Export kernel descriptor for amdgpu kernels

The host runtime (HIP or HSA) expects a kernel descriptor object for
each kernel in the ELF file. The amdgpu LLVM backend generates the
object. It is created as a symbol with the name of the kernel plus a
`.kd` suffix.

Add it to the exported symbols in the linker script, so that it can be
found.

Author: Flakebi

compiler/rustc_codegen_ssa/src/back/linker.rs

@@ -1776,6 +1776,7 @@ fn exported_symbols_for_non_proc_macro(tcx: TyCtxt<'_>, crate_type: CrateType) -
             symbols.push(symbol_export::exporting_symbol_name_for_instance_in_crate(
                 tcx, symbol, cnum,
             ));
+            symbol_export::extend_exported_symbols(&mut symbols, tcx, symbol, cnum);
         }
     });
 

compiler/rustc_codegen_ssa/src/back/symbol_export.rs

@@ -10,9 +10,10 @@ use rustc_middle::middle::exported_symbols::{
     ExportedSymbol, SymbolExportInfo, SymbolExportKind, SymbolExportLevel, metadata_symbol_name,
 };
 use rustc_middle::query::LocalCrate;
-use rustc_middle::ty::{self, GenericArgKind, GenericArgsRef, Instance, SymbolName, TyCtxt};
+use rustc_middle::ty::{self, GenericArgKind, GenericArgsRef, Instance, SymbolName, Ty, TyCtxt};
 use rustc_middle::util::Providers;
 use rustc_session::config::{CrateType, OomStrategy};
+use rustc_target::callconv::Conv;
 use rustc_target::spec::{SanitizerSet, TlsModel};
 use tracing::debug;
 
@@ -551,6 +552,42 @@ pub(crate) fn symbol_name_for_instance_in_crate<'tcx>(
     }
 }
 
+fn calling_convention_for_symbol<'tcx>(
+    tcx: TyCtxt<'tcx>,
+    symbol: ExportedSymbol<'tcx>,
+) -> (Conv, &'tcx [rustc_target::callconv::ArgAbi<'tcx, Ty<'tcx>>]) {
+    let instance = match symbol {
+        ExportedSymbol::NonGeneric(def_id) | ExportedSymbol::Generic(def_id, _)
+            if tcx.is_static(def_id) =>
+        {
+            None
+        }
+        ExportedSymbol::NonGeneric(def_id) => Some(Instance::mono(tcx, def_id)),
+        ExportedSymbol::Generic(def_id, args) => Some(Instance::new(def_id, args)),
+        // DropGlue always use the Rust calling convention and thus follow the target's default
+        // symbol decoration scheme.
+        ExportedSymbol::DropGlue(..) => None,
+        // AsyncDropGlueCtorShim always use the Rust calling convention and thus follow the
+        // target's default symbol decoration scheme.
+        ExportedSymbol::AsyncDropGlueCtorShim(..) => None,
+        // NoDefId always follow the target's default symbol decoration scheme.
+        ExportedSymbol::NoDefId(..) => None,
+        // ThreadLocalShim always follow the target's default symbol decoration scheme.
+        ExportedSymbol::ThreadLocalShim(..) => None,
+    };
+
+    instance
+        .map(|i| {
+            tcx.fn_abi_of_instance(
+                ty::TypingEnv::fully_monomorphized().as_query_input((i, ty::List::empty())),
+            )
+            .unwrap_or_else(|_| bug!("fn_abi_of_instance({i:?}) failed"))
+        })
+        .map(|fnabi| (fnabi.conv, &fnabi.args[..]))
+        // FIXME(workingjubilee): why don't we know the convention here?
+        .unwrap_or((Conv::Rust, &[]))
+}
+
 /// This is the symbol name of the given instance as seen by the linker.
 ///
 /// On 32-bit Windows symbols are decorated according to their calling conventions.
@@ -559,8 +596,6 @@ pub(crate) fn linking_symbol_name_for_instance_in_crate<'tcx>(
     symbol: ExportedSymbol<'tcx>,
     instantiating_crate: CrateNum,
 ) -> String {
-    use rustc_target::callconv::Conv;
-
     let mut undecorated = symbol_name_for_instance_in_crate(tcx, symbol, instantiating_crate);
 
     // thread local will not be a function call,
@@ -584,35 +619,7 @@ pub(crate) fn linking_symbol_name_for_instance_in_crate<'tcx>(
         _ => return undecorated,
     };
 
-    let instance = match symbol {
-        ExportedSymbol::NonGeneric(def_id) | ExportedSymbol::Generic(def_id, _)
-            if tcx.is_static(def_id) =>
-        {
-            None
-        }
-        ExportedSymbol::NonGeneric(def_id) => Some(Instance::mono(tcx, def_id)),
-        ExportedSymbol::Generic(def_id, args) => Some(Instance::new(def_id, args)),
-        // DropGlue always use the Rust calling convention and thus follow the target's default
-        // symbol decoration scheme.
-        ExportedSymbol::DropGlue(..) => None,
-        // AsyncDropGlueCtorShim always use the Rust calling convention and thus follow the
-        // target's default symbol decoration scheme.
-        ExportedSymbol::AsyncDropGlueCtorShim(..) => None,
-        // NoDefId always follow the target's default symbol decoration scheme.
-        ExportedSymbol::NoDefId(..) => None,
-        // ThreadLocalShim always follow the target's default symbol decoration scheme.
-        ExportedSymbol::ThreadLocalShim(..) => None,
-    };
-
-    let (conv, args) = instance
-        .map(|i| {
-            tcx.fn_abi_of_instance(
-                ty::TypingEnv::fully_monomorphized().as_query_input((i, ty::List::empty())),
-            )
-            .unwrap_or_else(|_| bug!("fn_abi_of_instance({i:?}) failed"))
-        })
-        .map(|fnabi| (fnabi.conv, &fnabi.args[..]))
-        .unwrap_or((Conv::Rust, &[]));
+    let (conv, args) = calling_convention_for_symbol(tcx, symbol);
 
     // Decorate symbols with prefixes, suffixes and total number of bytes of arguments.
     // Reference: https://docs.microsoft.com/en-us/cpp/build/reference/decorated-names?view=msvc-170
@@ -644,6 +651,27 @@ pub(crate) fn exporting_symbol_name_for_instance_in_crate<'tcx>(
     maybe_emutls_symbol_name(tcx, symbol, &undecorated).unwrap_or(undecorated)
 }
 
+/// On amdhsa, `gpu-kernel` functions have an associated metadata object with a `.kd` suffix.
+/// Add it to the symbols list for all kernel functions, so that it is exported in the linked
+/// object.
+pub(crate) fn extend_exported_symbols<'tcx>(
+    symbols: &mut Vec<String>,
+    tcx: TyCtxt<'tcx>,
+    symbol: ExportedSymbol<'tcx>,
+    instantiating_crate: CrateNum,
+) {
+    let (conv, _) = calling_convention_for_symbol(tcx, symbol);
+
+    if conv != Conv::GpuKernel || tcx.sess.target.os != "amdhsa" {
+        return;
+    }
+
+    let undecorated = symbol_name_for_instance_in_crate(tcx, symbol, instantiating_crate);
+
+    // Add the symbol for the kernel descriptor (with .kd suffix)
+    symbols.push(format!("{undecorated}.kd"));
+}
+
 fn maybe_emutls_symbol_name<'tcx>(
     tcx: TyCtxt<'tcx>,
     symbol: ExportedSymbol<'tcx>,

tests/run-make/amdgpu-kd/foo.rs

@@ -0,0 +1,11 @@
+#![allow(internal_features)]
+#![feature(no_core, lang_items, abi_gpu_kernel)]
+#![no_core]
+#![no_std]
+
+// This is needed because of #![no_core]:
+#[lang = "sized"]
+trait Sized {}
+
+#[no_mangle]
+extern "gpu-kernel" fn kernel() {}

tests/run-make/amdgpu-kd/rmake.rs

@@ -0,0 +1,20 @@
+// ON the amdhsa OS, the host runtime (HIP or HSA) expects a kernel descriptor object for each
+// kernel in the ELF file. The amdgpu LLVM backend generates the object. It is created as a symbol
+// with the name of the kernel plus a .kd suffix.
+// Check that the produced object has the .kd symbol exported.
+
+//@ needs-llvm-components: amdgpu
+//@ needs-rust-lld
+
+use run_make_support::{llvm_readobj, rustc};
+
+fn main() {
+    rustc()
+        .crate_name("foo")
+        .target("amdgcn-amd-amdhsa")
+        .arg("-Ctarget-cpu=gfx900")
+        .crate_type("cdylib")
+        .input("foo.rs")
+        .run();
+    llvm_readobj().input("foo.elf").symbols().run().assert_stdout_contains("kernel.kd");
+}