[Lldb-commits] [lldb] [lldb] Enable caching for BytecodeSyntheticChildren::FrontEnd::Update (PR #181199)

[Michael Buch via lldb-commits]

================
@@ -291,10 +292,41 @@ lldb::ChildCacheState BytecodeSyntheticChildren::FrontEnd::Update() {
     return ChildCacheState::eRefetch;
   }
 
+  // If the top of the stack is a 0 or 1, that value is popped of the stack and
+  // treated as a return value of eRefetch or eReuse respectively. If the top of
+  // the stack is any other value, it stays on the stack and becomes part of
+  // `self`.
+  //
+  // This dynamic logic can lead to errors for synthetic formatter authors.
+  // Consider the case where an `update` implementation places the number of
+  // children last on the stack. LLDB will _sometimes_ (but not always) consume
+  // that value as a ChildCacheState value. This would cause downstream problems
+  // in `num_children`, because the count won't be on the stack.
+  //
+  // Bytecode authors are encouraged to explicitly push a ChildCacheState value
+  // on to the stack.
----------------
Michael137 wrote:

Would it be possible to make it a requirement? If the byte code will mostly be compiler-generated, we can enforce it there. And make the interpreter expect it? I guess that would be a formatter byte-code ABI break?

https://github.com/llvm/llvm-project/pull/181199