Auto merge of #135047 - Flakebi:amdgpu-kernel-cc, r=workingjubilee

Add gpu-kernel calling convention

The amdgpu-kernel calling convention was reverted in commit f6b21e90d1 (#120495 and https://github.com/rust-lang/rust-analyzer/pull/16463) due to inactivity in the amdgpu target.

Introduce a `gpu-kernel` calling convention that translates to `ptx_kernel` or `amdgpu_kernel`, depending on the target that rust compiles for.

Tracking issue: #135467
amdgpu target tracking issue: #135024

compiler/rustc_codegen_llvm/src/abi.rs

@@ -1,3 +1,4 @@
+use std::borrow::Borrow;
 use std::cmp;
 
 use libc::c_uint;
@@ -312,7 +313,7 @@ impl<'ll, 'tcx> ArgAbiBuilderMethods<'tcx> for Builder<'_, 'll, 'tcx> {
 pub(crate) trait FnAbiLlvmExt<'ll, 'tcx> {
     fn llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
     fn ptr_to_llvm_type(&self, cx: &CodegenCx<'ll, 'tcx>) -> &'ll Type;
-    fn llvm_cconv(&self) -> llvm::CallConv;
+    fn llvm_cconv(&self, cx: &CodegenCx<'ll, 'tcx>) -> llvm::CallConv;
 
     /// Apply attributes to a function declaration/definition.
     fn apply_attrs_llfn(
@@ -404,8 +405,8 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
         cx.type_ptr_ext(cx.data_layout().instruction_address_space)
     }
 
-    fn llvm_cconv(&self) -> llvm::CallConv {
-        self.conv.into()
+    fn llvm_cconv(&self, cx: &CodegenCx<'ll, 'tcx>) -> llvm::CallConv {
+        llvm::CallConv::from_conv(self.conv, cx.tcx.sess.target.arch.borrow())
     }
 
     fn apply_attrs_llfn(
@@ -617,7 +618,7 @@ impl<'ll, 'tcx> FnAbiLlvmExt<'ll, 'tcx> for FnAbi<'tcx, Ty<'tcx>> {
             }
         }
 
-        let cconv = self.llvm_cconv();
+        let cconv = self.llvm_cconv(&bx.cx);
         if cconv != llvm::CCallConv {
             llvm::SetInstructionCallConv(callsite, cconv);
         }
@@ -655,8 +656,8 @@ impl<'tcx> AbiBuilderMethods<'tcx> for Builder<'_, '_, 'tcx> {
     }
 }
 
-impl From<Conv> for llvm::CallConv {
-    fn from(conv: Conv) -> Self {
+impl llvm::CallConv {
+    pub fn from_conv(conv: Conv, arch: &str) -> Self {
         match conv {
             Conv::C
             | Conv::Rust
@@ -666,6 +667,15 @@ impl From<Conv> for llvm::CallConv {
             Conv::Cold => llvm::ColdCallConv,
             Conv::PreserveMost => llvm::PreserveMost,
             Conv::PreserveAll => llvm::PreserveAll,
+            Conv::GpuKernel => {
+                if arch == "amdgpu" {
+                    llvm::AmdgpuKernel
+                } else if arch == "nvptx64" {
+                    llvm::PtxKernel
+                } else {
+                    panic!("Architecture {arch} does not support GpuKernel calling convention");
+                }
+            }
             Conv::AvrInterrupt => llvm::AvrInterrupt,
             Conv::AvrNonBlockingInterrupt => llvm::AvrNonBlockingInterrupt,
             Conv::ArmAapcs => llvm::ArmAapcsCallConv,

compiler/rustc_codegen_llvm/src/context.rs

@@ -741,7 +741,10 @@ impl<'ll, 'tcx> MiscCodegenMethods<'tcx> for CodegenCx<'ll, 'tcx> {
         if self.get_declared_value(entry_name).is_none() {
             Some(self.declare_entry_fn(
                 entry_name,
-                self.sess().target.entry_abi.into(),
+                llvm::CallConv::from_conv(
+                    self.sess().target.entry_abi,
+                    self.sess().target.arch.borrow(),
+                ),
                 llvm::UnnamedAddr::Global,
                 fn_type,
             ))

compiler/rustc_codegen_llvm/src/declare.rs

@@ -125,7 +125,7 @@ impl<'ll, 'tcx> CodegenCx<'ll, 'tcx> {
         let llfn = declare_raw_fn(
             self,
             name,
-            fn_abi.llvm_cconv(),
+            fn_abi.llvm_cconv(self),
             llvm::UnnamedAddr::Global,
             llvm::Visibility::Default,
             fn_abi.llvm_type(self),

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -120,6 +120,7 @@ pub enum CallConv {
     X86_Intr = 83,
     AvrNonBlockingInterrupt = 84,
     AvrInterrupt = 85,
+    AmdgpuKernel = 91,
 }
 
 /// Must match the layout of `LLVMLinkage`.