[Mlir-commits] [mlir] [MLIR][GPU-LLVM] Add in-pass signature update option for opencl kernels (PR #105664)

[Petr Kurapov]

================
@@ -515,3 +517,15 @@ gpu.module @kernels {
     gpu.return
   }
 }
+
+// -----
+
+gpu.module @kernels {
+// OPENCL-LABEL:   llvm.func spir_funccc @no_address_spaces(
+// OPENCL-SAME:                                             %{{[a-zA-Z_][a-zA-Z0-9_]*}}: !llvm.ptr<1>
+// OPENCL-SAME:                                             %{{[a-zA-Z_][a-zA-Z0-9_]*}}: !llvm.ptr<1>
+// OPENCL-SAME:                                             %{{[a-zA-Z_][a-zA-Z0-9_]*}}: !llvm.ptr<1>
+  gpu.func @no_address_spaces(%arg0: memref<f32>, %arg1: memref<f32, #gpu.address_space<global>>, %arg2: memref<f32>) {
+    gpu.return
+  }
+}
----------------
kurapov-peter wrote:

I added an example. I had to add func to llvm and finalize memref to llvm into the pass so that the signature and loads/stores are lowered correctly.

The function should have a spir_func attribute though. I think it'd be easiest to just go through all of them after func to llvm and attach it to those that don't have it. Kernels will have those set correctly, so we can always assume it's not a kernel.

https://github.com/llvm/llvm-project/pull/105664