[clang] [llvm] [Clang-Repl] Introduce Error Recovery for Clang-Repl. (PR #223305)
Jared Wyles via cfe-commits
cfe-commits at lists.llvm.org
Fri Oct 30 02:48:09 PDT 2026
================
@@ -54,6 +54,7 @@
namespace clang {
class ASTContext;
+class DeclStateReverter;
----------------
jaredwy wrote:
> Are you talking about PTUMutationAction or DeclStateReverter?
Both. It's the same problem in each, so start with `DeclStateReverter`.
I think you and Vassil agree, you just don't see it yet.
> It seems to be explicitly check and dispatching on decl kind which seems suboptimal to me.
> DeclStateReverter mainly provides a few privileged operations on a declaration. In that sense, it is more similar to ASTDeclMerger, which provides specific functionality that ASTDeclReader can use.
That's the same position. ASTDeclMerger has no kind dispatch in it, it's typed primitives (`mergeRedeclarable`, `mergeTemplatePattern`, ...). ASTDeclReader is the DeclVisitor and it picks the primitive from `VisitFunctionDecl` / `VisitCXXRecordDecl`. So "like ASTDeclMerger" means pull the `dyn_cast` chains out of `revertDefinitionArrival`, `tryDetachRedeclChain` and `removeSpecializations` and let the caller dispatch. Right now those are three different partitions. One has `FieldDecl`, one keys on `TagDecl` / `NamespaceDecl`, neither matches DeclShape. That's what I was getting at with extending ASTDeclUnmerger instead of duplicating it. It also gives you this:
> Ideally, most of the declaration-reverting logic should live directly in PTUMutationActions::restoreX.
So the only gap is what the caller is, which is where I started.
> DeclShape is not the completely same as Decl::kind.
Override the six `VisitXDecl` you care about and hierarchy fallthrough gives you DeclShape, with a typed pointer instead of `cast<TYPE>(D)` off a stored tag.
> we are not performing one single operation for each declaration kind. Instead, there are separate operations, and each operation has its own dispatch.
That's one visitor per operation. Reader, Writer and Unmerger already work that way. I still don't see what the switch buys over that.
Once the reverter is Merger shaped the friend question answers itself. Name it for the primitives it provides, like Merger, not for commit/restore/unlink. The ledger is the new part and it's orthogonal, `ASTDeclUnmerger` can take it and keep its dispatch.
https://github.com/llvm/llvm-project/pull/223305
More information about the cfe-commits
mailing list