Replace calls to compile_internal with compile_subroutine

Author: shwina

python/cuda_cccl/cuda/compute/_cccl_interop.py

@@ -248,7 +248,10 @@ def codegen(context, builder, impl_sig, args):
             input_vals = [builder.load(p) for p in input_ptrs]
 
             # Call the original operator
-            result = context.compile_internal(builder, op, sig, input_vals)
+            # See NVIDIA/numba-cuda#590 for why we need compile_subroutine
+            # vs compile_internal here:
+            cres = context.compile_subroutine(builder, op, sig, caching=False)
+            result = context.call_internal(builder, cres.fndesc, sig, input_vals)
 
             # Store the result
             builder.store(result, ret_ptr)
@@ -298,7 +301,8 @@ def codegen(context, builder, impl_sig, args):
             offset_val = builder.load(offset_ptr)
 
             sig = types.void(state_ptr_type, types.uint64)
-            context.compile_internal(builder, advance_fn, sig, [state_ptr, offset_val])
+            cres = context.compile_subroutine(builder, advance_fn, sig, caching=False)
+            result = context.call_internal(builder, cres.fndesc, sig, [state_ptr, offset_val])
 
             return context.get_dummy_value()
         return void_sig, codegen
@@ -342,7 +346,8 @@ def codegen(context, builder, impl_sig, args):
             result_ptr = builder.bitcast(args[1], value_type_llvm.as_pointer())
 
             sig = types.void(state_ptr_type, types.CPointer(value_type))
-            context.compile_internal(builder, deref_fn, sig, [state_ptr, result_ptr])
+            cres = context.compile_subroutine(builder, deref_fn, sig, caching=False)
+            result = context.call_internal(builder, cres.fndesc, sig, [state_ptr, result_ptr])
 
             return context.get_dummy_value()
         return void_sig, codegen
@@ -388,7 +393,8 @@ def codegen(context, builder, impl_sig, args):
             value_val = builder.load(value_ptr)
 
             sig = types.void(state_ptr_type, value_type)
-            context.compile_internal(builder, deref_fn, sig, [state_ptr, value_val])
+            cres = context.compile_subroutine(builder, deref_fn, sig, caching=False)
+            result = context.call_internal(builder, cres.fndesc, sig, [state_ptr, value_val])
 
             return context.get_dummy_value()
         return void_sig, codegen