Fix nvvm_internal attribute usage. add example now compiles and runs!

`register_attr` was removed in favor of `register_tool`. This updates how the nvvm attributes are applied in our proc_macros and how the nvvm backend parses them. Also removes redundant applications of no_std, as cuda_builder applies it to all crates built on the cuda target via rustc flags.

The add example crate now compiles and runs the geneated ptx! :D

Author: jorge-ortega

crates/cuda_builder/src/lib.rs

@@ -379,7 +379,7 @@ fn invoke_rustc(builder: &CudaBuilder) -> Result<PathBuf, CudaBuilderError> {
     let mut rustflags = vec![
         format!("-Zcodegen-backend={}", rustc_codegen_nvvm.display()),
         "-Zcrate-attr=feature(register_tool)".into(),
-        "-Zcrate-attr=register_tool(rust_cuda)".into(),
+        "-Zcrate-attr=register_tool(nvvm_internal)".into(),
         "-Zcrate-attr=no_std".into(),
         "-Zsaturating_float_casts=false".into(),
     ];
@@ -440,8 +440,7 @@ fn invoke_rustc(builder: &CudaBuilder) -> Result<PathBuf, CudaBuilderError> {
         "--lib",
         "--message-format=json-render-diagnostics",
         "-Zbuild-std=core,alloc",
-        "--target",
-        "nvptx64-nvidia-cuda",
+        "--target=nvptx64-nvidia-cuda",
     ]);
 
     cargo.args(&builder.build_args);

crates/cuda_std/src/lib.rs

@@ -5,7 +5,7 @@
 //!
 //! This library will build on non-nvptx targets or targets not using the nvvm backend. However, it will not
 //! be usable, and it will throw linker errors if you attempt to use most of the functions in the library.
-//! However, [`kernel`] automatically cfg-gates the function annotated for `nvptx64` or `nvptx`, therefore,
+//! However, [`kernel`] automatically cfg-gates the function annotated for `nvptx64`, therefore,
 //! no "actual" functions from this crate should be used when compiling for a non-nvptx target.
 //!
 //! This crate cannot be used with the llvm ptx backend either, it heavily relies on external functions implicitly
@@ -24,9 +24,7 @@
 #![allow(internal_features)]
 #![cfg_attr(
     target_os = "cuda",
-    no_std,
     feature(alloc_error_handler, asm_experimental_arch, link_llvm_intrinsics),
-    register_attr(nvvm_internal)
 )]
 
 extern crate alloc;

crates/cuda_std/src/shared.rs

@@ -68,7 +68,7 @@ macro_rules! shared_array {
 /// Gets a pointer to the dynamic shared memory that was allocated by the caller of the kernel. The
 /// data is left uninitialized.
 ///
-/// **Calling this function multiple times will yield the same pointer**.  
+/// **Calling this function multiple times will yield the same pointer**.
 #[gpu_only]
 pub fn dynamic_shared_mem<T>() -> *mut T {
     // it is unclear whether an alignment of 16 is actually required for correctness, however,
@@ -77,7 +77,7 @@ pub fn dynamic_shared_mem<T>() -> *mut T {
     extern "C" {
         // need to use nvvm_internal and not address_space because address_space only parses
         // static definitions, not extern static definitions.
-        #[nvvm_internal(addrspace(3))]
+        #[nvvm_internal::addrspace(3)]
         #[allow(improper_ctypes)]
         // mangle it a bit to make sure nobody makes the same thing
         #[link_name = "_Zcuda_std_dyn_shared"]

crates/cuda_std_macros/src/lib.rs

@@ -27,7 +27,7 @@ pub fn kernel(input: proc_macro::TokenStream, item: proc_macro::TokenStream) ->
     let mut item = parse_macro_input!(item as ItemFn);
     let no_mangle = parse_quote!(#[no_mangle]);
     item.attrs.push(no_mangle);
-    let internal = parse_quote!(#[cfg_attr(target_arch="nvptx64", nvvm_internal(kernel(#input)))]);
+    let internal = parse_quote!(#[cfg_attr(target_arch="nvptx64", nvvm_internal::kernel(#input))]);
     item.attrs.push(internal);
 
     // used to guarantee some things about how params are passed in the codegen.
@@ -155,12 +155,7 @@ pub fn gpu_only(_attr: proc_macro::TokenStream, item: proc_macro::TokenStream) -
     } = syn::parse_macro_input!(item as syn::ItemFn);
 
     let mut cloned_attrs = attrs.clone();
-    cloned_attrs.retain(|a| {
-        !a.path()
-            .get_ident()
-            .map(|x| *x == "nvvm_internal")
-            .unwrap_or_default()
-    });
+    cloned_attrs.retain(|a| a.path().segments[0].ident != "nvvm");
 
     let fn_name = sig.ident.clone();
 
@@ -203,7 +198,7 @@ pub fn externally_visible(
         "#[externally_visible] function should also be #[no_mangle]"
     );
 
-    let new_attr = parse_quote!(#[cfg_attr(target_os = "cuda", nvvm_internal(used))]);
+    let new_attr = parse_quote!(#[cfg_attr(target_os = "cuda", nvvm_internal::used)]);
     func.attrs.push(new_attr);
 
     func.into_token_stream().into()
@@ -231,7 +226,7 @@ pub fn address_space(attr: proc_macro::TokenStream, item: proc_macro::TokenStrea
     };
 
     let new_attr =
-        parse_quote!(#[cfg_attr(target_os = "cuda", nvvm_internal(addrspace(#addrspace_num)))]);
+        parse_quote!(#[cfg_attr(target_os = "cuda", nvvm_internal::addrspace(#addrspace_num))]);
     global.attrs.push(new_attr);
 
     global.into_token_stream().into()

crates/rustc_codegen_nvvm/src/attributes.rs

@@ -116,31 +116,23 @@ impl NvvmAttributes {
         let mut nvvm_attrs = Self::default();
 
         for attr in attrs {
-            if attr.has_name(cx.symbols.nvvm_internal) {
+            if attr.path_matches(&[cx.symbols.nvvm_internal, cx.symbols.kernel]) {
+                nvvm_attrs.kernel = true;
+            } else if attr.path_matches(&[cx.symbols.nvvm_internal, sym::used]) {
+                nvvm_attrs.used = true;
+            } else if attr.path_matches(&[cx.symbols.nvvm_internal, cx.symbols.addrspace]) {
                 let args = attr.meta_item_list().unwrap_or_default();
-                if let Some(arg) = args.first() {
-                    if arg.has_name(cx.symbols.kernel) {
-                        nvvm_attrs.kernel = true;
-                    }
-                    if arg.has_name(sym::used) {
-                        nvvm_attrs.used = true;
-                    }
-                    if arg.has_name(cx.symbols.addrspace) {
-                        let args = arg.meta_item_list().unwrap_or_default();
-                        if let Some(MetaItemInner::Lit(MetaItemLit {
-                            kind: LitKind::Int(val, _),
-                            ..
-                        })) = args.first()
-                        {
-                            nvvm_attrs.addrspace = Some(val.get() as u8);
-                        } else {
-                            panic!();
-                        }
-                    }
+                if let Some(MetaItemInner::Lit(MetaItemLit {
+                    kind: LitKind::Int(val, _),
+                    ..
+                })) = args.first()
+                {
+                    nvvm_attrs.addrspace = Some(val.get() as u8);
+                } else {
+                    panic!();
                 }
             }
         }
-
         nvvm_attrs
     }
 }

crates/rustc_codegen_nvvm/src/context.rs

@@ -257,7 +257,7 @@ impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> {
 impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> {
     /// Computes the address space for a static.
     pub fn static_addrspace(&self, instance: Instance<'tcx>) -> AddressSpace {
-        let ty = instance.ty(self.tcx, ty::TypingEnv::fully_monomorphized());
+        let ty = instance.ty(self.tcx, self.typing_env());
         let is_mutable = self.tcx().is_mutable_static(instance.def_id());
         let attrs = self.tcx.get_attrs_unchecked(instance.def_id()); // TODO: replace with get_attrs
         let nvvm_attrs = NvvmAttributes::parse(self, attrs);

crates/rustc_codegen_nvvm/src/mono_item.rs

@@ -90,7 +90,7 @@ impl<'ll, 'tcx> PreDefineCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> {
         attributes::from_fn_attrs(self, lldecl, instance);
 
         let def_id = instance.def_id();
-        let attrs = self.tcx.get_attrs_unchecked(def_id); // FIXME(jorge): Replace with get_attrs
+        let attrs = self.tcx.get_attrs_unchecked(def_id); // TODO: Replace with get_attrs
         let nvvm_attrs = NvvmAttributes::parse(self, attrs);
 
         unsafe {

examples/cuda/gpu/add_gpu/src/lib.rs

@@ -1,5 +1,3 @@
-#![cfg_attr(target_os = "cuda", no_std, register_attr(nvvm_internal))]
-
 use cuda_std::prelude::*;
 
 #[kernel]