diff --git a/examples/python/hello-loopy.py b/examples/python/hello-loopy.py
index 82ff2e60dee345fe16771d09cb39d2d56e9f493d..7c5de5a1b1d7042498a12204959a59021ac5e0d8 100644
--- a/examples/python/hello-loopy.py
+++ b/examples/python/hello-loopy.py
@@ -26,5 +26,5 @@ knl = lp.split_iname(knl, "i", 128, outer_tag="g.0", inner_tag="l.0")
 evt, (out,) = knl(queue, a=a)
 # ENDEXAMPLE
 
-cknl = lp.CompiledKernel(ctx, knl)
-print(cknl.get_highlighted_code({"a": np.float32}))
+knl = lp.add_and_infer_dtypes(knl, {"a": np.dtype(np.float32)})
+print(lp.generate_code_v2(knl).device_code())
diff --git a/loopy/preprocess.py b/loopy/preprocess.py
index 17226b63addb9e2e30d556730aa326d2ed59128c..c331ccc8259645029866cad4a518cb8198428836 100644
--- a/loopy/preprocess.py
+++ b/loopy/preprocess.py
@@ -48,20 +48,22 @@ def prepare_for_caching(kernel):
     import loopy as lp
     new_args = []
 
+    tgt = kernel.target
+
     for arg in kernel.args:
         dtype = arg.dtype
-        if dtype is not None and dtype is not lp.auto:
-            dtype = dtype.with_target(kernel.target)
+        if dtype is not None and dtype is not lp.auto and dtype.target is not tgt:
+            arg = arg.copy(dtype=dtype.with_target(kernel.target))
 
-        new_args.append(arg.copy(dtype=dtype))
+        new_args.append(arg)
 
     new_temporary_variables = {}
     for name, temp in six.iteritems(kernel.temporary_variables):
         dtype = temp.dtype
-        if dtype is not None and dtype is not lp.auto:
-            dtype = dtype.with_target(kernel.target)
+        if dtype is not None and dtype is not lp.auto and dtype.target is not tgt:
+            temp = temp.copy(dtype=dtype.with_target(tgt))
 
-        new_temporary_variables[name] = temp.copy(dtype=dtype)
+        new_temporary_variables[name] = temp
 
     kernel = kernel.copy(
             args=new_args,
diff --git a/loopy/target/c/codegen/expression.py b/loopy/target/c/codegen/expression.py
index 68cc32e56be077c7e45d11b9e2aade86b04494cc..8f924d3aee3b9f2982006fdb7b558cccac6785e3 100644
--- a/loopy/target/c/codegen/expression.py
+++ b/loopy/target/c/codegen/expression.py
@@ -167,6 +167,13 @@ class ExpressionToCExpressionMapper(IdentityMapper):
         def base_impl(expr, type_context):
             return self.rec(expr.aggregate, type_context)[self.rec(expr.index, 'i')]
 
+        def make_var(name):
+            from loopy import TaggedVariable
+            if isinstance(expr.aggregate, TaggedVariable):
+                return TaggedVariable(name, expr.aggregate.tag)
+            else:
+                return var(name)
+
         from pymbolic.primitives import Variable
         if not isinstance(expr.aggregate, Variable):
             return base_impl(expr, type_context)
@@ -224,16 +231,16 @@ class ExpressionToCExpressionMapper(IdentityMapper):
                         (isinstance(ary, (ConstantArg, GlobalArg)) or
                          (isinstance(ary, TemporaryVariable) and ary.base_storage))):
                     # unsubscripted global args are pointers
-                    result = var(access_info.array_name)[0]
+                    result = make_var(access_info.array_name)[0]
 
                 else:
                     # unsubscripted temp vars are scalars
                     # (unless they use base_storage)
-                    result = var(access_info.array_name)
+                    result = make_var(access_info.array_name)
 
             else:
                 subscript, = access_info.subscripts
-                result = var(access_info.array_name)[self.rec(subscript, 'i')]
+                result = make_var(access_info.array_name)[self.rec(subscript, 'i')]
 
             if access_info.vector_index is not None:
                 return self.codegen_state.ast_builder.add_vector_access(
@@ -716,6 +723,8 @@ class CExpressionToCodeMapper(RecursiveMapper):
     def map_variable(self, expr, enclosing_prec):
         return expr.name
 
+    map_tagged_variable = map_variable
+
     def map_lookup(self, expr, enclosing_prec):
         return self.parenthesize_if_needed(
                 "%s.%s" % (