[Mlir-commits] [mlir] [MLIR][LLVM] Support Recursive DITypes (PR #80251)

[Tobias Gysi]

================
@@ -200,22 +208,66 @@ DebugTranslation::translateImpl(DIGlobalVariableAttr attr) {
       attr.getIsDefined(), nullptr, nullptr, attr.getAlignInBits(), nullptr);
 }
 
+llvm::DIType *DebugTranslation::translateImpl(DIRecursiveTypeAttr attr) {
+  if (attr.isRecSelf()) {
+    auto *iter = recursiveTypeMap.find(attr.getId());
+    assert(iter != recursiveTypeMap.end() && "unbound DI recursive self type");
+    return iter->second;
+  }
+
+  size_t recursiveStackSize = recursiveTypeMap.size();
+  auto setRecursivePlaceholderFn = [&](llvm::DIType *node) {
+    auto [iter, inserted] = recursiveTypeMap.try_emplace(attr.getId(), node);
+    assert(inserted && "illegal reuse of recursive id");
+  };
+
+  llvm::DIType *node =
+      TypeSwitch<DITypeAttr, llvm::DIType *>(attr.getBaseType())
+          .Case<DICompositeTypeAttr>([&](auto attr) {
+            return translateImpl(attr, setRecursivePlaceholderFn);
----------------
gysit wrote:

Did you consider to split the translateImpl in two function calls? E.g. one that computes the paceholder type without elements and one that adds the translated elements? The translateImpl could then use these two helper functions without the callback, while this function could call the two helper functions and in-between store the place holder on the stack. I hope this is understandable/makes sense I am starting to get tired...

It feels like this is a bit more direct than the extra setRecursivePlaceholderFn parameter. 

https://github.com/llvm/llvm-project/pull/80251