Try removing __rust_cuda_ffi_safe_assert module

Author: juntyr

rust-cuda-kernel/src/kernel/wrapper/generate/cuda_wrapper.rs

@@ -61,13 +61,28 @@ pub(in super::super) fn quote_cuda_wrapper(
         },
     );
 
+    let private_func_params = func_params
+        .iter()
+        .map(|param| {
+            let mut private = syn::Ident::clone(param);
+            private.set_span(proc_macro::Span::def_site().into());
+            private
+        })
+        .collect::<Vec<_>>();
+
     quote! {
         #[cfg(target_os = "cuda")]
         #[#crate_path::device::specialise_kernel_function(#func_ident)]
         #[no_mangle]
         #[allow(unused_unsafe)]
         #(#func_attrs)*
         pub unsafe extern "ptx-kernel" fn #func_ident_hash(#(#ffi_inputs),*) {
+            extern "C" { #(
+                #[allow(dead_code)]
+                #[deny(improper_ctypes)]
+                static #private_func_params: #ffi_types;
+            )* }
+
             unsafe {
                 // Initialise the dynamically-sized thread-block shared memory
                 //  and the thread-local offset pointer that points to it
@@ -89,17 +104,6 @@ pub(in super::super) fn quote_cuda_wrapper(
                 ::core::arch::asm!(#KERNEL_TYPE_USE_END_CANARY);
             }
 
-            #[deny(improper_ctypes)]
-            mod __rust_cuda_ffi_safe_assert {
-                #[allow(unused_imports)]
-                use super::*;
-
-                extern "C" { #(
-                    #[allow(dead_code)]
-                    static #func_params: #ffi_types;
-                )* }
-            }
-
             #ffi_param_ptx_jit_wrap
         }
     }

rust-cuda-kernel/src/kernel/wrapper/generate/host_linker_macro/get_ptx.rs

@@ -80,10 +80,24 @@ pub(super) fn quote_get_ptx(
             .collect::<Vec<_>>()
     };
 
+    let private_func_params = func_params
+        .iter()
+        .map(|param| {
+            let mut private = syn::Ident::clone(param);
+            private.set_span(proc_macro::Span::def_site().into());
+            private
+        })
+        .collect::<Vec<_>>();
+
     quote! {
         fn get_ptx() -> &'static ::core::ffi::CStr {
-            #[allow(unused_imports)]
-            use __rust_cuda_ffi_safe_assert::#args;
+            #args_trait
+
+            extern "C" { #(
+                #[allow(dead_code)]
+                #[deny(improper_ctypes)]
+                static #private_func_params: #cpu_func_lifetime_erased_types;
+            )* }
 
             #crate_path::kernel::link_kernel!{
                 #func_ident #func_ident_hash #crate_name #crate_manifest_dir #generic_start_token
@@ -95,19 +109,6 @@ pub(super) fn quote_get_ptx(
 
             #(#type_layout_asserts)*
 
-            #[deny(improper_ctypes)]
-            mod __rust_cuda_ffi_safe_assert {
-                #[allow(unused_imports)]
-                use super::*;
-
-                #args_trait
-
-                extern "C" { #(
-                    #[allow(dead_code)]
-                    static #func_params: #cpu_func_lifetime_erased_types;
-                )* }
-            }
-
             PTX_CSTR
         }
     }

src/host/mod.rs

@@ -152,7 +152,6 @@ impl<'a, T: PortableBitSemantics + TypeGraphLayout> HostAndDeviceMutRef<'a, T> {
         }
     }
 
-    #[allow(dead_code)] // FIXME
     #[must_use]
     pub(crate) fn for_device<'b>(&'b mut self) -> DeviceMutRef<'a, T>
     where
@@ -164,7 +163,6 @@ impl<'a, T: PortableBitSemantics + TypeGraphLayout> HostAndDeviceMutRef<'a, T> {
         }
     }
 
-    #[allow(dead_code)] // FIXME
     #[must_use]
     pub(crate) fn for_host<'b: 'a>(&'b self) -> &'a T {
         self.host_ref