[Mlir-commits] [mlir] [MLIR][LLVM] Make DISubprogramAttr cyclic (PR #106571)

[Billy Zhu]

================
@@ -577,13 +579,28 @@ def LLVM_DISubprogramAttr : LLVM_Attr<"DISubprogram", "di_subprogram",
       "ArrayRef<DINodeAttr>":$retainedNodes
     ), [{
       MLIRContext *ctx = file.getContext();
-      return $_get(ctx, id, compileUnit, scope, StringAttr::get(ctx, name),
+      return $_get(ctx, id, /*recId=*/nullptr, compileUnit, scope,
+                   StringAttr::get(ctx, name),
                    StringAttr::get(ctx, linkageName), file, line,
                    scopeLine, subprogramFlags, type, retainedNodes);
     }]>
   ];
-
   let assemblyFormat = "`<` struct(params) `>`";
+  let extraClassDeclaration = [{
+    /// Requirements of DIRecursiveTypeAttrInterface.
+    /// @{
+
+    /// Get whether this attr describes a recursive self reference.
+    bool isRecSelf() { return !getScope(); }
----------------
zyx-billy wrote:

My only slight concern is if it's possible to see a DISubprogram without a scope naturally (it looks like it's allowed to be null [here](https://github.com/llvm/llvm-project/blob/9ce4af5cadc24060f3c3674e01902d374afea983/llvm/lib/IR/Verifier.cpp#L1443) but I'm not sure if it's illegal/impractical elsewhere...). If it's not supposed to happen, can we enforce during import that all llvm's DISubprograms have a scope?

The concern is if we categorize a node as RecSelf when it isn't, we won't be able to tell the outer real node from the inner recursive reference (which triggers an assertion during export).

https://github.com/llvm/llvm-project/pull/106571