[PATCH] D53220: Remove possibility to change compile database path at runtime

[Simon Marchi via Phabricator via cfe-commits]

simark marked 7 inline comments as done.
simark added inline comments.


================
Comment at: clangd/ClangdLSPServer.cpp:433
 
     reparseOpenedFiles();
   }
----------------
sammccall wrote:
> sammccall wrote:
> > This isn't needed, the compilation database can only be set during initialization.
> It's still here... maybe forgot to upload a new diff?
> (Just to be clear: I meant `reparseOpenFiles` doesn't need to be called, as there are none.)
Oh I have only done the change locally, I have not uploaded a new diff yet.


================
Comment at: clangd/ClangdLSPServer.h:90
   void reparseOpenedFiles();
+  void applyConfiguration(const ClangdInitializationOptions &Settings);
   void applyConfiguration(const ClangdConfigurationParamsChange &Settings);
----------------
sammccall wrote:
> simark wrote:
> > sammccall wrote:
> > > Prefer a different name for this function - an overload set should have similar semantics and these are quite different (pseudo-constructor vs mutation allowed at any time).
> > Ok, I'll think about a better name.
> (In fact, this could also live directly in `onInitialize`, I think?)
Indeed, since it's things that can only be set during initialization.


================
Comment at: clangd/Protocol.h:422
+struct ClangdInitializationOptions : public ClangdConfigurationParamsChange {
+  llvm::Optional<std::string> compilationDatabasePath;
+};
----------------
sammccall wrote:
> simark wrote:
> > sammccall wrote:
> > > Can we just move this to InitializeParams as a clangd extension?
> > > Doing tricks with inheritance here makes the protocol harder to understand.
> > > Can we just move this to InitializeParams as a clangd extension?
> > 
> > Do you mean move it in the JSON, so it looks like this on the wire?
> > 
> > ```
> > {
> >   "method": "initialize",
> >   "params": {
> >     "compilationDatabasePath": "<path>",
> >     ...
> >   }
> > }
> > ```
> > 
> > instead of 
> > 
> > ```
> > {
> >   "method": "initialize",
> >   "params": {
> >     "initializationOptions": {
> >           "compilationDatabasePath": "<path>"
> >     },
> >     ...
> >   }
> > }
> > ```
> > 
> > ?
> > 
> > I think `initializationOptions` is a good place for this to be, I wouldn't change that..  If you don't like the inheritance, we can just get rid of it in our code and have two separate versions of the deserializing code.  We designed it so `didChangeConfiguration` and the initialization options would share the same structure, but it doesn't have to stay that way.
> > Do you mean move it in the JSON, so it looks like this on the wire?
> 
> Well, since you asked... :-) I'm not going to push hard for it (this patch is certainly a win already), but I do think that would be much clearer.
> 
> The current protocol has `InitializeParams` and `ClangdInitializationOptions`, and it's not clear semantically what the distinction between them is.
> 
> With hindsight, I think something like this would be easier to follow:
> ```
> // Clangd options that may be set at startup.
> struct InitializeParams {
>   // Clangd extension: override the path used to load the CDB.
>   Optional<string> compilationDatabasePath;
>   // Provides initial configuration as if by workspace/updateConfiguration.
>   Optional<ClangdConfigurationParamsChange> initialConfiguration;
> }
> // Clangd options that may be set dynamically at runtime.
> struct ClangdConfigurationParamsChange { ... }
> ```
> though even here, the benefit from being able to inline the initial configuration into the initalize message is unclear to me. The implementation has to support dynamic updates in any case, so why not make use of that?
> 
> > We designed it so didChangeConfiguration and the initialization options would share the same structure
> 
> This makes sense, but if they're diverging, I'm not sure that keeping them *mostly* the same brings more benefits than confusion.
> 
> ------
> 
> That said, if you prefer to keep the JSON as it is, that's fine. (If we grow more extensions, we may want to reorganize in future though?)
> My main concern is the use of inheritance here, and how it provides a default (configuration-change options can be provided at startup) that doesn't seem obviously correct and is hard to opt out of.
> The current protocol has InitializeParams and ClangdInitializationOptions, and it's not clear semantically what the distinction between them is.

`InitializeParams` is the type for the parameters of the `initialize` request.  In it, there is this field:

```
	/**
	 * User provided initialization options.
	 */
	initializationOptions?: any;
```

which is made to pass such language-server-specific options.  Since it's of the `any` type, we gave it a name, `ClangdInitializationOptions`.

> With hindsight, I think something like this would be easier to follow:
>
> [snippet]

I don't really understand why putting fields as extra fields in `InitializeParams` would be any easier than putting them in the object that exists specifically for that purpose.

> though even here, the benefit from being able to inline the initial configuration into the initalize message is unclear to me. The implementation has to support dynamic updates in any case, so why not make use of that?

I'd say, to avoid having the server start some work (like indexing some files using a certain set of flags) that will be invalidated when it receives some config changes moments later.

> This makes sense, but if they're diverging, I'm not sure that keeping them *mostly* the same brings more benefits than confusion.

I think you are exaggerating the confusion.  It is pretty straightforward: everything you can change during execution, you can also specify at initialize time.


Repository:
  rCTE Clang Tools Extra

https://reviews.llvm.org/D53220