:(+)(x::Integer, y::DevicePtr) = y + x
     T_int = convert(LLVMType, Int)
     T_ptr = convert(LLVMType, DevicePtr{T,A})
 
-    T_actual_ptr = LLVM.PointerType(eltyp)
+    T_actual_ptr = LLVM.PointerType(eltyp, convert(Int, A))
 
     # create a function
     param_types = [T_ptr, T_int]
@@ -130,10 +130,8 @@ Base.:(+)(x::Integer, y::DevicePtr) = y + x
         position!(builder, entry)
 
         ptr = inttoptr!(builder, parameters(llvm_f)[1], T_actual_ptr)
-
         ptr = gep!(builder, ptr, [parameters(llvm_f)[2]])
-        ptr_with_as = addrspacecast!(builder, ptr, LLVM.PointerType(eltyp, convert(Int, A)))
-        ld = load!(builder, ptr_with_as)
+        ld = load!(builder, ptr)
 
         if A != AS.Generic
             metadata(ld)[LLVM.MD_tbaa] = tbaa_addrspace(A)
@@ -153,7 +151,7 @@ end
     T_int = convert(LLVMType, Int)
     T_ptr = convert(LLVMType, DevicePtr{T,A})
 
-    T_actual_ptr = LLVM.PointerType(eltyp)
+    T_actual_ptr = LLVM.PointerType(eltyp, convert(Int, A))
 
     # create a function
     param_types = [T_ptr, eltyp, T_int]
@@ -165,11 +163,9 @@ end
         position!(builder, entry)
 
         ptr = inttoptr!(builder, parameters(llvm_f)[1], T_actual_ptr)
-
         ptr = gep!(builder, ptr, [parameters(llvm_f)[3]])
-        ptr_with_as = addrspacecast!(builder, ptr, LLVM.PointerType(eltyp, convert(Int, A)))
         val = parameters(llvm_f)[2]
-        st = store!(builder, val, ptr_with_as)
+        st = store!(builder, val, ptr)
 
         if A != AS.Generic
             metadata(st)[LLVM.MD_tbaa] = tbaa_addrspace(A)
@@ -201,8 +197,7 @@ const LDGTypes = Union{UInt8, UInt16, UInt32, UInt64,
     T_int32 = LLVM.Int32Type(JuliaContext())
     T_ptr = convert(LLVMType, DevicePtr{T,AS.Global})
 
-    T_actual_ptr = LLVM.PointerType(eltyp)
-    T_actual_ptr_as = LLVM.PointerType(eltyp, convert(Int, AS.Global))
+    T_actual_ptr = LLVM.PointerType(eltyp, convert(Int, AS.Global))
 
     # create a function
     param_types = [T_ptr, T_int]
@@ -222,7 +217,7 @@ const LDGTypes = Union{UInt8, UInt16, UInt32, UInt64,
         "llvm.nvvm.ldg.global.$class.$typ.p1$typ"
     end
     mod = LLVM.parent(llvm_f)
-    intrinsic_typ = LLVM.FunctionType(eltyp, [T_actual_ptr_as, T_int32])
+    intrinsic_typ = LLVM.FunctionType(eltyp, [T_actual_ptr, T_int32])
     intrinsic = LLVM.Function(mod, intrinsic_name, intrinsic_typ)
 
     # generate IR
@@ -231,11 +226,9 @@ const LDGTypes = Union{UInt8, UInt16, UInt32, UInt64,
         position!(builder, entry)
 
         ptr = inttoptr!(builder, parameters(llvm_f)[1], T_actual_ptr)
-
         ptr = gep!(builder, ptr, [parameters(llvm_f)[2]])
-        ptr_with_as = addrspacecast!(builder, ptr, T_actual_ptr_as)
         ld = call!(builder, intrinsic,
-                   [ptr_with_as, ConstantInt(Int32(align), JuliaContext())])
+                   [ptr, ConstantInt(Int32(align), JuliaContext())])
 
         metadata(ld)[LLVM.MD_tbaa] = tbaa_addrspace(AS.Global)
 
diff --git a/test/device/cuda.jl b/test/device/cuda.jl
index 43e51037..6d43e6b1 100644
--- a/test/device/cuda.jl
+++ b/test/device/cuda.jl
@@ -1133,6 +1133,17 @@ end
     end
 end
 
+@testset "shared memory" begin
+    function kernel()
+        shared = @cuStaticSharedMem(Float32, 1)
+        @atomic shared[threadIdx().x] += 0f0
+        return
+    end
+
+    @cuda kernel()
+    synchronize()
+end
+
 end
 
 end
diff --git a/test/device/wmma.jl b/test/device/wmma.jl
index 44f0a261..dfe9ae5e 100644
--- a/test/device/wmma.jl
+++ b/test/device/wmma.jl
@@ -1,6 +1,5 @@
 # Need https://github.com/JuliaLang/julia/pull/33970
 # and  https://github.com/JuliaLang/julia/pull/34043
-if VERSION >= v"1.4.0-DEV.666" && capability(device()) >= v"7.0"
 
 using CUDAnative.WMMA
 
@@ -294,4 +293,3 @@ end
 ################################################################################
 
 end
-end
diff --git a/test/runtests.jl b/test/runtests.jl
index 1cff84a2..38424346 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -33,7 +33,9 @@ include("device/execution.jl")
 include("device/pointer.jl")
 include("device/array.jl")
 include("device/cuda.jl")
+if VERSION >= v"1.4.1" && capability(device()) >= v"7.0"
 include("device/wmma.jl")
+end
 
 include("nvtx.jl")