[clang] [clang-tools-extra] [clangd] Autocomplete fixes for methods (PR #165916)

[Nathan Ridge via cfe-commits]

HighCommander4 wrote:

Thanks!

One question: is the new flag `CodeCompletionResult::IsAddressOfOperand` giving us anything that setting `DeclaringEntity=false` and `FunctionCanBeCall=false` wouldn't?

My other high-level piece of feedback is that I think we could implement some of this functionality more cleanly at the time a `CodeCompletionString` is constructed. Specifically, in [`CodeCompletionResult::createCodeCompletionStringForDecl`](https://searchfox.org/llvm/rev/b397c9d24196fe4103ad7cfe6bbfdfc08496364b/clang/lib/Sema/SemaCodeComplete.cpp#3748), we have access to flags such as `DeclaringEntity` and `FunctionCanBeCall` (they are fields of `this`), and we can pass them in to helpers such as `AddFunctionParameterChunks` and `AddFunctionTypeQualsToCompletionString` and let them influence the string chunks directly. For example:

 * If `DeclaringEntity=true`, instead of adding `CK_Placeholder` chunks for function parameter, we can create `CK_Text` chunks.
 * For parameters with default values, instead of adding the parameter and default value as a single chunk, we can add them as two separate chunks, with the default value being `CK_Informative`.
    * Clangd puts `CK_Informative` chunks into the signature but not the snippet, which is what we want.
 * For the function qualifiers, we can add them as `CK_Text` if `DeclaringEntity=true`, and otherwise as `CK_Informative`. (We don't need to introduce the new `CK_FunctionQualifier` chunk kind.)

With this approach, most (all?) changes to the clangd `getSignature` function should become unnecessary, and other consumers of the Sema code completion API (e.g. libclang) get to benefit from our improvements as well.

WDYT?

https://github.com/llvm/llvm-project/pull/165916