[PATCH] D96717: [clangd] Bind outgoing calls through LSPBinder too. NFC

[Sam McCall via Phabricator via cfe-commits]

sammccall marked 2 inline comments as done.
sammccall added inline comments.


================
Comment at: clang-tools-extra/clangd/LSPBinder.h:88
+  template <typename Request, typename Response>
+  void outgoingMethod(llvm::StringLiteral Method,
+                      OutgoingMethod<Request, Response> &Handler) {
----------------
kadircet wrote:
> well this one filling an out-parameter definitely tripped me over while reading the code a couple times, and once more on the tests.
> 
> what about making this return the handler instead? E.g. `Edit = Bind.outgoingMethod<EditParams, EditResult>("edit")`, I know it is ugly that we duplicate template params on declaration + here now. Maybe we can get away by making `OutgoingMethod` a class with a call operator and a `RawOutgoing *Out` member that can be bound later on ? e.g:
> 
> ```
> template <StringLiteral Method, typename P, typename R>
> class OutgoingMethod {
>   RawOutgoing *Out = nullptr;
> public:
>   void operator()(const P& Params, Callback<R> CB) {
>     assert(Out && "Method haven't bound");
>     Out->callMethod(Method, JSON(Params), ....);
>   }
> };
> ```
> 
> then we either make LSPBinder a friend of OutgoingMethod and set Out through it, or the other way around and have a `OutgoingMethod::bindOut(LSPBinder&)` ?
> 
> not sure if it is any better though, as we still construct a somewhat invalid object :/
> well this one filling an out-parameter definitely tripped me over while reading the code a couple times, and once more on the tests.

Yup. My feeling is the wins are:
 - avoid writing the types over and over (we can't really avoid it on the instance variable)
 - resembles syntax for binding incoming things

And the main loss is definitely that we're not using assignment syntax for assignment.

> what about making this return the handler instead? E.g. Edit = Bind.outgoingMethod<EditParams, EditResult>("edit")

This is a reasonable alternative, as you say the duplication is the downside.

> Maybe we can get away by making OutgoingMethod a class with a call operator

Yeah, I don't really understand what this achieves - the usage looks the same, just now the logical assignment doesn't use assignment syntax *or* assignment semantics :-)

Other ideas I had:
 - the object is untyped, and the call operator is templated/typed - no type-safety at callsite, yuck
 - `outgoingMethod` returns an untyped proxy object with a template conversion operator to unique_function<Req, Rsp>. This gives us assignment syntax... and horrible compiler errors on mismatch.

I'm going to try the latter out to see how bad it is.


================
Comment at: clang-tools-extra/clangd/LSPBinder.h:93
+          Method, JSON(R),
+          // FIXME: why keep ctx alive but not restore it for the callback?
+          [Reply(std::move(Reply)), Ctx(Context::current().clone()),
----------------
kadircet wrote:
> haha feels like confusing `Context` and `WithContext` issue i talked about yesterday :D
Exactly. I'm fairly sure we need a WithContext here but don't want to risk changing it at the same time.


================
Comment at: clang-tools-extra/clangd/unittests/ClangdLSPServerTests.cpp:237
+      Value += X;
+      Changed(Value);
+    }
----------------
whoops, after rebasing against 40cc63ea6eec7874d3a358f9fa549ef2f6543512 this uncovered a bug - we're testing the public API from the wrong thread (server thread != main thread).

So I've had to remove that test in this patch. We can test public APIs again in D96755.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D96717/new/

https://reviews.llvm.org/D96717