[Mlir-commits] [mlir] [mlir][GPUToNVVM] Fix memref function args/results (PR #96392)

[Guray Ozen]

================
@@ -679,6 +680,28 @@ gpu.module @gpumodule {
 }
 
 
+gpu.module @test_module_34 {
+  // CHECK-LABEL: llvm.func @memref_signature(
+  //  CHECK-SAME:     %{{.*}}: !llvm.ptr, %{{.*}}: !llvm.ptr, %{{.*}}: i64, %{{.*}}: i64, %{{.*}}: i64, %{{.*}}: f32) -> !llvm.struct<(struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>, f32)>
+  //       CHECK:   llvm.mlir.undef
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.mlir.undef
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.insertvalue
+  //       CHECK:   llvm.return
+
+  // CHECK-BARE-LABEL: llvm.func @memref_signature(
+  //  CHECK-BARE-SAME:     %{{.*}}: !llvm.ptr, %{{.*}}: f32) -> !llvm.struct<(ptr, f32)>
+  gpu.func @memref_signature(%m: memref<2xf32>, %f: f32) -> (memref<2xf32>, f32) {
+    gpu.return %m, %f : memref<2xf32>, f32
----------------
grypp wrote:

This example is nice. I think here gpu.func is a regular device function, so it can return. However, a GPU kernel cannot return anything. For nvidia, we differentiate kernel when nvvm.kernel attribute is present. I'm not sure how does it work for other models. 
Does it make sense to add a verifier for that? What do you think?

https://github.com/llvm/llvm-project/pull/96392