[PATCH] D107696: [CodeComplete] Basic code completion for attribute names.

[Aaron Ballman via Phabricator via cfe-commits]

aaron.ballman added inline comments.


================
Comment at: clang/lib/Sema/SemaCodeComplete.cpp:4390
+  auto AddCompletions = [&](const ParsedAttrInfo &A) {
+    if (A.IsTargetSpecific && !A.existsInTarget(Context.getTargetInfo()))
+      return;
----------------
Should we also early return if the attribute is ignored? (See `IgnoredAttr` in `Attr.td`) I'm assuming that attributes which do nothing are unlikely to be high-value attributes to autocomplete (so maybe they'd go at the end of the list if we wanted to keep them).


================
Comment at: clang/lib/Sema/SemaCodeComplete.cpp:4436-4446
+        llvm::SmallString<32> Guarded;
+        if (!Scope.empty()) {
+          const char *GuardedScope = underscoreAttrScope(Scope);
+          if (!GuardedScope)
+            continue;
+          Guarded.append(GuardedScope);
+          Guarded.append("::");
----------------
This could probably use a `Twine` and save some typing, but I don't insist.


================
Comment at: clang/lib/Sema/SemaCodeComplete.cpp:4357
+        continue;
+      llvm::StringRef Name = S.NormalizedFullName;
+      if (Completion == AttributeCompletion::Scope) {
----------------
sammccall wrote:
> aaron.ballman wrote:
> > sammccall wrote:
> > > aaron.ballman wrote:
> > > > Should we also add some special handling for attributes optionally starting with double underscores? e.g., `__attribute__((const))` and `__attribute__((__const__))`` are both equally useful to complete.
> > > > 
> > > > Also, we should add some explicit testing for `omp::sequence` and `omp::directive` as those are handled very specially and won't appear in the parsed attribute map. I think the OpenMP code completion would be super useful, but can be handled in a follow-up if you want.
> > > > Should we also add some special handling for attributes optionally starting with double underscores?
> > > 
> > > I think so. Two questions:
> > >  - Do I understand right that this is "just" a matter of adding leading/trailing `__` as a second option, for AS_GNU?
> > >  - are there similar rules for other syntaxes I've missed?
> > > 
> > > Offering both seems potentially confusing for users who don't care (especially in the case of const!). But I guess enough users will care about macros. At least in clangd the underscore versions will get ranked lower for having "internal" names though.
> > > 
> > > FWIW The no-underscores version appears to be ~3x more common (87k vs 27k hits in third-party code in google's repo). Among headers, no-underscores is "only" 2x more common (40k vs 21k). 
> > > 
> > > ---
> > > 
> > > > Also, we should add some explicit testing for omp::sequence and omp::directive as those are handled very specially and won't appear in the parsed attribute map. 
> > > 
> > > Yeah, I punted on these because it seems they will need special case logic, I'll add some tests that they don't do anything.
> > > Do I understand right that this is "just" a matter of adding leading/trailing __ as a second option, for AS_GNU?
> > > are there similar rules for other syntaxes I've missed?
> > 
> > Clang supports GNU attributes in either `__attribute__((foo))` or `__attribute__((__foo__))` forms. So I'd say that autocompleting after the second `(` should either suggest attributes (preferred) or `__` (for the poor folks writing libraries). If the user wants to autocomplete after `__attribute__((__`, I think it should suggest `foo__` as the rest of the attribute name. (Basically, if the user looks like they want underscores, give them all the underscores.)
> > 
> > Clang also supports `[[]]` attributes but with somewhat different rules. We support `[[gnu::attr]]`, `[[__gnu__::attr]]`, `[[gnu::__attr__]]`, and `[[__gnu__::__attr__]]` for GCC attributes. We support `[[clang::attr]]`, `[[_Clang::attr]]`, `[[clang::__attr__]]`, and `[[_Clang::__attr__]]` for Clang attributes. For vendors other than Clang and GCC, we don't support any additional underscores for either the scope or the attribute name. I would say that if the user asked for underscores for the vendor scope, they likely want the underscores for the attribute as well.
> > 
> > I suppose there's a third case. That horrible `using` syntax that I've never really seen used in the wild. e.g., ``[[using clang: attr]``. We do support the underscore behavior there as well.
> > 
> > > Offering both seems potentially confusing for users who don't care (especially in the case of const!). But I guess enough users will care about macros.
> > 
> > Yeah, users who are writing portable libraries are far more likely to care than users writing typical application code.
> > So I'd say that autocompleting after the second ( should either suggest attributes (preferred) or __ (for the poor folks writing libraries).
> 
> This is clever, unfortunately it doesn't really work for other reasons:
>  - LSP allows clients to "narrow down" results as an identifier is typed with client-side filtering, and `_` is treated as an identifier character.
>  - clang's code completion similarly returns all matches and leaves partial-identifier-filtering them to the CodeCompleteConsumer
> 
> I can't see a better plan than including both the underscore and non-underscore versions, with the latter downranked. Doesn't seem *too* bad.
> 
> > Clang also supports [[]] attributes...
> 
> Yeah, I think what we should do here is pretend the scoped namespace contains only the scoped attribute and vice versa:
>  - When we want qualified attributes, emit `gnu::attr` and `__gnu__::__attr__` but no mixing
>  - When we want scopes (after `using`), emit `gnu` and `__gnu__`
>  - When we want unqualified attributes (after `ns::` or `using ns:`), emit either `attr` or `__attr__` depending on the scope
> 
> > That horrible using syntax
> 
> I really don't understand why this is worthwhile. Maybe feature parity with XMLNS? :-)
> In any case, it happens to be easy to handle.
> I can't see a better plan than including both the underscore and non-underscore versions, with the latter downranked. Doesn't seem *too* bad.

Seems reasonable enough to me!

> Yeah, I think what we should do here is pretend the scoped namespace contains only the scoped attribute and vice versa:

+1, that seems good to me.

> I really don't understand why this is worthwhile. 

IMHO, it absolutely wasn't worthwhile. :-D


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D107696/new/

https://reviews.llvm.org/D107696