[Mlir-commits] [mlir] [MLIR][LLVM] Recursion importer handle repeated self-references (PR #87295)

[Tobias Gysi]

================
@@ -423,3 +423,31 @@ llvm.mlir.global @global_variable() {dbg_expr = #di_global_variable_expression}
 // CHECK: ![[SCOPE]] = !DISubprogram({{.*}}type: ![[SUBROUTINE:[0-9]+]],
 // CHECK: ![[SUBROUTINE]] = !DISubroutineType(types: ![[SR_TYPES:[0-9]+]])
 // CHECK: ![[SR_TYPES]] = !{![[COMP]]}
+
+// -----
+
+// Ensures nested recursive decls work.
+// The output should be identical to if the inner composite type decl was
+// replaced with the recursive self reference.
+
+#di_file = #llvm.di_file<"test.mlir" in "/">
+#di_composite_type_self = #llvm.di_composite_type<tag = DW_TAG_null, recId = distinct[0]<>>
+
+#di_subroutine_type_inner = #llvm.di_subroutine_type<types = #di_composite_type_self>
+#di_subprogram_inner = #llvm.di_subprogram<scope = #di_file, file = #di_file, subprogramFlags = Optimized, type = #di_subroutine_type_inner>
+#di_composite_type_inner = #llvm.di_composite_type<tag = DW_TAG_class_type, recId = distinct[0]<>, scope = #di_subprogram_inner>
+
+#di_subroutine_type = #llvm.di_subroutine_type<types = #di_composite_type_inner>
----------------
gysit wrote:

This is not relevant for this revision, but I wonder if `#di_composite_type_inner` could be replaced by `#di_composite_type_self` in a second traversal after the import. That should allow us to prune all redundancies and thanks the having the full translation stack seems potentially easier than during the import? It may also not be to bad performance-wise since the redundant subtrees anyways come from the cache if I understand correctly.

https://github.com/llvm/llvm-project/pull/87295