xref: /llvm-project/clang-tools-extra/clangd/CodeComplete.cpp (revision 030d48b7db9845f42bf3ef365193bfdbb23f5440)
1 //===--- CodeComplete.cpp ----------------------------------------*- C++-*-===//
2 //
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 // See https://llvm.org/LICENSE.txt for license information.
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 //
7 //===----------------------------------------------------------------------===//
8 //
9 // Code completion has several moving parts:
10 //  - AST-based completions are provided using the completion hooks in Sema.
11 //  - external completions are retrieved from the index (using hints from Sema)
12 //  - the two sources overlap, and must be merged and overloads bundled
13 //  - results must be scored and ranked (see Quality.h) before rendering
14 //
15 // Signature help works in a similar way as code completion, but it is simpler:
16 // it's purely AST-based, and there are few candidates.
17 //
18 //===----------------------------------------------------------------------===//
19 
20 #include "CodeComplete.h"
21 #include "AST.h"
22 #include "CodeCompletionStrings.h"
23 #include "Compiler.h"
24 #include "Config.h"
25 #include "ExpectedTypes.h"
26 #include "Feature.h"
27 #include "FileDistance.h"
28 #include "FuzzyMatch.h"
29 #include "Headers.h"
30 #include "Hover.h"
31 #include "Preamble.h"
32 #include "Protocol.h"
33 #include "Quality.h"
34 #include "SourceCode.h"
35 #include "URI.h"
36 #include "index/Index.h"
37 #include "index/Symbol.h"
38 #include "index/SymbolOrigin.h"
39 #include "support/Logger.h"
40 #include "support/Markup.h"
41 #include "support/Threading.h"
42 #include "support/ThreadsafeFS.h"
43 #include "support/Trace.h"
44 #include "clang/AST/Decl.h"
45 #include "clang/AST/DeclBase.h"
46 #include "clang/Basic/CharInfo.h"
47 #include "clang/Basic/LangOptions.h"
48 #include "clang/Basic/SourceLocation.h"
49 #include "clang/Basic/TokenKinds.h"
50 #include "clang/Format/Format.h"
51 #include "clang/Frontend/CompilerInstance.h"
52 #include "clang/Frontend/FrontendActions.h"
53 #include "clang/Lex/ExternalPreprocessorSource.h"
54 #include "clang/Lex/Lexer.h"
55 #include "clang/Lex/Preprocessor.h"
56 #include "clang/Lex/PreprocessorOptions.h"
57 #include "clang/Sema/CodeCompleteConsumer.h"
58 #include "clang/Sema/DeclSpec.h"
59 #include "clang/Sema/Sema.h"
60 #include "llvm/ADT/ArrayRef.h"
61 #include "llvm/ADT/SmallVector.h"
62 #include "llvm/ADT/StringExtras.h"
63 #include "llvm/ADT/StringRef.h"
64 #include "llvm/Support/Casting.h"
65 #include "llvm/Support/Compiler.h"
66 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/Error.h"
68 #include "llvm/Support/FormatVariadic.h"
69 #include "llvm/Support/ScopedPrinter.h"
70 #include <algorithm>
71 #include <iterator>
72 #include <limits>
73 #include <optional>
74 #include <utility>
75 
76 // We log detailed candidate here if you run with -debug-only=codecomplete.
77 #define DEBUG_TYPE "CodeComplete"
78 
79 namespace clang {
80 namespace clangd {
81 
82 #if CLANGD_DECISION_FOREST
83 const CodeCompleteOptions::CodeCompletionRankingModel
84     CodeCompleteOptions::DefaultRankingModel =
85         CodeCompleteOptions::DecisionForest;
86 #else
87 const CodeCompleteOptions::CodeCompletionRankingModel
88     CodeCompleteOptions::DefaultRankingModel = CodeCompleteOptions::Heuristics;
89 #endif
90 
91 namespace {
92 
93 // Note: changes to this function should also be reflected in the
94 // CodeCompletionResult overload where appropriate.
95 CompletionItemKind
96 toCompletionItemKind(index::SymbolKind Kind,
97                      const llvm::StringRef *Signature = nullptr) {
98   using SK = index::SymbolKind;
99   switch (Kind) {
100   case SK::Unknown:
101     return CompletionItemKind::Missing;
102   case SK::Module:
103   case SK::Namespace:
104   case SK::NamespaceAlias:
105     return CompletionItemKind::Module;
106   case SK::Macro:
107     // Use macro signature (if provided) to tell apart function-like and
108     // object-like macros.
109     return Signature && Signature->contains('(') ? CompletionItemKind::Function
110                                                  : CompletionItemKind::Constant;
111   case SK::Enum:
112     return CompletionItemKind::Enum;
113   case SK::Struct:
114     return CompletionItemKind::Struct;
115   case SK::Class:
116   case SK::Extension:
117   case SK::Union:
118     return CompletionItemKind::Class;
119   case SK::Protocol:
120     // Use interface instead of class for differentiation of classes and
121     // protocols with the same name (e.g. @interface NSObject vs. @protocol
122     // NSObject).
123     return CompletionItemKind::Interface;
124   case SK::TypeAlias:
125     // We use the same kind as the VSCode C++ extension.
126     // FIXME: pick a better option when we have one.
127     return CompletionItemKind::Interface;
128   case SK::Using:
129     return CompletionItemKind::Reference;
130   case SK::Function:
131   case SK::ConversionFunction:
132     return CompletionItemKind::Function;
133   case SK::Variable:
134   case SK::Parameter:
135   case SK::NonTypeTemplateParm:
136     return CompletionItemKind::Variable;
137   case SK::Field:
138     return CompletionItemKind::Field;
139   case SK::EnumConstant:
140     return CompletionItemKind::EnumMember;
141   case SK::InstanceMethod:
142   case SK::ClassMethod:
143   case SK::StaticMethod:
144   case SK::Destructor:
145     return CompletionItemKind::Method;
146   case SK::InstanceProperty:
147   case SK::ClassProperty:
148   case SK::StaticProperty:
149     return CompletionItemKind::Property;
150   case SK::Constructor:
151     return CompletionItemKind::Constructor;
152   case SK::TemplateTypeParm:
153   case SK::TemplateTemplateParm:
154     return CompletionItemKind::TypeParameter;
155   case SK::Concept:
156     return CompletionItemKind::Interface;
157   }
158   llvm_unreachable("Unhandled clang::index::SymbolKind.");
159 }
160 
161 // Note: changes to this function should also be reflected in the
162 // index::SymbolKind overload where appropriate.
163 CompletionItemKind toCompletionItemKind(const CodeCompletionResult &Res,
164                                         CodeCompletionContext::Kind CtxKind) {
165   if (Res.Declaration)
166     return toCompletionItemKind(index::getSymbolInfo(Res.Declaration).Kind);
167   if (CtxKind == CodeCompletionContext::CCC_IncludedFile)
168     return CompletionItemKind::File;
169   switch (Res.Kind) {
170   case CodeCompletionResult::RK_Declaration:
171     llvm_unreachable("RK_Declaration without Decl");
172   case CodeCompletionResult::RK_Keyword:
173     return CompletionItemKind::Keyword;
174   case CodeCompletionResult::RK_Macro:
175     // There is no 'Macro' kind in LSP.
176     // Avoid using 'Text' to avoid confusion with client-side word-based
177     // completion proposals.
178     return Res.MacroDefInfo && Res.MacroDefInfo->isFunctionLike()
179                ? CompletionItemKind::Function
180                : CompletionItemKind::Constant;
181   case CodeCompletionResult::RK_Pattern:
182     return CompletionItemKind::Snippet;
183   }
184   llvm_unreachable("Unhandled CodeCompletionResult::ResultKind.");
185 }
186 
187 // FIXME: find a home for this (that can depend on both markup and Protocol).
188 MarkupContent renderDoc(const markup::Document &Doc, MarkupKind Kind) {
189   MarkupContent Result;
190   Result.kind = Kind;
191   switch (Kind) {
192   case MarkupKind::PlainText:
193     Result.value.append(Doc.asPlainText());
194     break;
195   case MarkupKind::Markdown:
196     Result.value.append(Doc.asMarkdown());
197     break;
198   }
199   return Result;
200 }
201 
202 Symbol::IncludeDirective insertionDirective(const CodeCompleteOptions &Opts) {
203   if (!Opts.ImportInsertions || !Opts.MainFileSignals)
204     return Symbol::IncludeDirective::Include;
205   return Opts.MainFileSignals->InsertionDirective;
206 }
207 
208 // Identifier code completion result.
209 struct RawIdentifier {
210   llvm::StringRef Name;
211   unsigned References; // # of usages in file.
212 };
213 
214 /// A code completion result, in clang-native form.
215 /// It may be promoted to a CompletionItem if it's among the top-ranked results.
216 struct CompletionCandidate {
217   llvm::StringRef Name; // Used for filtering and sorting.
218   // We may have a result from Sema, from the index, or both.
219   const CodeCompletionResult *SemaResult = nullptr;
220   const Symbol *IndexResult = nullptr;
221   const RawIdentifier *IdentifierResult = nullptr;
222   llvm::SmallVector<SymbolInclude, 1> RankedIncludeHeaders;
223 
224   // Returns a token identifying the overload set this is part of.
225   // 0 indicates it's not part of any overload set.
226   size_t overloadSet(const CodeCompleteOptions &Opts, llvm::StringRef FileName,
227                      IncludeInserter *Inserter,
228                      CodeCompletionContext::Kind CCContextKind) const {
229     if (!Opts.BundleOverloads.value_or(false))
230       return 0;
231 
232     // Depending on the index implementation, we can see different header
233     // strings (literal or URI) mapping to the same file. We still want to
234     // bundle those, so we must resolve the header to be included here.
235     std::string HeaderForHash;
236     if (Inserter) {
237       if (auto Header = headerToInsertIfAllowed(Opts, CCContextKind)) {
238         if (auto HeaderFile = toHeaderFile(*Header, FileName)) {
239           if (auto Spelled =
240                   Inserter->calculateIncludePath(*HeaderFile, FileName))
241             HeaderForHash = *Spelled;
242         } else {
243           vlog("Code completion header path manipulation failed {0}",
244                HeaderFile.takeError());
245         }
246       }
247     }
248 
249     llvm::SmallString<256> Scratch;
250     if (IndexResult) {
251       switch (IndexResult->SymInfo.Kind) {
252       case index::SymbolKind::ClassMethod:
253       case index::SymbolKind::InstanceMethod:
254       case index::SymbolKind::StaticMethod:
255 #ifndef NDEBUG
256         llvm_unreachable("Don't expect members from index in code completion");
257 #else
258         [[fallthrough]];
259 #endif
260       case index::SymbolKind::Function:
261         // We can't group overloads together that need different #includes.
262         // This could break #include insertion.
263         return llvm::hash_combine(
264             (IndexResult->Scope + IndexResult->Name).toStringRef(Scratch),
265             HeaderForHash);
266       default:
267         return 0;
268       }
269     }
270     if (SemaResult) {
271       // We need to make sure we're consistent with the IndexResult case!
272       const NamedDecl *D = SemaResult->Declaration;
273       if (!D || !D->isFunctionOrFunctionTemplate())
274         return 0;
275       {
276         llvm::raw_svector_ostream OS(Scratch);
277         D->printQualifiedName(OS);
278       }
279       return llvm::hash_combine(Scratch, HeaderForHash);
280     }
281     assert(IdentifierResult);
282     return 0;
283   }
284 
285   bool contextAllowsHeaderInsertion(CodeCompletionContext::Kind Kind) const {
286     // Explicitly disable insertions for forward declarations since they don't
287     // reference the declaration.
288     if (Kind == CodeCompletionContext::CCC_ObjCClassForwardDecl)
289       return false;
290     return true;
291   }
292 
293   // The best header to include if include insertion is allowed.
294   std::optional<llvm::StringRef>
295   headerToInsertIfAllowed(const CodeCompleteOptions &Opts,
296                           CodeCompletionContext::Kind ContextKind) const {
297     if (Opts.InsertIncludes == CodeCompleteOptions::NeverInsert ||
298         RankedIncludeHeaders.empty() ||
299         !contextAllowsHeaderInsertion(ContextKind))
300       return std::nullopt;
301     if (SemaResult && SemaResult->Declaration) {
302       // Avoid inserting new #include if the declaration is found in the current
303       // file e.g. the symbol is forward declared.
304       auto &SM = SemaResult->Declaration->getASTContext().getSourceManager();
305       for (const Decl *RD : SemaResult->Declaration->redecls())
306         if (SM.isInMainFile(SM.getExpansionLoc(RD->getBeginLoc())))
307           return std::nullopt;
308     }
309     Symbol::IncludeDirective Directive = insertionDirective(Opts);
310     for (const auto &Inc : RankedIncludeHeaders)
311       if ((Inc.Directive & Directive) != 0)
312         return Inc.Header;
313     return std::nullopt;
314   }
315 
316   using Bundle = llvm::SmallVector<CompletionCandidate, 4>;
317 };
318 using ScoredBundle =
319     std::pair<CompletionCandidate::Bundle, CodeCompletion::Scores>;
320 struct ScoredBundleGreater {
321   bool operator()(const ScoredBundle &L, const ScoredBundle &R) {
322     if (L.second.Total != R.second.Total)
323       return L.second.Total > R.second.Total;
324     return L.first.front().Name <
325            R.first.front().Name; // Earlier name is better.
326   }
327 };
328 
329 // Remove the first template argument from Signature.
330 // If Signature only contains a single argument an empty string is returned.
331 std::string removeFirstTemplateArg(llvm::StringRef Signature) {
332   auto Rest = Signature.split(",").second;
333   if (Rest.empty())
334     return "";
335   return ("<" + Rest.ltrim()).str();
336 }
337 
338 // Assembles a code completion out of a bundle of >=1 completion candidates.
339 // Many of the expensive strings are only computed at this point, once we know
340 // the candidate bundle is going to be returned.
341 //
342 // Many fields are the same for all candidates in a bundle (e.g. name), and are
343 // computed from the first candidate, in the constructor.
344 // Others vary per candidate, so add() must be called for remaining candidates.
345 struct CodeCompletionBuilder {
346   CodeCompletionBuilder(ASTContext *ASTCtx, const CompletionCandidate &C,
347                         CodeCompletionString *SemaCCS,
348                         llvm::ArrayRef<std::string> AccessibleScopes,
349                         const IncludeInserter &Includes,
350                         llvm::StringRef FileName,
351                         CodeCompletionContext::Kind ContextKind,
352                         const CodeCompleteOptions &Opts,
353                         bool IsUsingDeclaration, tok::TokenKind NextTokenKind)
354       : ASTCtx(ASTCtx), ArgumentLists(Opts.ArgumentLists),
355         IsUsingDeclaration(IsUsingDeclaration), NextTokenKind(NextTokenKind) {
356     Completion.Deprecated = true; // cleared by any non-deprecated overload.
357     add(C, SemaCCS, ContextKind);
358     if (C.SemaResult) {
359       assert(ASTCtx);
360       Completion.Origin |= SymbolOrigin::AST;
361       Completion.Name = std::string(llvm::StringRef(SemaCCS->getTypedText()));
362       Completion.FilterText = SemaCCS->getAllTypedText();
363       if (Completion.Scope.empty()) {
364         if ((C.SemaResult->Kind == CodeCompletionResult::RK_Declaration) ||
365             (C.SemaResult->Kind == CodeCompletionResult::RK_Pattern))
366           if (const auto *D = C.SemaResult->getDeclaration())
367             if (const auto *ND = dyn_cast<NamedDecl>(D))
368               Completion.Scope = std::string(
369                   splitQualifiedName(printQualifiedName(*ND)).first);
370       }
371       Completion.Kind = toCompletionItemKind(*C.SemaResult, ContextKind);
372       // Sema could provide more info on whether the completion was a file or
373       // folder.
374       if (Completion.Kind == CompletionItemKind::File &&
375           Completion.Name.back() == '/')
376         Completion.Kind = CompletionItemKind::Folder;
377       for (const auto &FixIt : C.SemaResult->FixIts) {
378         Completion.FixIts.push_back(toTextEdit(
379             FixIt, ASTCtx->getSourceManager(), ASTCtx->getLangOpts()));
380       }
381       llvm::sort(Completion.FixIts, [](const TextEdit &X, const TextEdit &Y) {
382         return std::tie(X.range.start.line, X.range.start.character) <
383                std::tie(Y.range.start.line, Y.range.start.character);
384       });
385     }
386     if (C.IndexResult) {
387       Completion.Origin |= C.IndexResult->Origin;
388       if (Completion.Scope.empty())
389         Completion.Scope = std::string(C.IndexResult->Scope);
390       if (Completion.Kind == CompletionItemKind::Missing)
391         Completion.Kind = toCompletionItemKind(C.IndexResult->SymInfo.Kind,
392                                                &C.IndexResult->Signature);
393       if (Completion.Name.empty())
394         Completion.Name = std::string(C.IndexResult->Name);
395       if (Completion.FilterText.empty())
396         Completion.FilterText = Completion.Name;
397       // If the completion was visible to Sema, no qualifier is needed. This
398       // avoids unneeded qualifiers in cases like with `using ns::X`.
399       if (Completion.RequiredQualifier.empty() && !C.SemaResult) {
400         llvm::StringRef ShortestQualifier = C.IndexResult->Scope;
401         for (llvm::StringRef Scope : AccessibleScopes) {
402           llvm::StringRef Qualifier = C.IndexResult->Scope;
403           if (Qualifier.consume_front(Scope) &&
404               Qualifier.size() < ShortestQualifier.size())
405             ShortestQualifier = Qualifier;
406         }
407         Completion.RequiredQualifier = std::string(ShortestQualifier);
408       }
409     }
410     if (C.IdentifierResult) {
411       Completion.Origin |= SymbolOrigin::Identifier;
412       Completion.Kind = CompletionItemKind::Text;
413       Completion.Name = std::string(C.IdentifierResult->Name);
414       Completion.FilterText = Completion.Name;
415     }
416 
417     // Turn absolute path into a literal string that can be #included.
418     auto Inserted = [&](llvm::StringRef Header)
419         -> llvm::Expected<std::pair<std::string, bool>> {
420       auto ResolvedDeclaring =
421           URI::resolve(C.IndexResult->CanonicalDeclaration.FileURI, FileName);
422       if (!ResolvedDeclaring)
423         return ResolvedDeclaring.takeError();
424       auto ResolvedInserted = toHeaderFile(Header, FileName);
425       if (!ResolvedInserted)
426         return ResolvedInserted.takeError();
427       auto Spelled = Includes.calculateIncludePath(*ResolvedInserted, FileName);
428       if (!Spelled)
429         return error("Header not on include path");
430       return std::make_pair(
431           std::move(*Spelled),
432           Includes.shouldInsertInclude(*ResolvedDeclaring, *ResolvedInserted));
433     };
434     bool ShouldInsert =
435         C.headerToInsertIfAllowed(Opts, ContextKind).has_value();
436     Symbol::IncludeDirective Directive = insertionDirective(Opts);
437     // Calculate include paths and edits for all possible headers.
438     for (const auto &Inc : C.RankedIncludeHeaders) {
439       if ((Inc.Directive & Directive) == 0)
440         continue;
441 
442       if (auto ToInclude = Inserted(Inc.Header)) {
443         CodeCompletion::IncludeCandidate Include;
444         Include.Header = ToInclude->first;
445         if (ToInclude->second && ShouldInsert)
446           Include.Insertion = Includes.insert(
447               ToInclude->first, Directive == Symbol::Import
448                                     ? tooling::IncludeDirective::Import
449                                     : tooling::IncludeDirective::Include);
450         Completion.Includes.push_back(std::move(Include));
451       } else
452         log("Failed to generate include insertion edits for adding header "
453             "(FileURI='{0}', IncludeHeader='{1}') into {2}: {3}",
454             C.IndexResult->CanonicalDeclaration.FileURI, Inc.Header, FileName,
455             ToInclude.takeError());
456     }
457     // Prefer includes that do not need edits (i.e. already exist).
458     std::stable_partition(Completion.Includes.begin(),
459                           Completion.Includes.end(),
460                           [](const CodeCompletion::IncludeCandidate &I) {
461                             return !I.Insertion.has_value();
462                           });
463   }
464 
465   void add(const CompletionCandidate &C, CodeCompletionString *SemaCCS,
466            CodeCompletionContext::Kind ContextKind) {
467     assert(bool(C.SemaResult) == bool(SemaCCS));
468     Bundled.emplace_back();
469     BundledEntry &S = Bundled.back();
470     bool IsConcept = false;
471     if (C.SemaResult) {
472       getSignature(*SemaCCS, &S.Signature, &S.SnippetSuffix, C.SemaResult->Kind,
473                    C.SemaResult->CursorKind,
474                    /*IncludeFunctionArguments=*/C.SemaResult->FunctionCanBeCall,
475                    /*RequiredQualifiers=*/&Completion.RequiredQualifier);
476       S.ReturnType = getReturnType(*SemaCCS);
477       if (C.SemaResult->Kind == CodeCompletionResult::RK_Declaration)
478         if (const auto *D = C.SemaResult->getDeclaration())
479           if (isa<ConceptDecl>(D))
480             IsConcept = true;
481     } else if (C.IndexResult) {
482       S.Signature = std::string(C.IndexResult->Signature);
483       S.SnippetSuffix = std::string(C.IndexResult->CompletionSnippetSuffix);
484       S.ReturnType = std::string(C.IndexResult->ReturnType);
485       if (C.IndexResult->SymInfo.Kind == index::SymbolKind::Concept)
486         IsConcept = true;
487     }
488 
489     /// When a concept is used as a type-constraint (e.g. `Iterator auto x`),
490     /// and in some other contexts, its first type argument is not written.
491     /// Drop the parameter from the signature.
492     if (IsConcept && ContextKind == CodeCompletionContext::CCC_TopLevel) {
493       S.Signature = removeFirstTemplateArg(S.Signature);
494       // Dropping the first placeholder from the suffix will leave a $2
495       // with no $1.
496       S.SnippetSuffix = removeFirstTemplateArg(S.SnippetSuffix);
497     }
498 
499     if (!Completion.Documentation) {
500       auto SetDoc = [&](llvm::StringRef Doc) {
501         if (!Doc.empty()) {
502           Completion.Documentation.emplace();
503           parseDocumentation(Doc, *Completion.Documentation);
504         }
505       };
506       if (C.IndexResult) {
507         SetDoc(C.IndexResult->Documentation);
508       } else if (C.SemaResult) {
509         const auto DocComment = getDocComment(*ASTCtx, *C.SemaResult,
510                                               /*CommentsFromHeaders=*/false);
511         SetDoc(formatDocumentation(*SemaCCS, DocComment));
512       }
513     }
514     if (Completion.Deprecated) {
515       if (C.SemaResult)
516         Completion.Deprecated &=
517             C.SemaResult->Availability == CXAvailability_Deprecated;
518       if (C.IndexResult)
519         Completion.Deprecated &=
520             bool(C.IndexResult->Flags & Symbol::Deprecated);
521     }
522   }
523 
524   CodeCompletion build() {
525     Completion.ReturnType = summarizeReturnType();
526     Completion.Signature = summarizeSignature();
527     Completion.SnippetSuffix = summarizeSnippet();
528     Completion.BundleSize = Bundled.size();
529     return std::move(Completion);
530   }
531 
532 private:
533   struct BundledEntry {
534     std::string SnippetSuffix;
535     std::string Signature;
536     std::string ReturnType;
537   };
538 
539   // If all BundledEntries have the same value for a property, return it.
540   template <std::string BundledEntry::*Member>
541   const std::string *onlyValue() const {
542     auto B = Bundled.begin(), E = Bundled.end();
543     for (auto *I = B + 1; I != E; ++I)
544       if (I->*Member != B->*Member)
545         return nullptr;
546     return &(B->*Member);
547   }
548 
549   template <bool BundledEntry::*Member> const bool *onlyValue() const {
550     auto B = Bundled.begin(), E = Bundled.end();
551     for (auto *I = B + 1; I != E; ++I)
552       if (I->*Member != B->*Member)
553         return nullptr;
554     return &(B->*Member);
555   }
556 
557   std::string summarizeReturnType() const {
558     if (auto *RT = onlyValue<&BundledEntry::ReturnType>())
559       return *RT;
560     return "";
561   }
562 
563   std::string summarizeSnippet() const {
564     /// localize ArgumentLists tests for better readability
565     const bool None = ArgumentLists == Config::ArgumentListsPolicy::None;
566     const bool Open =
567         ArgumentLists == Config::ArgumentListsPolicy::OpenDelimiter;
568     const bool Delim = ArgumentLists == Config::ArgumentListsPolicy::Delimiters;
569     const bool Full =
570         ArgumentLists == Config::ArgumentListsPolicy::FullPlaceholders ||
571         (!None && !Open && !Delim); // <-- failsafe: Full is default
572 
573     if (IsUsingDeclaration)
574       return "";
575     auto *Snippet = onlyValue<&BundledEntry::SnippetSuffix>();
576     if (!Snippet)
577       // All bundles are function calls.
578       // FIXME(ibiryukov): sometimes add template arguments to a snippet, e.g.
579       // we need to complete 'forward<$1>($0)'.
580       return None ? "" : (Open ? "(" : "($0)");
581 
582     if (Snippet->empty())
583       return "";
584 
585     bool MayHaveArgList = Completion.Kind == CompletionItemKind::Function ||
586                           Completion.Kind == CompletionItemKind::Method ||
587                           Completion.Kind == CompletionItemKind::Constructor ||
588                           Completion.Kind == CompletionItemKind::Text /*Macro*/;
589     // If likely arg list already exists, don't add new parens & placeholders.
590     //   Snippet: function(int x, int y)
591     //   func^(1,2) -> function(1, 2)
592     //             NOT function(int x, int y)(1, 2)
593     if (MayHaveArgList) {
594       // Check for a template argument list in the code.
595       //   Snippet: function<class T>(int x)
596       //   fu^<int>(1) -> function<int>(1)
597       if (NextTokenKind == tok::less && Snippet->front() == '<')
598         return "";
599       // Potentially followed by regular argument list.
600       if (NextTokenKind == tok::l_paren) {
601         //   Snippet: function<class T>(int x)
602         //   fu^(1,2) -> function<class T>(1, 2)
603         if (Snippet->front() == '<') {
604           // Find matching '>', handling nested brackets.
605           int Balance = 0;
606           size_t I = 0;
607           do {
608             if (Snippet->at(I) == '>')
609               --Balance;
610             else if (Snippet->at(I) == '<')
611               ++Balance;
612             ++I;
613           } while (Balance > 0);
614           return Snippet->substr(0, I);
615         }
616         return "";
617       }
618     }
619     if (Full)
620       return *Snippet;
621 
622     // Replace argument snippets with a simplified pattern.
623     if (MayHaveArgList) {
624       // Functions snippets can be of 2 types:
625       // - containing only function arguments, e.g.
626       //   foo(${1:int p1}, ${2:int p2});
627       //   We transform this pattern to '($0)' or '()'.
628       // - template arguments and function arguments, e.g.
629       //   foo<${1:class}>(${2:int p1}).
630       //   We transform this pattern to '<$1>()$0' or '<$0>()'.
631 
632       bool EmptyArgs = llvm::StringRef(*Snippet).ends_with("()");
633       if (Snippet->front() == '<')
634         return None ? "" : (Open ? "<" : (EmptyArgs ? "<$1>()$0" : "<$1>($0)"));
635       if (Snippet->front() == '(')
636         return None ? "" : (Open ? "(" : (EmptyArgs ? "()" : "($0)"));
637       return *Snippet; // Not an arg snippet?
638     }
639     // 'CompletionItemKind::Interface' matches template type aliases.
640     if (Completion.Kind == CompletionItemKind::Interface ||
641         Completion.Kind == CompletionItemKind::Class ||
642         Completion.Kind == CompletionItemKind::Variable) {
643       if (Snippet->front() != '<')
644         return *Snippet; // Not an arg snippet?
645 
646       // Classes and template using aliases can only have template arguments,
647       // e.g. Foo<${1:class}>.
648       if (llvm::StringRef(*Snippet).ends_with("<>"))
649         return "<>"; // can happen with defaulted template arguments.
650       return None ? "" : (Open ? "<" : "<$0>");
651     }
652     return *Snippet;
653   }
654 
655   std::string summarizeSignature() const {
656     if (auto *Signature = onlyValue<&BundledEntry::Signature>())
657       return *Signature;
658     // All bundles are function calls.
659     return "(…)";
660   }
661 
662   // ASTCtx can be nullptr if not run with sema.
663   ASTContext *ASTCtx;
664   CodeCompletion Completion;
665   llvm::SmallVector<BundledEntry, 1> Bundled;
666   /// the way argument lists are handled.
667   Config::ArgumentListsPolicy ArgumentLists;
668   // No snippets will be generated for using declarations and when the function
669   // arguments are already present.
670   bool IsUsingDeclaration;
671   tok::TokenKind NextTokenKind;
672 };
673 
674 // Determine the symbol ID for a Sema code completion result, if possible.
675 SymbolID getSymbolID(const CodeCompletionResult &R, const SourceManager &SM) {
676   switch (R.Kind) {
677   case CodeCompletionResult::RK_Declaration:
678   case CodeCompletionResult::RK_Pattern: {
679     // Computing USR caches linkage, which may change after code completion.
680     if (hasUnstableLinkage(R.Declaration))
681       return {};
682     return clang::clangd::getSymbolID(R.Declaration);
683   }
684   case CodeCompletionResult::RK_Macro:
685     return clang::clangd::getSymbolID(R.Macro->getName(), R.MacroDefInfo, SM);
686   case CodeCompletionResult::RK_Keyword:
687     return {};
688   }
689   llvm_unreachable("unknown CodeCompletionResult kind");
690 }
691 
692 // Scopes of the partial identifier we're trying to complete.
693 // It is used when we query the index for more completion results.
694 struct SpecifiedScope {
695   // The scopes we should look in, determined by Sema.
696   //
697   // If the qualifier was fully resolved, we look for completions in these
698   // scopes; if there is an unresolved part of the qualifier, it should be
699   // resolved within these scopes.
700   //
701   // Examples of qualified completion:
702   //
703   //   "::vec"                                      => {""}
704   //   "using namespace std; ::vec^"                => {"", "std::"}
705   //   "namespace ns {using namespace std;} ns::^"  => {"ns::", "std::"}
706   //   "std::vec^"                                  => {""}  // "std" unresolved
707   //
708   // Examples of unqualified completion:
709   //
710   //   "vec^"                                        => {""}
711   //   "using namespace std; vec^"                   => {"", "std::"}
712   //   "namespace ns {inline namespace ni { struct Foo {}}}
713   //    using namespace ns::ni; Fo^ "                => {"", "ns::ni::"}
714   //   "using namespace std; namespace ns { vec^ }"  => {"ns::", "std::", ""}
715   //
716   // "" for global namespace, "ns::" for normal namespace.
717   std::vector<std::string> AccessibleScopes;
718   // This is an overestimate of AccessibleScopes, e.g. it ignores inline
719   // namespaces, to fetch more relevant symbols from index.
720   std::vector<std::string> QueryScopes;
721   // The full scope qualifier as typed by the user (without the leading "::").
722   // Set if the qualifier is not fully resolved by Sema.
723   std::optional<std::string> UnresolvedQualifier;
724 
725   std::optional<std::string> EnclosingNamespace;
726 
727   bool AllowAllScopes = false;
728 
729   // Scopes that are accessible from current context. Used for dropping
730   // unnecessary namespecifiers.
731   std::vector<std::string> scopesForQualification() {
732     std::set<std::string> Results;
733     for (llvm::StringRef AS : AccessibleScopes)
734       Results.insert(
735           (AS + (UnresolvedQualifier ? *UnresolvedQualifier : "")).str());
736     return {Results.begin(), Results.end()};
737   }
738 
739   // Construct scopes being queried in indexes. The results are deduplicated.
740   // This method formats the scopes to match the index request representation.
741   std::vector<std::string> scopesForIndexQuery() {
742     // The enclosing namespace must be first, it gets a quality boost.
743     std::vector<std::string> EnclosingAtFront;
744     if (EnclosingNamespace.has_value())
745       EnclosingAtFront.push_back(*EnclosingNamespace);
746     std::set<std::string> Deduplicated;
747     for (llvm::StringRef S : QueryScopes)
748       if (S != EnclosingNamespace)
749         Deduplicated.insert((S + UnresolvedQualifier.value_or("")).str());
750 
751     EnclosingAtFront.reserve(EnclosingAtFront.size() + Deduplicated.size());
752     llvm::copy(Deduplicated, std::back_inserter(EnclosingAtFront));
753 
754     return EnclosingAtFront;
755   }
756 };
757 
758 // Get all scopes that will be queried in indexes and whether symbols from
759 // any scope is allowed. The first scope in the list is the preferred scope
760 // (e.g. enclosing namespace).
761 SpecifiedScope getQueryScopes(CodeCompletionContext &CCContext,
762                               const Sema &CCSema,
763                               const CompletionPrefix &HeuristicPrefix,
764                               const CodeCompleteOptions &Opts) {
765   SpecifiedScope Scopes;
766   for (auto *Context : CCContext.getVisitedContexts()) {
767     if (isa<TranslationUnitDecl>(Context)) {
768       Scopes.QueryScopes.push_back("");
769       Scopes.AccessibleScopes.push_back("");
770     } else if (const auto *ND = dyn_cast<NamespaceDecl>(Context)) {
771       Scopes.QueryScopes.push_back(printNamespaceScope(*Context));
772       Scopes.AccessibleScopes.push_back(printQualifiedName(*ND) + "::");
773     }
774   }
775 
776   const CXXScopeSpec *SemaSpecifier =
777       CCContext.getCXXScopeSpecifier().value_or(nullptr);
778   // Case 1: unqualified completion.
779   if (!SemaSpecifier) {
780     // Case 2 (exception): sema saw no qualifier, but there appears to be one!
781     // This can happen e.g. in incomplete macro expansions. Use heuristics.
782     if (!HeuristicPrefix.Qualifier.empty()) {
783       vlog("Sema said no scope specifier, but we saw {0} in the source code",
784            HeuristicPrefix.Qualifier);
785       StringRef SpelledSpecifier = HeuristicPrefix.Qualifier;
786       if (SpelledSpecifier.consume_front("::")) {
787         Scopes.AccessibleScopes = {""};
788         Scopes.QueryScopes = {""};
789       }
790       Scopes.UnresolvedQualifier = std::string(SpelledSpecifier);
791       return Scopes;
792     }
793     /// FIXME: When the enclosing namespace contains an inline namespace,
794     /// it's dropped here. This leads to a behavior similar to
795     /// https://github.com/clangd/clangd/issues/1451
796     Scopes.EnclosingNamespace = printNamespaceScope(*CCSema.CurContext);
797     // Allow AllScopes completion as there is no explicit scope qualifier.
798     Scopes.AllowAllScopes = Opts.AllScopes;
799     return Scopes;
800   }
801   // Case 3: sema saw and resolved a scope qualifier.
802   if (SemaSpecifier && SemaSpecifier->isValid())
803     return Scopes;
804 
805   // Case 4: There was a qualifier, and Sema didn't resolve it.
806   Scopes.QueryScopes.push_back(""); // Make sure global scope is included.
807   llvm::StringRef SpelledSpecifier = Lexer::getSourceText(
808       CharSourceRange::getCharRange(SemaSpecifier->getRange()),
809       CCSema.SourceMgr, clang::LangOptions());
810   if (SpelledSpecifier.consume_front("::"))
811     Scopes.QueryScopes = {""};
812   Scopes.UnresolvedQualifier = std::string(SpelledSpecifier);
813   // Sema excludes the trailing "::".
814   if (!Scopes.UnresolvedQualifier->empty())
815     *Scopes.UnresolvedQualifier += "::";
816 
817   Scopes.AccessibleScopes = Scopes.QueryScopes;
818 
819   return Scopes;
820 }
821 
822 // Should we perform index-based completion in a context of the specified kind?
823 // FIXME: consider allowing completion, but restricting the result types.
824 bool contextAllowsIndex(enum CodeCompletionContext::Kind K) {
825   switch (K) {
826   case CodeCompletionContext::CCC_TopLevel:
827   case CodeCompletionContext::CCC_ObjCInterface:
828   case CodeCompletionContext::CCC_ObjCImplementation:
829   case CodeCompletionContext::CCC_ObjCIvarList:
830   case CodeCompletionContext::CCC_ClassStructUnion:
831   case CodeCompletionContext::CCC_Statement:
832   case CodeCompletionContext::CCC_Expression:
833   case CodeCompletionContext::CCC_ObjCMessageReceiver:
834   case CodeCompletionContext::CCC_EnumTag:
835   case CodeCompletionContext::CCC_UnionTag:
836   case CodeCompletionContext::CCC_ClassOrStructTag:
837   case CodeCompletionContext::CCC_ObjCProtocolName:
838   case CodeCompletionContext::CCC_Namespace:
839   case CodeCompletionContext::CCC_Type:
840   case CodeCompletionContext::CCC_ParenthesizedExpression:
841   case CodeCompletionContext::CCC_ObjCInterfaceName:
842   case CodeCompletionContext::CCC_Symbol:
843   case CodeCompletionContext::CCC_SymbolOrNewName:
844   case CodeCompletionContext::CCC_ObjCClassForwardDecl:
845   case CodeCompletionContext::CCC_TopLevelOrExpression:
846     return true;
847   case CodeCompletionContext::CCC_OtherWithMacros:
848   case CodeCompletionContext::CCC_DotMemberAccess:
849   case CodeCompletionContext::CCC_ArrowMemberAccess:
850   case CodeCompletionContext::CCC_ObjCCategoryName:
851   case CodeCompletionContext::CCC_ObjCPropertyAccess:
852   case CodeCompletionContext::CCC_MacroName:
853   case CodeCompletionContext::CCC_MacroNameUse:
854   case CodeCompletionContext::CCC_PreprocessorExpression:
855   case CodeCompletionContext::CCC_PreprocessorDirective:
856   case CodeCompletionContext::CCC_SelectorName:
857   case CodeCompletionContext::CCC_TypeQualifiers:
858   case CodeCompletionContext::CCC_ObjCInstanceMessage:
859   case CodeCompletionContext::CCC_ObjCClassMessage:
860   case CodeCompletionContext::CCC_IncludedFile:
861   case CodeCompletionContext::CCC_Attribute:
862   // FIXME: Provide identifier based completions for the following contexts:
863   case CodeCompletionContext::CCC_Other: // Be conservative.
864   case CodeCompletionContext::CCC_NaturalLanguage:
865   case CodeCompletionContext::CCC_Recovery:
866   case CodeCompletionContext::CCC_NewName:
867     return false;
868   }
869   llvm_unreachable("unknown code completion context");
870 }
871 
872 static bool isInjectedClass(const NamedDecl &D) {
873   if (auto *R = dyn_cast_or_null<RecordDecl>(&D))
874     if (R->isInjectedClassName())
875       return true;
876   return false;
877 }
878 
879 // Some member calls are excluded because they're so rarely useful.
880 static bool isExcludedMember(const NamedDecl &D) {
881   // Destructor completion is rarely useful, and works inconsistently.
882   // (s.^ completes ~string, but s.~st^ is an error).
883   if (D.getKind() == Decl::CXXDestructor)
884     return true;
885   // Injected name may be useful for A::foo(), but who writes A::A::foo()?
886   if (isInjectedClass(D))
887     return true;
888   // Explicit calls to operators are also rare.
889   auto NameKind = D.getDeclName().getNameKind();
890   if (NameKind == DeclarationName::CXXOperatorName ||
891       NameKind == DeclarationName::CXXLiteralOperatorName ||
892       NameKind == DeclarationName::CXXConversionFunctionName)
893     return true;
894   return false;
895 }
896 
897 // The CompletionRecorder captures Sema code-complete output, including context.
898 // It filters out ignored results (but doesn't apply fuzzy-filtering yet).
899 // It doesn't do scoring or conversion to CompletionItem yet, as we want to
900 // merge with index results first.
901 // Generally the fields and methods of this object should only be used from
902 // within the callback.
903 struct CompletionRecorder : public CodeCompleteConsumer {
904   CompletionRecorder(const CodeCompleteOptions &Opts,
905                      llvm::unique_function<void()> ResultsCallback)
906       : CodeCompleteConsumer(Opts.getClangCompleteOpts()),
907         CCContext(CodeCompletionContext::CCC_Other), Opts(Opts),
908         CCAllocator(std::make_shared<GlobalCodeCompletionAllocator>()),
909         CCTUInfo(CCAllocator), ResultsCallback(std::move(ResultsCallback)) {
910     assert(this->ResultsCallback);
911   }
912 
913   std::vector<CodeCompletionResult> Results;
914   CodeCompletionContext CCContext;
915   Sema *CCSema = nullptr; // Sema that created the results.
916   // FIXME: Sema is scary. Can we store ASTContext and Preprocessor, instead?
917 
918   void ProcessCodeCompleteResults(class Sema &S, CodeCompletionContext Context,
919                                   CodeCompletionResult *InResults,
920                                   unsigned NumResults) final {
921     // Results from recovery mode are generally useless, and the callback after
922     // recovery (if any) is usually more interesting. To make sure we handle the
923     // future callback from sema, we just ignore all callbacks in recovery mode,
924     // as taking only results from recovery mode results in poor completion
925     // results.
926     // FIXME: in case there is no future sema completion callback after the
927     // recovery mode, we might still want to provide some results (e.g. trivial
928     // identifier-based completion).
929     if (Context.getKind() == CodeCompletionContext::CCC_Recovery) {
930       log("Code complete: Ignoring sema code complete callback with Recovery "
931           "context.");
932       return;
933     }
934     // If a callback is called without any sema result and the context does not
935     // support index-based completion, we simply skip it to give way to
936     // potential future callbacks with results.
937     if (NumResults == 0 && !contextAllowsIndex(Context.getKind()))
938       return;
939     if (CCSema) {
940       log("Multiple code complete callbacks (parser backtracked?). "
941           "Dropping results from context {0}, keeping results from {1}.",
942           getCompletionKindString(Context.getKind()),
943           getCompletionKindString(this->CCContext.getKind()));
944       return;
945     }
946     // Record the completion context.
947     CCSema = &S;
948     CCContext = Context;
949 
950     // Retain the results we might want.
951     for (unsigned I = 0; I < NumResults; ++I) {
952       auto &Result = InResults[I];
953       // Class members that are shadowed by subclasses are usually noise.
954       if (Result.Hidden && Result.Declaration &&
955           Result.Declaration->isCXXClassMember())
956         continue;
957       if (!Opts.IncludeIneligibleResults &&
958           (Result.Availability == CXAvailability_NotAvailable ||
959            Result.Availability == CXAvailability_NotAccessible))
960         continue;
961       if (Result.Declaration &&
962           !Context.getBaseType().isNull() // is this a member-access context?
963           && isExcludedMember(*Result.Declaration))
964         continue;
965       // Skip injected class name when no class scope is not explicitly set.
966       // E.g. show injected A::A in `using A::A^` but not in "A^".
967       if (Result.Declaration && !Context.getCXXScopeSpecifier() &&
968           isInjectedClass(*Result.Declaration))
969         continue;
970       // We choose to never append '::' to completion results in clangd.
971       Result.StartsNestedNameSpecifier = false;
972       Results.push_back(Result);
973     }
974     ResultsCallback();
975   }
976 
977   CodeCompletionAllocator &getAllocator() override { return *CCAllocator; }
978   CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; }
979 
980   // Returns the filtering/sorting name for Result, which must be from Results.
981   // Returned string is owned by this recorder (or the AST).
982   llvm::StringRef getName(const CodeCompletionResult &Result) {
983     switch (Result.Kind) {
984     case CodeCompletionResult::RK_Declaration:
985       if (auto *ID = Result.Declaration->getIdentifier())
986         return ID->getName();
987       break;
988     case CodeCompletionResult::RK_Keyword:
989       return Result.Keyword;
990     case CodeCompletionResult::RK_Macro:
991       return Result.Macro->getName();
992     case CodeCompletionResult::RK_Pattern:
993       break;
994     }
995     auto *CCS = codeCompletionString(Result);
996     const CodeCompletionString::Chunk *OnlyText = nullptr;
997     for (auto &C : *CCS) {
998       if (C.Kind != CodeCompletionString::CK_TypedText)
999         continue;
1000       if (OnlyText)
1001         return CCAllocator->CopyString(CCS->getAllTypedText());
1002       OnlyText = &C;
1003     }
1004     return OnlyText ? OnlyText->Text : llvm::StringRef();
1005   }
1006 
1007   // Build a CodeCompletion string for R, which must be from Results.
1008   // The CCS will be owned by this recorder.
1009   CodeCompletionString *codeCompletionString(const CodeCompletionResult &R) {
1010     // CodeCompletionResult doesn't seem to be const-correct. We own it, anyway.
1011     return const_cast<CodeCompletionResult &>(R).CreateCodeCompletionString(
1012         *CCSema, CCContext, *CCAllocator, CCTUInfo,
1013         /*IncludeBriefComments=*/false);
1014   }
1015 
1016 private:
1017   CodeCompleteOptions Opts;
1018   std::shared_ptr<GlobalCodeCompletionAllocator> CCAllocator;
1019   CodeCompletionTUInfo CCTUInfo;
1020   llvm::unique_function<void()> ResultsCallback;
1021 };
1022 
1023 struct ScoredSignature {
1024   // When not null, requires documentation to be requested from the index with
1025   // this ID.
1026   SymbolID IDForDoc;
1027   SignatureInformation Signature;
1028   SignatureQualitySignals Quality;
1029 };
1030 
1031 // Returns the index of the parameter matching argument number "Arg.
1032 // This is usually just "Arg", except for variadic functions/templates, where
1033 // "Arg" might be higher than the number of parameters. When that happens, we
1034 // assume the last parameter is variadic and assume all further args are
1035 // part of it.
1036 int paramIndexForArg(const CodeCompleteConsumer::OverloadCandidate &Candidate,
1037                      int Arg) {
1038   int NumParams = Candidate.getNumParams();
1039   if (auto *T = Candidate.getFunctionType()) {
1040     if (auto *Proto = T->getAs<FunctionProtoType>()) {
1041       if (Proto->isVariadic())
1042         ++NumParams;
1043     }
1044   }
1045   return std::min(Arg, std::max(NumParams - 1, 0));
1046 }
1047 
1048 class SignatureHelpCollector final : public CodeCompleteConsumer {
1049 public:
1050   SignatureHelpCollector(const clang::CodeCompleteOptions &CodeCompleteOpts,
1051                          MarkupKind DocumentationFormat,
1052                          const SymbolIndex *Index, SignatureHelp &SigHelp)
1053       : CodeCompleteConsumer(CodeCompleteOpts), SigHelp(SigHelp),
1054         Allocator(std::make_shared<clang::GlobalCodeCompletionAllocator>()),
1055         CCTUInfo(Allocator), Index(Index),
1056         DocumentationFormat(DocumentationFormat) {}
1057 
1058   void ProcessOverloadCandidates(Sema &S, unsigned CurrentArg,
1059                                  OverloadCandidate *Candidates,
1060                                  unsigned NumCandidates,
1061                                  SourceLocation OpenParLoc,
1062                                  bool Braced) override {
1063     assert(!OpenParLoc.isInvalid());
1064     SourceManager &SrcMgr = S.getSourceManager();
1065     OpenParLoc = SrcMgr.getFileLoc(OpenParLoc);
1066     if (SrcMgr.isInMainFile(OpenParLoc))
1067       SigHelp.argListStart = sourceLocToPosition(SrcMgr, OpenParLoc);
1068     else
1069       elog("Location oustide main file in signature help: {0}",
1070            OpenParLoc.printToString(SrcMgr));
1071 
1072     std::vector<ScoredSignature> ScoredSignatures;
1073     SigHelp.signatures.reserve(NumCandidates);
1074     ScoredSignatures.reserve(NumCandidates);
1075     // FIXME(rwols): How can we determine the "active overload candidate"?
1076     // Right now the overloaded candidates seem to be provided in a "best fit"
1077     // order, so I'm not too worried about this.
1078     SigHelp.activeSignature = 0;
1079     assert(CurrentArg <= (unsigned)std::numeric_limits<int>::max() &&
1080            "too many arguments");
1081 
1082     SigHelp.activeParameter = static_cast<int>(CurrentArg);
1083 
1084     for (unsigned I = 0; I < NumCandidates; ++I) {
1085       OverloadCandidate Candidate = Candidates[I];
1086       // We want to avoid showing instantiated signatures, because they may be
1087       // long in some cases (e.g. when 'T' is substituted with 'std::string', we
1088       // would get 'std::basic_string<char>').
1089       if (auto *Func = Candidate.getFunction()) {
1090         if (auto *Pattern = Func->getTemplateInstantiationPattern())
1091           Candidate = OverloadCandidate(Pattern);
1092       }
1093       if (static_cast<int>(I) == SigHelp.activeSignature) {
1094         // The activeParameter in LSP relates to the activeSignature. There is
1095         // another, per-signature field, but we currently do not use it and not
1096         // all clients might support it.
1097         // FIXME: Add support for per-signature activeParameter field.
1098         SigHelp.activeParameter =
1099             paramIndexForArg(Candidate, SigHelp.activeParameter);
1100       }
1101 
1102       const auto *CCS = Candidate.CreateSignatureString(
1103           CurrentArg, S, *Allocator, CCTUInfo,
1104           /*IncludeBriefComments=*/true, Braced);
1105       assert(CCS && "Expected the CodeCompletionString to be non-null");
1106       ScoredSignatures.push_back(processOverloadCandidate(
1107           Candidate, *CCS,
1108           Candidate.getFunction()
1109               ? getDeclComment(S.getASTContext(), *Candidate.getFunction())
1110               : ""));
1111     }
1112 
1113     // Sema does not load the docs from the preamble, so we need to fetch extra
1114     // docs from the index instead.
1115     llvm::DenseMap<SymbolID, std::string> FetchedDocs;
1116     if (Index) {
1117       LookupRequest IndexRequest;
1118       for (const auto &S : ScoredSignatures) {
1119         if (!S.IDForDoc)
1120           continue;
1121         IndexRequest.IDs.insert(S.IDForDoc);
1122       }
1123       Index->lookup(IndexRequest, [&](const Symbol &S) {
1124         if (!S.Documentation.empty())
1125           FetchedDocs[S.ID] = std::string(S.Documentation);
1126       });
1127       vlog("SigHelp: requested docs for {0} symbols from the index, got {1} "
1128            "symbols with non-empty docs in the response",
1129            IndexRequest.IDs.size(), FetchedDocs.size());
1130     }
1131 
1132     llvm::sort(ScoredSignatures, [](const ScoredSignature &L,
1133                                     const ScoredSignature &R) {
1134       // Ordering follows:
1135       // - Less number of parameters is better.
1136       // - Aggregate > Function > FunctionType > FunctionTemplate
1137       // - High score is better.
1138       // - Shorter signature is better.
1139       // - Alphabetically smaller is better.
1140       if (L.Quality.NumberOfParameters != R.Quality.NumberOfParameters)
1141         return L.Quality.NumberOfParameters < R.Quality.NumberOfParameters;
1142       if (L.Quality.NumberOfOptionalParameters !=
1143           R.Quality.NumberOfOptionalParameters)
1144         return L.Quality.NumberOfOptionalParameters <
1145                R.Quality.NumberOfOptionalParameters;
1146       if (L.Quality.Kind != R.Quality.Kind) {
1147         using OC = CodeCompleteConsumer::OverloadCandidate;
1148         auto KindPriority = [&](OC::CandidateKind K) {
1149           switch (K) {
1150           case OC::CK_Aggregate:
1151             return 0;
1152           case OC::CK_Function:
1153             return 1;
1154           case OC::CK_FunctionType:
1155             return 2;
1156           case OC::CK_FunctionProtoTypeLoc:
1157             return 3;
1158           case OC::CK_FunctionTemplate:
1159             return 4;
1160           case OC::CK_Template:
1161             return 5;
1162           }
1163           llvm_unreachable("Unknown overload candidate type.");
1164         };
1165         return KindPriority(L.Quality.Kind) < KindPriority(R.Quality.Kind);
1166       }
1167       if (L.Signature.label.size() != R.Signature.label.size())
1168         return L.Signature.label.size() < R.Signature.label.size();
1169       return L.Signature.label < R.Signature.label;
1170     });
1171 
1172     for (auto &SS : ScoredSignatures) {
1173       auto IndexDocIt =
1174           SS.IDForDoc ? FetchedDocs.find(SS.IDForDoc) : FetchedDocs.end();
1175       if (IndexDocIt != FetchedDocs.end()) {
1176         markup::Document SignatureComment;
1177         parseDocumentation(IndexDocIt->second, SignatureComment);
1178         SS.Signature.documentation =
1179             renderDoc(SignatureComment, DocumentationFormat);
1180       }
1181 
1182       SigHelp.signatures.push_back(std::move(SS.Signature));
1183     }
1184   }
1185 
1186   GlobalCodeCompletionAllocator &getAllocator() override { return *Allocator; }
1187 
1188   CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; }
1189 
1190 private:
1191   void processParameterChunk(llvm::StringRef ChunkText,
1192                              SignatureInformation &Signature) const {
1193     // (!) this is O(n), should still be fast compared to building ASTs.
1194     unsigned ParamStartOffset = lspLength(Signature.label);
1195     unsigned ParamEndOffset = ParamStartOffset + lspLength(ChunkText);
1196     // A piece of text that describes the parameter that corresponds to
1197     // the code-completion location within a function call, message send,
1198     // macro invocation, etc.
1199     Signature.label += ChunkText;
1200     ParameterInformation Info;
1201     Info.labelOffsets.emplace(ParamStartOffset, ParamEndOffset);
1202     // FIXME: only set 'labelOffsets' when all clients migrate out of it.
1203     Info.labelString = std::string(ChunkText);
1204 
1205     Signature.parameters.push_back(std::move(Info));
1206   }
1207 
1208   void processOptionalChunk(const CodeCompletionString &CCS,
1209                             SignatureInformation &Signature,
1210                             SignatureQualitySignals &Signal) const {
1211     for (const auto &Chunk : CCS) {
1212       switch (Chunk.Kind) {
1213       case CodeCompletionString::CK_Optional:
1214         assert(Chunk.Optional &&
1215                "Expected the optional code completion string to be non-null.");
1216         processOptionalChunk(*Chunk.Optional, Signature, Signal);
1217         break;
1218       case CodeCompletionString::CK_VerticalSpace:
1219         break;
1220       case CodeCompletionString::CK_CurrentParameter:
1221       case CodeCompletionString::CK_Placeholder:
1222         processParameterChunk(Chunk.Text, Signature);
1223         Signal.NumberOfOptionalParameters++;
1224         break;
1225       default:
1226         Signature.label += Chunk.Text;
1227         break;
1228       }
1229     }
1230   }
1231 
1232   // FIXME(ioeric): consider moving CodeCompletionString logic here to
1233   // CompletionString.h.
1234   ScoredSignature processOverloadCandidate(const OverloadCandidate &Candidate,
1235                                            const CodeCompletionString &CCS,
1236                                            llvm::StringRef DocComment) const {
1237     SignatureInformation Signature;
1238     SignatureQualitySignals Signal;
1239     const char *ReturnType = nullptr;
1240 
1241     markup::Document OverloadComment;
1242     parseDocumentation(formatDocumentation(CCS, DocComment), OverloadComment);
1243     Signature.documentation = renderDoc(OverloadComment, DocumentationFormat);
1244     Signal.Kind = Candidate.getKind();
1245 
1246     for (const auto &Chunk : CCS) {
1247       switch (Chunk.Kind) {
1248       case CodeCompletionString::CK_ResultType:
1249         // A piece of text that describes the type of an entity or,
1250         // for functions and methods, the return type.
1251         assert(!ReturnType && "Unexpected CK_ResultType");
1252         ReturnType = Chunk.Text;
1253         break;
1254       case CodeCompletionString::CK_CurrentParameter:
1255       case CodeCompletionString::CK_Placeholder:
1256         processParameterChunk(Chunk.Text, Signature);
1257         Signal.NumberOfParameters++;
1258         break;
1259       case CodeCompletionString::CK_Optional: {
1260         // The rest of the parameters are defaulted/optional.
1261         assert(Chunk.Optional &&
1262                "Expected the optional code completion string to be non-null.");
1263         processOptionalChunk(*Chunk.Optional, Signature, Signal);
1264         break;
1265       }
1266       case CodeCompletionString::CK_VerticalSpace:
1267         break;
1268       default:
1269         Signature.label += Chunk.Text;
1270         break;
1271       }
1272     }
1273     if (ReturnType) {
1274       Signature.label += " -> ";
1275       Signature.label += ReturnType;
1276     }
1277     dlog("Signal for {0}: {1}", Signature, Signal);
1278     ScoredSignature Result;
1279     Result.Signature = std::move(Signature);
1280     Result.Quality = Signal;
1281     const FunctionDecl *Func = Candidate.getFunction();
1282     if (Func && Result.Signature.documentation.value.empty()) {
1283       // Computing USR caches linkage, which may change after code completion.
1284       if (!hasUnstableLinkage(Func))
1285         Result.IDForDoc = clangd::getSymbolID(Func);
1286     }
1287     return Result;
1288   }
1289 
1290   SignatureHelp &SigHelp;
1291   std::shared_ptr<clang::GlobalCodeCompletionAllocator> Allocator;
1292   CodeCompletionTUInfo CCTUInfo;
1293   const SymbolIndex *Index;
1294   MarkupKind DocumentationFormat;
1295 }; // SignatureHelpCollector
1296 
1297 // Used only for completion of C-style comments in function call (i.e.
1298 // /*foo=*/7). Similar to SignatureHelpCollector, but needs to do less work.
1299 class ParamNameCollector final : public CodeCompleteConsumer {
1300 public:
1301   ParamNameCollector(const clang::CodeCompleteOptions &CodeCompleteOpts,
1302                      std::set<std::string> &ParamNames)
1303       : CodeCompleteConsumer(CodeCompleteOpts),
1304         Allocator(std::make_shared<clang::GlobalCodeCompletionAllocator>()),
1305         CCTUInfo(Allocator), ParamNames(ParamNames) {}
1306 
1307   void ProcessOverloadCandidates(Sema &S, unsigned CurrentArg,
1308                                  OverloadCandidate *Candidates,
1309                                  unsigned NumCandidates,
1310                                  SourceLocation OpenParLoc,
1311                                  bool Braced) override {
1312     assert(CurrentArg <= (unsigned)std::numeric_limits<int>::max() &&
1313            "too many arguments");
1314 
1315     for (unsigned I = 0; I < NumCandidates; ++I) {
1316       if (const NamedDecl *ND = Candidates[I].getParamDecl(CurrentArg))
1317         if (const auto *II = ND->getIdentifier())
1318           ParamNames.emplace(II->getName());
1319     }
1320   }
1321 
1322 private:
1323   GlobalCodeCompletionAllocator &getAllocator() override { return *Allocator; }
1324 
1325   CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; }
1326 
1327   std::shared_ptr<clang::GlobalCodeCompletionAllocator> Allocator;
1328   CodeCompletionTUInfo CCTUInfo;
1329   std::set<std::string> &ParamNames;
1330 };
1331 
1332 struct SemaCompleteInput {
1333   PathRef FileName;
1334   size_t Offset;
1335   const PreambleData &Preamble;
1336   const std::optional<PreamblePatch> Patch;
1337   const ParseInputs &ParseInput;
1338 };
1339 
1340 void loadMainFilePreambleMacros(const Preprocessor &PP,
1341                                 const PreambleData &Preamble) {
1342   // The ExternalPreprocessorSource has our macros, if we know where to look.
1343   // We can read all the macros using PreambleMacros->ReadDefinedMacros(),
1344   // but this includes transitively included files, so may deserialize a lot.
1345   ExternalPreprocessorSource *PreambleMacros = PP.getExternalSource();
1346   // As we have the names of the macros, we can look up their IdentifierInfo
1347   // and then use this to load just the macros we want.
1348   const auto &ITable = PP.getIdentifierTable();
1349   IdentifierInfoLookup *PreambleIdentifiers =
1350       ITable.getExternalIdentifierLookup();
1351 
1352   if (!PreambleIdentifiers || !PreambleMacros)
1353     return;
1354   for (const auto &MacroName : Preamble.Macros.Names) {
1355     if (ITable.find(MacroName.getKey()) != ITable.end())
1356       continue;
1357     if (auto *II = PreambleIdentifiers->get(MacroName.getKey()))
1358       if (II->isOutOfDate())
1359         PreambleMacros->updateOutOfDateIdentifier(*II);
1360   }
1361 }
1362 
1363 // Invokes Sema code completion on a file.
1364 // If \p Includes is set, it will be updated based on the compiler invocation.
1365 bool semaCodeComplete(std::unique_ptr<CodeCompleteConsumer> Consumer,
1366                       const clang::CodeCompleteOptions &Options,
1367                       const SemaCompleteInput &Input,
1368                       IncludeStructure *Includes = nullptr) {
1369   trace::Span Tracer("Sema completion");
1370 
1371   IgnoreDiagnostics IgnoreDiags;
1372   auto CI = buildCompilerInvocation(Input.ParseInput, IgnoreDiags);
1373   if (!CI) {
1374     elog("Couldn't create CompilerInvocation");
1375     return false;
1376   }
1377   auto &FrontendOpts = CI->getFrontendOpts();
1378   FrontendOpts.SkipFunctionBodies = true;
1379   // Disable typo correction in Sema.
1380   CI->getLangOpts().SpellChecking = false;
1381   // Code completion won't trigger in delayed template bodies.
1382   // This is on-by-default in windows to allow parsing SDK headers; we're only
1383   // disabling it for the main-file (not preamble).
1384   CI->getLangOpts().DelayedTemplateParsing = false;
1385   // Setup code completion.
1386   FrontendOpts.CodeCompleteOpts = Options;
1387   FrontendOpts.CodeCompletionAt.FileName = std::string(Input.FileName);
1388   std::tie(FrontendOpts.CodeCompletionAt.Line,
1389            FrontendOpts.CodeCompletionAt.Column) =
1390       offsetToClangLineColumn(Input.ParseInput.Contents, Input.Offset);
1391 
1392   std::unique_ptr<llvm::MemoryBuffer> ContentsBuffer =
1393       llvm::MemoryBuffer::getMemBuffer(Input.ParseInput.Contents,
1394                                        Input.FileName);
1395   // The diagnostic options must be set before creating a CompilerInstance.
1396   CI->getDiagnosticOpts().IgnoreWarnings = true;
1397   // We reuse the preamble whether it's valid or not. This is a
1398   // correctness/performance tradeoff: building without a preamble is slow, and
1399   // completion is latency-sensitive.
1400   // However, if we're completing *inside* the preamble section of the draft,
1401   // overriding the preamble will break sema completion. Fortunately we can just
1402   // skip all includes in this case; these completions are really simple.
1403   PreambleBounds PreambleRegion =
1404       ComputePreambleBounds(CI->getLangOpts(), *ContentsBuffer, 0);
1405   bool CompletingInPreamble = Input.Offset < PreambleRegion.Size ||
1406                               (!PreambleRegion.PreambleEndsAtStartOfLine &&
1407                                Input.Offset == PreambleRegion.Size);
1408   if (Input.Patch)
1409     Input.Patch->apply(*CI);
1410   // NOTE: we must call BeginSourceFile after prepareCompilerInstance. Otherwise
1411   // the remapped buffers do not get freed.
1412   llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS =
1413       Input.ParseInput.TFS->view(Input.ParseInput.CompileCommand.Directory);
1414   if (Input.Preamble.StatCache)
1415     VFS = Input.Preamble.StatCache->getConsumingFS(std::move(VFS));
1416   auto Clang = prepareCompilerInstance(
1417       std::move(CI), !CompletingInPreamble ? &Input.Preamble.Preamble : nullptr,
1418       std::move(ContentsBuffer), std::move(VFS), IgnoreDiags);
1419   Clang->getPreprocessorOpts().SingleFileParseMode = CompletingInPreamble;
1420   Clang->setCodeCompletionConsumer(Consumer.release());
1421 
1422   if (Input.Preamble.RequiredModules)
1423     Input.Preamble.RequiredModules->adjustHeaderSearchOptions(Clang->getHeaderSearchOpts());
1424 
1425   SyntaxOnlyAction Action;
1426   if (!Action.BeginSourceFile(*Clang, Clang->getFrontendOpts().Inputs[0])) {
1427     log("BeginSourceFile() failed when running codeComplete for {0}",
1428         Input.FileName);
1429     return false;
1430   }
1431   // Macros can be defined within the preamble region of the main file.
1432   // They don't fall nicely into our index/Sema dichotomy:
1433   //  - they're not indexed for completion (they're not available across files)
1434   //  - but Sema code complete won't see them: as part of the preamble, they're
1435   //    deserialized only when mentioned.
1436   // Force them to be deserialized so SemaCodeComplete sees them.
1437   loadMainFilePreambleMacros(Clang->getPreprocessor(), Input.Preamble);
1438   if (Includes)
1439     Includes->collect(*Clang);
1440   if (llvm::Error Err = Action.Execute()) {
1441     log("Execute() failed when running codeComplete for {0}: {1}",
1442         Input.FileName, toString(std::move(Err)));
1443     return false;
1444   }
1445   Action.EndSourceFile();
1446 
1447   return true;
1448 }
1449 
1450 // Should we allow index completions in the specified context?
1451 bool allowIndex(CodeCompletionContext &CC) {
1452   if (!contextAllowsIndex(CC.getKind()))
1453     return false;
1454   // We also avoid ClassName::bar (but allow namespace::bar).
1455   auto Scope = CC.getCXXScopeSpecifier();
1456   if (!Scope)
1457     return true;
1458   NestedNameSpecifier *NameSpec = (*Scope)->getScopeRep();
1459   if (!NameSpec)
1460     return true;
1461   // We only query the index when qualifier is a namespace.
1462   // If it's a class, we rely solely on sema completions.
1463   switch (NameSpec->getKind()) {
1464   case NestedNameSpecifier::Global:
1465   case NestedNameSpecifier::Namespace:
1466   case NestedNameSpecifier::NamespaceAlias:
1467     return true;
1468   case NestedNameSpecifier::Super:
1469   case NestedNameSpecifier::TypeSpec:
1470   case NestedNameSpecifier::TypeSpecWithTemplate:
1471   // Unresolved inside a template.
1472   case NestedNameSpecifier::Identifier:
1473     return false;
1474   }
1475   llvm_unreachable("invalid NestedNameSpecifier kind");
1476 }
1477 
1478 // Should we include a symbol from the index given the completion kind?
1479 // FIXME: Ideally we can filter in the fuzzy find request itself.
1480 bool includeSymbolFromIndex(CodeCompletionContext::Kind Kind,
1481                             const Symbol &Sym) {
1482   // Objective-C protocols are only useful in ObjC protocol completions,
1483   // in other places they're confusing, especially when they share the same
1484   // identifier with a class.
1485   if (Sym.SymInfo.Kind == index::SymbolKind::Protocol &&
1486       Sym.SymInfo.Lang == index::SymbolLanguage::ObjC)
1487     return Kind == CodeCompletionContext::CCC_ObjCProtocolName;
1488   else if (Kind == CodeCompletionContext::CCC_ObjCProtocolName)
1489     // Don't show anything else in ObjC protocol completions.
1490     return false;
1491 
1492   if (Kind == CodeCompletionContext::CCC_ObjCClassForwardDecl)
1493     return Sym.SymInfo.Kind == index::SymbolKind::Class &&
1494            Sym.SymInfo.Lang == index::SymbolLanguage::ObjC;
1495   return true;
1496 }
1497 
1498 std::future<std::pair<bool, SymbolSlab>>
1499 startAsyncFuzzyFind(const SymbolIndex &Index, const FuzzyFindRequest &Req) {
1500   return runAsync<std::pair<bool, SymbolSlab>>([&Index, Req]() {
1501     trace::Span Tracer("Async fuzzyFind");
1502     SymbolSlab::Builder Syms;
1503     bool Incomplete =
1504         Index.fuzzyFind(Req, [&Syms](const Symbol &Sym) { Syms.insert(Sym); });
1505     return std::make_pair(Incomplete, std::move(Syms).build());
1506   });
1507 }
1508 
1509 // Creates a `FuzzyFindRequest` based on the cached index request from the
1510 // last completion, if any, and the speculated completion filter text in the
1511 // source code.
1512 FuzzyFindRequest speculativeFuzzyFindRequestForCompletion(
1513     FuzzyFindRequest CachedReq, const CompletionPrefix &HeuristicPrefix) {
1514   CachedReq.Query = std::string(HeuristicPrefix.Name);
1515   return CachedReq;
1516 }
1517 
1518 // This function is similar to Lexer::findNextToken(), but assumes
1519 // that the input SourceLocation is the completion point (which is
1520 // a case findNextToken() does not handle).
1521 std::optional<Token>
1522 findTokenAfterCompletionPoint(SourceLocation CompletionPoint,
1523                               const SourceManager &SM,
1524                               const LangOptions &LangOpts) {
1525   SourceLocation Loc = CompletionPoint;
1526   if (Loc.isMacroID()) {
1527     if (!Lexer::isAtEndOfMacroExpansion(Loc, SM, LangOpts, &Loc))
1528       return std::nullopt;
1529   }
1530 
1531   // Advance to the next SourceLocation after the completion point.
1532   // Lexer::findNextToken() would call MeasureTokenLength() here,
1533   // which does not handle the completion point (and can't, because
1534   // the Lexer instance it constructs internally doesn't have a
1535   // Preprocessor and so doesn't know about the completion point).
1536   Loc = Loc.getLocWithOffset(1);
1537 
1538   // Break down the source location.
1539   std::pair<FileID, unsigned> LocInfo = SM.getDecomposedLoc(Loc);
1540 
1541   // Try to load the file buffer.
1542   bool InvalidTemp = false;
1543   StringRef File = SM.getBufferData(LocInfo.first, &InvalidTemp);
1544   if (InvalidTemp)
1545     return std::nullopt;
1546 
1547   const char *TokenBegin = File.data() + LocInfo.second;
1548 
1549   // Lex from the start of the given location.
1550   Lexer TheLexer(SM.getLocForStartOfFile(LocInfo.first), LangOpts, File.begin(),
1551                  TokenBegin, File.end());
1552   // Find the token.
1553   Token Tok;
1554   TheLexer.LexFromRawLexer(Tok);
1555   return Tok;
1556 }
1557 
1558 // Runs Sema-based (AST) and Index-based completion, returns merged results.
1559 //
1560 // There are a few tricky considerations:
1561 //   - the AST provides information needed for the index query (e.g. which
1562 //     namespaces to search in). So Sema must start first.
1563 //   - we only want to return the top results (Opts.Limit).
1564 //     Building CompletionItems for everything else is wasteful, so we want to
1565 //     preserve the "native" format until we're done with scoring.
1566 //   - the data underlying Sema completion items is owned by the AST and various
1567 //     other arenas, which must stay alive for us to build CompletionItems.
1568 //   - we may get duplicate results from Sema and the Index, we need to merge.
1569 //
1570 // So we start Sema completion first, and do all our work in its callback.
1571 // We use the Sema context information to query the index.
1572 // Then we merge the two result sets, producing items that are Sema/Index/Both.
1573 // These items are scored, and the top N are synthesized into the LSP response.
1574 // Finally, we can clean up the data structures created by Sema completion.
1575 //
1576 // Main collaborators are:
1577 //   - semaCodeComplete sets up the compiler machinery to run code completion.
1578 //   - CompletionRecorder captures Sema completion results, including context.
1579 //   - SymbolIndex (Opts.Index) provides index completion results as Symbols
1580 //   - CompletionCandidates are the result of merging Sema and Index results.
1581 //     Each candidate points to an underlying CodeCompletionResult (Sema), a
1582 //     Symbol (Index), or both. It computes the result quality score.
1583 //     CompletionCandidate also does conversion to CompletionItem (at the end).
1584 //   - FuzzyMatcher scores how the candidate matches the partial identifier.
1585 //     This score is combined with the result quality score for the final score.
1586 //   - TopN determines the results with the best score.
1587 class CodeCompleteFlow {
1588   PathRef FileName;
1589   IncludeStructure Includes;           // Complete once the compiler runs.
1590   SpeculativeFuzzyFind *SpecFuzzyFind; // Can be nullptr.
1591   const CodeCompleteOptions &Opts;
1592 
1593   // Sema takes ownership of Recorder. Recorder is valid until Sema cleanup.
1594   CompletionRecorder *Recorder = nullptr;
1595   CodeCompletionContext::Kind CCContextKind = CodeCompletionContext::CCC_Other;
1596   bool IsUsingDeclaration = false;
1597   // The snippets will not be generated if the token following completion
1598   // location is an opening parenthesis (tok::l_paren) because this would add
1599   // extra parenthesis.
1600   tok::TokenKind NextTokenKind = tok::eof;
1601   // Counters for logging.
1602   int NSema = 0, NIndex = 0, NSemaAndIndex = 0, NIdent = 0;
1603   bool Incomplete = false; // Would more be available with a higher limit?
1604   CompletionPrefix HeuristicPrefix;
1605   std::optional<FuzzyMatcher> Filter; // Initialized once Sema runs.
1606   Range ReplacedRange;
1607   std::vector<std::string> QueryScopes;      // Initialized once Sema runs.
1608   std::vector<std::string> AccessibleScopes; // Initialized once Sema runs.
1609   // Initialized once QueryScopes is initialized, if there are scopes.
1610   std::optional<ScopeDistance> ScopeProximity;
1611   std::optional<OpaqueType> PreferredType; // Initialized once Sema runs.
1612   // Whether to query symbols from any scope. Initialized once Sema runs.
1613   bool AllScopes = false;
1614   llvm::StringSet<> ContextWords;
1615   // Include-insertion and proximity scoring rely on the include structure.
1616   // This is available after Sema has run.
1617   std::optional<IncludeInserter> Inserter;  // Available during runWithSema.
1618   std::optional<URIDistance> FileProximity; // Initialized once Sema runs.
1619   /// Speculative request based on the cached request and the filter text before
1620   /// the cursor.
1621   /// Initialized right before sema run. This is only set if `SpecFuzzyFind` is
1622   /// set and contains a cached request.
1623   std::optional<FuzzyFindRequest> SpecReq;
1624 
1625 public:
1626   // A CodeCompleteFlow object is only useful for calling run() exactly once.
1627   CodeCompleteFlow(PathRef FileName, const IncludeStructure &Includes,
1628                    SpeculativeFuzzyFind *SpecFuzzyFind,
1629                    const CodeCompleteOptions &Opts)
1630       : FileName(FileName), Includes(Includes), SpecFuzzyFind(SpecFuzzyFind),
1631         Opts(Opts) {}
1632 
1633   CodeCompleteResult run(const SemaCompleteInput &SemaCCInput) && {
1634     trace::Span Tracer("CodeCompleteFlow");
1635     HeuristicPrefix = guessCompletionPrefix(SemaCCInput.ParseInput.Contents,
1636                                             SemaCCInput.Offset);
1637     populateContextWords(SemaCCInput.ParseInput.Contents);
1638     if (Opts.Index && SpecFuzzyFind && SpecFuzzyFind->CachedReq) {
1639       assert(!SpecFuzzyFind->Result.valid());
1640       SpecReq = speculativeFuzzyFindRequestForCompletion(
1641           *SpecFuzzyFind->CachedReq, HeuristicPrefix);
1642       SpecFuzzyFind->Result = startAsyncFuzzyFind(*Opts.Index, *SpecReq);
1643     }
1644 
1645     // We run Sema code completion first. It builds an AST and calculates:
1646     //   - completion results based on the AST.
1647     //   - partial identifier and context. We need these for the index query.
1648     CodeCompleteResult Output;
1649     auto RecorderOwner = std::make_unique<CompletionRecorder>(Opts, [&]() {
1650       assert(Recorder && "Recorder is not set");
1651       CCContextKind = Recorder->CCContext.getKind();
1652       IsUsingDeclaration = Recorder->CCContext.isUsingDeclaration();
1653       auto Style = getFormatStyleForFile(SemaCCInput.FileName,
1654                                          SemaCCInput.ParseInput.Contents,
1655                                          *SemaCCInput.ParseInput.TFS, false);
1656       const auto NextToken = findTokenAfterCompletionPoint(
1657           Recorder->CCSema->getPreprocessor().getCodeCompletionLoc(),
1658           Recorder->CCSema->getSourceManager(), Recorder->CCSema->LangOpts);
1659       if (NextToken)
1660         NextTokenKind = NextToken->getKind();
1661       // If preprocessor was run, inclusions from preprocessor callback should
1662       // already be added to Includes.
1663       Inserter.emplace(
1664           SemaCCInput.FileName, SemaCCInput.ParseInput.Contents, Style,
1665           SemaCCInput.ParseInput.CompileCommand.Directory,
1666           &Recorder->CCSema->getPreprocessor().getHeaderSearchInfo(),
1667           Config::current().Style.QuotedHeaders,
1668           Config::current().Style.AngledHeaders);
1669       for (const auto &Inc : Includes.MainFileIncludes)
1670         Inserter->addExisting(Inc);
1671 
1672       // Most of the cost of file proximity is in initializing the FileDistance
1673       // structures based on the observed includes, once per query. Conceptually
1674       // that happens here (though the per-URI-scheme initialization is lazy).
1675       // The per-result proximity scoring is (amortized) very cheap.
1676       FileDistanceOptions ProxOpts{}; // Use defaults.
1677       const auto &SM = Recorder->CCSema->getSourceManager();
1678       llvm::StringMap<SourceParams> ProxSources;
1679       auto MainFileID =
1680           Includes.getID(SM.getFileEntryForID(SM.getMainFileID()));
1681       assert(MainFileID);
1682       for (auto &HeaderIDAndDepth : Includes.includeDepth(*MainFileID)) {
1683         auto &Source =
1684             ProxSources[Includes.getRealPath(HeaderIDAndDepth.getFirst())];
1685         Source.Cost = HeaderIDAndDepth.getSecond() * ProxOpts.IncludeCost;
1686         // Symbols near our transitive includes are good, but only consider
1687         // things in the same directory or below it. Otherwise there can be
1688         // many false positives.
1689         if (HeaderIDAndDepth.getSecond() > 0)
1690           Source.MaxUpTraversals = 1;
1691       }
1692       FileProximity.emplace(ProxSources, ProxOpts);
1693 
1694       Output = runWithSema();
1695       Inserter.reset(); // Make sure this doesn't out-live Clang.
1696       SPAN_ATTACH(Tracer, "sema_completion_kind",
1697                   getCompletionKindString(CCContextKind));
1698       log("Code complete: sema context {0}, query scopes [{1}] (AnyScope={2}), "
1699           "expected type {3}{4}",
1700           getCompletionKindString(CCContextKind),
1701           llvm::join(QueryScopes.begin(), QueryScopes.end(), ","), AllScopes,
1702           PreferredType ? Recorder->CCContext.getPreferredType().getAsString()
1703                         : "<none>",
1704           IsUsingDeclaration ? ", inside using declaration" : "");
1705     });
1706 
1707     Recorder = RecorderOwner.get();
1708 
1709     semaCodeComplete(std::move(RecorderOwner), Opts.getClangCompleteOpts(),
1710                      SemaCCInput, &Includes);
1711     logResults(Output, Tracer);
1712     return Output;
1713   }
1714 
1715   void logResults(const CodeCompleteResult &Output, const trace::Span &Tracer) {
1716     SPAN_ATTACH(Tracer, "sema_results", NSema);
1717     SPAN_ATTACH(Tracer, "index_results", NIndex);
1718     SPAN_ATTACH(Tracer, "merged_results", NSemaAndIndex);
1719     SPAN_ATTACH(Tracer, "identifier_results", NIdent);
1720     SPAN_ATTACH(Tracer, "returned_results", int64_t(Output.Completions.size()));
1721     SPAN_ATTACH(Tracer, "incomplete", Output.HasMore);
1722     log("Code complete: {0} results from Sema, {1} from Index, "
1723         "{2} matched, {3} from identifiers, {4} returned{5}.",
1724         NSema, NIndex, NSemaAndIndex, NIdent, Output.Completions.size(),
1725         Output.HasMore ? " (incomplete)" : "");
1726     assert(!Opts.Limit || Output.Completions.size() <= Opts.Limit);
1727     // We don't assert that isIncomplete means we hit a limit.
1728     // Indexes may choose to impose their own limits even if we don't have one.
1729   }
1730 
1731   CodeCompleteResult runWithoutSema(llvm::StringRef Content, size_t Offset,
1732                                     const ThreadsafeFS &TFS) && {
1733     trace::Span Tracer("CodeCompleteWithoutSema");
1734     // Fill in fields normally set by runWithSema()
1735     HeuristicPrefix = guessCompletionPrefix(Content, Offset);
1736     populateContextWords(Content);
1737     CCContextKind = CodeCompletionContext::CCC_Recovery;
1738     IsUsingDeclaration = false;
1739     Filter = FuzzyMatcher(HeuristicPrefix.Name);
1740     auto Pos = offsetToPosition(Content, Offset);
1741     ReplacedRange.start = ReplacedRange.end = Pos;
1742     ReplacedRange.start.character -= HeuristicPrefix.Name.size();
1743 
1744     llvm::StringMap<SourceParams> ProxSources;
1745     ProxSources[FileName].Cost = 0;
1746     FileProximity.emplace(ProxSources);
1747 
1748     auto Style = getFormatStyleForFile(FileName, Content, TFS, false);
1749     // This will only insert verbatim headers.
1750     Inserter.emplace(FileName, Content, Style,
1751                      /*BuildDir=*/"", /*HeaderSearchInfo=*/nullptr,
1752                      Config::current().Style.QuotedHeaders,
1753                      Config::current().Style.AngledHeaders);
1754 
1755     auto Identifiers = collectIdentifiers(Content, Style);
1756     std::vector<RawIdentifier> IdentifierResults;
1757     for (const auto &IDAndCount : Identifiers) {
1758       RawIdentifier ID;
1759       ID.Name = IDAndCount.first();
1760       ID.References = IDAndCount.second;
1761       // Avoid treating typed filter as an identifier.
1762       if (ID.Name == HeuristicPrefix.Name)
1763         --ID.References;
1764       if (ID.References > 0)
1765         IdentifierResults.push_back(std::move(ID));
1766     }
1767 
1768     // Simplified version of getQueryScopes():
1769     //  - accessible scopes are determined heuristically.
1770     //  - all-scopes query if no qualifier was typed (and it's allowed).
1771     SpecifiedScope Scopes;
1772     Scopes.QueryScopes = visibleNamespaces(
1773         Content.take_front(Offset), format::getFormattingLangOpts(Style));
1774     for (std::string &S : Scopes.QueryScopes)
1775       if (!S.empty())
1776         S.append("::"); // visibleNamespaces doesn't include trailing ::.
1777     if (HeuristicPrefix.Qualifier.empty())
1778       AllScopes = Opts.AllScopes;
1779     else if (HeuristicPrefix.Qualifier.starts_with("::")) {
1780       Scopes.QueryScopes = {""};
1781       Scopes.UnresolvedQualifier =
1782           std::string(HeuristicPrefix.Qualifier.drop_front(2));
1783     } else
1784       Scopes.UnresolvedQualifier = std::string(HeuristicPrefix.Qualifier);
1785     // First scope is the (modified) enclosing scope.
1786     QueryScopes = Scopes.scopesForIndexQuery();
1787     AccessibleScopes = QueryScopes;
1788     ScopeProximity.emplace(QueryScopes);
1789 
1790     SymbolSlab IndexResults = Opts.Index ? queryIndex() : SymbolSlab();
1791 
1792     CodeCompleteResult Output = toCodeCompleteResult(mergeResults(
1793         /*SemaResults=*/{}, IndexResults, IdentifierResults));
1794     Output.RanParser = false;
1795     logResults(Output, Tracer);
1796     return Output;
1797   }
1798 
1799 private:
1800   void populateContextWords(llvm::StringRef Content) {
1801     // Take last 3 lines before the completion point.
1802     unsigned RangeEnd = HeuristicPrefix.Qualifier.begin() - Content.data(),
1803              RangeBegin = RangeEnd;
1804     for (size_t I = 0; I < 3 && RangeBegin > 0; ++I) {
1805       auto PrevNL = Content.rfind('\n', RangeBegin);
1806       if (PrevNL == StringRef::npos) {
1807         RangeBegin = 0;
1808         break;
1809       }
1810       RangeBegin = PrevNL;
1811     }
1812 
1813     ContextWords = collectWords(Content.slice(RangeBegin, RangeEnd));
1814     dlog("Completion context words: {0}",
1815          llvm::join(ContextWords.keys(), ", "));
1816   }
1817 
1818   // This is called by run() once Sema code completion is done, but before the
1819   // Sema data structures are torn down. It does all the real work.
1820   CodeCompleteResult runWithSema() {
1821     const auto &CodeCompletionRange = CharSourceRange::getCharRange(
1822         Recorder->CCSema->getPreprocessor().getCodeCompletionTokenRange());
1823     // When we are getting completions with an empty identifier, for example
1824     //    std::vector<int> asdf;
1825     //    asdf.^;
1826     // Then the range will be invalid and we will be doing insertion, use
1827     // current cursor position in such cases as range.
1828     if (CodeCompletionRange.isValid()) {
1829       ReplacedRange = halfOpenToRange(Recorder->CCSema->getSourceManager(),
1830                                       CodeCompletionRange);
1831     } else {
1832       const auto &Pos = sourceLocToPosition(
1833           Recorder->CCSema->getSourceManager(),
1834           Recorder->CCSema->getPreprocessor().getCodeCompletionLoc());
1835       ReplacedRange.start = ReplacedRange.end = Pos;
1836     }
1837     Filter = FuzzyMatcher(
1838         Recorder->CCSema->getPreprocessor().getCodeCompletionFilter());
1839     auto SpecifiedScopes = getQueryScopes(
1840         Recorder->CCContext, *Recorder->CCSema, HeuristicPrefix, Opts);
1841 
1842     QueryScopes = SpecifiedScopes.scopesForIndexQuery();
1843     AccessibleScopes = SpecifiedScopes.scopesForQualification();
1844     AllScopes = SpecifiedScopes.AllowAllScopes;
1845     if (!QueryScopes.empty())
1846       ScopeProximity.emplace(QueryScopes);
1847     PreferredType =
1848         OpaqueType::fromType(Recorder->CCSema->getASTContext(),
1849                              Recorder->CCContext.getPreferredType());
1850     // Sema provides the needed context to query the index.
1851     // FIXME: in addition to querying for extra/overlapping symbols, we should
1852     //        explicitly request symbols corresponding to Sema results.
1853     //        We can use their signals even if the index can't suggest them.
1854     // We must copy index results to preserve them, but there are at most Limit.
1855     auto IndexResults = (Opts.Index && allowIndex(Recorder->CCContext))
1856                             ? queryIndex()
1857                             : SymbolSlab();
1858     trace::Span Tracer("Populate CodeCompleteResult");
1859     // Merge Sema and Index results, score them, and pick the winners.
1860     auto Top =
1861         mergeResults(Recorder->Results, IndexResults, /*Identifiers*/ {});
1862     return toCodeCompleteResult(Top);
1863   }
1864 
1865   CodeCompleteResult
1866   toCodeCompleteResult(const std::vector<ScoredBundle> &Scored) {
1867     CodeCompleteResult Output;
1868 
1869     // Convert the results to final form, assembling the expensive strings.
1870     // If necessary, search the index for documentation comments.
1871     LookupRequest Req;
1872     llvm::DenseMap<SymbolID, uint32_t> SymbolToCompletion;
1873     for (auto &C : Scored) {
1874       Output.Completions.push_back(toCodeCompletion(C.first));
1875       Output.Completions.back().Score = C.second;
1876       Output.Completions.back().CompletionTokenRange = ReplacedRange;
1877       if (Opts.Index && !Output.Completions.back().Documentation) {
1878         for (auto &Cand : C.first) {
1879           if (Cand.SemaResult &&
1880               Cand.SemaResult->Kind == CodeCompletionResult::RK_Declaration) {
1881             auto ID = clangd::getSymbolID(Cand.SemaResult->getDeclaration());
1882             if (!ID)
1883               continue;
1884             Req.IDs.insert(ID);
1885             SymbolToCompletion[ID] = Output.Completions.size() - 1;
1886           }
1887         }
1888       }
1889     }
1890     Output.HasMore = Incomplete;
1891     Output.Context = CCContextKind;
1892     Output.CompletionRange = ReplacedRange;
1893 
1894     // Look up documentation from the index.
1895     if (Opts.Index) {
1896       Opts.Index->lookup(Req, [&](const Symbol &S) {
1897         if (S.Documentation.empty())
1898           return;
1899         auto &C = Output.Completions[SymbolToCompletion.at(S.ID)];
1900         C.Documentation.emplace();
1901         parseDocumentation(S.Documentation, *C.Documentation);
1902       });
1903     }
1904 
1905     return Output;
1906   }
1907 
1908   SymbolSlab queryIndex() {
1909     trace::Span Tracer("Query index");
1910     SPAN_ATTACH(Tracer, "limit", int64_t(Opts.Limit));
1911 
1912     // Build the query.
1913     FuzzyFindRequest Req;
1914     if (Opts.Limit)
1915       Req.Limit = Opts.Limit;
1916     Req.Query = std::string(Filter->pattern());
1917     Req.RestrictForCodeCompletion = true;
1918     Req.Scopes = QueryScopes;
1919     Req.AnyScope = AllScopes;
1920     // FIXME: we should send multiple weighted paths here.
1921     Req.ProximityPaths.push_back(std::string(FileName));
1922     if (PreferredType)
1923       Req.PreferredTypes.push_back(std::string(PreferredType->raw()));
1924     vlog("Code complete: fuzzyFind({0:2})", toJSON(Req));
1925 
1926     if (SpecFuzzyFind)
1927       SpecFuzzyFind->NewReq = Req;
1928     if (SpecFuzzyFind && SpecFuzzyFind->Result.valid() && (*SpecReq == Req)) {
1929       vlog("Code complete: speculative fuzzy request matches the actual index "
1930            "request. Waiting for the speculative index results.");
1931       SPAN_ATTACH(Tracer, "Speculative results", true);
1932 
1933       trace::Span WaitSpec("Wait speculative results");
1934       auto SpecRes = SpecFuzzyFind->Result.get();
1935       Incomplete |= SpecRes.first;
1936       return std::move(SpecRes.second);
1937     }
1938 
1939     SPAN_ATTACH(Tracer, "Speculative results", false);
1940 
1941     // Run the query against the index.
1942     SymbolSlab::Builder ResultsBuilder;
1943     Incomplete |= Opts.Index->fuzzyFind(
1944         Req, [&](const Symbol &Sym) { ResultsBuilder.insert(Sym); });
1945     return std::move(ResultsBuilder).build();
1946   }
1947 
1948   // Merges Sema and Index results where possible, to form CompletionCandidates.
1949   // \p Identifiers is raw identifiers that can also be completion candidates.
1950   // Identifiers are not merged with results from index or sema.
1951   // Groups overloads if desired, to form CompletionCandidate::Bundles. The
1952   // bundles are scored and top results are returned, best to worst.
1953   std::vector<ScoredBundle>
1954   mergeResults(const std::vector<CodeCompletionResult> &SemaResults,
1955                const SymbolSlab &IndexResults,
1956                const std::vector<RawIdentifier> &IdentifierResults) {
1957     trace::Span Tracer("Merge and score results");
1958     std::vector<CompletionCandidate::Bundle> Bundles;
1959     llvm::DenseMap<size_t, size_t> BundleLookup;
1960     auto AddToBundles = [&](const CodeCompletionResult *SemaResult,
1961                             const Symbol *IndexResult,
1962                             const RawIdentifier *IdentifierResult) {
1963       CompletionCandidate C;
1964       C.SemaResult = SemaResult;
1965       C.IndexResult = IndexResult;
1966       C.IdentifierResult = IdentifierResult;
1967       if (C.IndexResult) {
1968         C.Name = IndexResult->Name;
1969         C.RankedIncludeHeaders = getRankedIncludes(*C.IndexResult);
1970       } else if (C.SemaResult) {
1971         C.Name = Recorder->getName(*SemaResult);
1972       } else {
1973         assert(IdentifierResult);
1974         C.Name = IdentifierResult->Name;
1975       }
1976       if (auto OverloadSet = C.overloadSet(
1977               Opts, FileName, Inserter ? &*Inserter : nullptr, CCContextKind)) {
1978         auto Ret = BundleLookup.try_emplace(OverloadSet, Bundles.size());
1979         if (Ret.second)
1980           Bundles.emplace_back();
1981         Bundles[Ret.first->second].push_back(std::move(C));
1982       } else {
1983         Bundles.emplace_back();
1984         Bundles.back().push_back(std::move(C));
1985       }
1986     };
1987     llvm::DenseSet<const Symbol *> UsedIndexResults;
1988     auto CorrespondingIndexResult =
1989         [&](const CodeCompletionResult &SemaResult) -> const Symbol * {
1990       if (auto SymID =
1991               getSymbolID(SemaResult, Recorder->CCSema->getSourceManager())) {
1992         auto I = IndexResults.find(SymID);
1993         if (I != IndexResults.end()) {
1994           UsedIndexResults.insert(&*I);
1995           return &*I;
1996         }
1997       }
1998       return nullptr;
1999     };
2000     // Emit all Sema results, merging them with Index results if possible.
2001     for (auto &SemaResult : SemaResults)
2002       AddToBundles(&SemaResult, CorrespondingIndexResult(SemaResult), nullptr);
2003     // Now emit any Index-only results.
2004     for (const auto &IndexResult : IndexResults) {
2005       if (UsedIndexResults.count(&IndexResult))
2006         continue;
2007       if (!includeSymbolFromIndex(CCContextKind, IndexResult))
2008         continue;
2009       AddToBundles(/*SemaResult=*/nullptr, &IndexResult, nullptr);
2010     }
2011     // Emit identifier results.
2012     for (const auto &Ident : IdentifierResults)
2013       AddToBundles(/*SemaResult=*/nullptr, /*IndexResult=*/nullptr, &Ident);
2014     // We only keep the best N results at any time, in "native" format.
2015     TopN<ScoredBundle, ScoredBundleGreater> Top(
2016         Opts.Limit == 0 ? std::numeric_limits<size_t>::max() : Opts.Limit);
2017     for (auto &Bundle : Bundles)
2018       addCandidate(Top, std::move(Bundle));
2019     return std::move(Top).items();
2020   }
2021 
2022   std::optional<float> fuzzyScore(const CompletionCandidate &C) {
2023     // Macros can be very spammy, so we only support prefix completion.
2024     if (((C.SemaResult &&
2025           C.SemaResult->Kind == CodeCompletionResult::RK_Macro) ||
2026          (C.IndexResult &&
2027           C.IndexResult->SymInfo.Kind == index::SymbolKind::Macro)) &&
2028         !C.Name.starts_with_insensitive(Filter->pattern()))
2029       return std::nullopt;
2030     return Filter->match(C.Name);
2031   }
2032 
2033   CodeCompletion::Scores
2034   evaluateCompletion(const SymbolQualitySignals &Quality,
2035                      const SymbolRelevanceSignals &Relevance) {
2036     using RM = CodeCompleteOptions::CodeCompletionRankingModel;
2037     CodeCompletion::Scores Scores;
2038     switch (Opts.RankingModel) {
2039     case RM::Heuristics:
2040       Scores.Quality = Quality.evaluateHeuristics();
2041       Scores.Relevance = Relevance.evaluateHeuristics();
2042       Scores.Total =
2043           evaluateSymbolAndRelevance(Scores.Quality, Scores.Relevance);
2044       // NameMatch is in fact a multiplier on total score, so rescoring is
2045       // sound.
2046       Scores.ExcludingName =
2047           Relevance.NameMatch > std::numeric_limits<float>::epsilon()
2048               ? Scores.Total / Relevance.NameMatch
2049               : Scores.Quality;
2050       return Scores;
2051 
2052     case RM::DecisionForest:
2053       DecisionForestScores DFScores = Opts.DecisionForestScorer(
2054           Quality, Relevance, Opts.DecisionForestBase);
2055       Scores.ExcludingName = DFScores.ExcludingName;
2056       Scores.Total = DFScores.Total;
2057       return Scores;
2058     }
2059     llvm_unreachable("Unhandled CodeCompletion ranking model.");
2060   }
2061 
2062   // Scores a candidate and adds it to the TopN structure.
2063   void addCandidate(TopN<ScoredBundle, ScoredBundleGreater> &Candidates,
2064                     CompletionCandidate::Bundle Bundle) {
2065     SymbolQualitySignals Quality;
2066     SymbolRelevanceSignals Relevance;
2067     Relevance.Context = CCContextKind;
2068     Relevance.Name = Bundle.front().Name;
2069     Relevance.FilterLength = HeuristicPrefix.Name.size();
2070     Relevance.Query = SymbolRelevanceSignals::CodeComplete;
2071     Relevance.FileProximityMatch = &*FileProximity;
2072     if (ScopeProximity)
2073       Relevance.ScopeProximityMatch = &*ScopeProximity;
2074     if (PreferredType)
2075       Relevance.HadContextType = true;
2076     Relevance.ContextWords = &ContextWords;
2077     Relevance.MainFileSignals = Opts.MainFileSignals;
2078 
2079     auto &First = Bundle.front();
2080     if (auto FuzzyScore = fuzzyScore(First))
2081       Relevance.NameMatch = *FuzzyScore;
2082     else
2083       return;
2084     SymbolOrigin Origin = SymbolOrigin::Unknown;
2085     bool FromIndex = false;
2086     for (const auto &Candidate : Bundle) {
2087       if (Candidate.IndexResult) {
2088         Quality.merge(*Candidate.IndexResult);
2089         Relevance.merge(*Candidate.IndexResult);
2090         Origin |= Candidate.IndexResult->Origin;
2091         FromIndex = true;
2092         if (!Candidate.IndexResult->Type.empty())
2093           Relevance.HadSymbolType |= true;
2094         if (PreferredType &&
2095             PreferredType->raw() == Candidate.IndexResult->Type) {
2096           Relevance.TypeMatchesPreferred = true;
2097         }
2098       }
2099       if (Candidate.SemaResult) {
2100         Quality.merge(*Candidate.SemaResult);
2101         Relevance.merge(*Candidate.SemaResult);
2102         if (PreferredType) {
2103           if (auto CompletionType = OpaqueType::fromCompletionResult(
2104                   Recorder->CCSema->getASTContext(), *Candidate.SemaResult)) {
2105             Relevance.HadSymbolType |= true;
2106             if (PreferredType == CompletionType)
2107               Relevance.TypeMatchesPreferred = true;
2108           }
2109         }
2110         Origin |= SymbolOrigin::AST;
2111       }
2112       if (Candidate.IdentifierResult) {
2113         Quality.References = Candidate.IdentifierResult->References;
2114         Relevance.Scope = SymbolRelevanceSignals::FileScope;
2115         Origin |= SymbolOrigin::Identifier;
2116       }
2117     }
2118 
2119     CodeCompletion::Scores Scores = evaluateCompletion(Quality, Relevance);
2120     if (Opts.RecordCCResult)
2121       Opts.RecordCCResult(toCodeCompletion(Bundle), Quality, Relevance,
2122                           Scores.Total);
2123 
2124     dlog("CodeComplete: {0} ({1}) = {2}\n{3}{4}\n", First.Name,
2125          llvm::to_string(Origin), Scores.Total, llvm::to_string(Quality),
2126          llvm::to_string(Relevance));
2127 
2128     NSema += bool(Origin & SymbolOrigin::AST);
2129     NIndex += FromIndex;
2130     NSemaAndIndex += bool(Origin & SymbolOrigin::AST) && FromIndex;
2131     NIdent += bool(Origin & SymbolOrigin::Identifier);
2132     if (Candidates.push({std::move(Bundle), Scores}))
2133       Incomplete = true;
2134   }
2135 
2136   CodeCompletion toCodeCompletion(const CompletionCandidate::Bundle &Bundle) {
2137     std::optional<CodeCompletionBuilder> Builder;
2138     for (const auto &Item : Bundle) {
2139       CodeCompletionString *SemaCCS =
2140           Item.SemaResult ? Recorder->codeCompletionString(*Item.SemaResult)
2141                           : nullptr;
2142       if (!Builder)
2143         Builder.emplace(Recorder ? &Recorder->CCSema->getASTContext() : nullptr,
2144                         Item, SemaCCS, AccessibleScopes, *Inserter, FileName,
2145                         CCContextKind, Opts, IsUsingDeclaration, NextTokenKind);
2146       else
2147         Builder->add(Item, SemaCCS, CCContextKind);
2148     }
2149     return Builder->build();
2150   }
2151 };
2152 
2153 } // namespace
2154 
2155 clang::CodeCompleteOptions CodeCompleteOptions::getClangCompleteOpts() const {
2156   clang::CodeCompleteOptions Result;
2157   Result.IncludeCodePatterns = EnableSnippets;
2158   Result.IncludeMacros = true;
2159   Result.IncludeGlobals = true;
2160   // We choose to include full comments and not do doxygen parsing in
2161   // completion.
2162   // FIXME: ideally, we should support doxygen in some form, e.g. do markdown
2163   // formatting of the comments.
2164   Result.IncludeBriefComments = false;
2165 
2166   // When an is used, Sema is responsible for completing the main file,
2167   // the index can provide results from the preamble.
2168   // Tell Sema not to deserialize the preamble to look for results.
2169   Result.LoadExternal = ForceLoadPreamble || !Index;
2170   Result.IncludeFixIts = IncludeFixIts;
2171 
2172   return Result;
2173 }
2174 
2175 CompletionPrefix guessCompletionPrefix(llvm::StringRef Content,
2176                                        unsigned Offset) {
2177   assert(Offset <= Content.size());
2178   StringRef Rest = Content.take_front(Offset);
2179   CompletionPrefix Result;
2180 
2181   // Consume the unqualified name. We only handle ASCII characters.
2182   // isAsciiIdentifierContinue will let us match "0invalid", but we don't mind.
2183   while (!Rest.empty() && isAsciiIdentifierContinue(Rest.back()))
2184     Rest = Rest.drop_back();
2185   Result.Name = Content.slice(Rest.size(), Offset);
2186 
2187   // Consume qualifiers.
2188   while (Rest.consume_back("::") && !Rest.ends_with(":")) // reject ::::
2189     while (!Rest.empty() && isAsciiIdentifierContinue(Rest.back()))
2190       Rest = Rest.drop_back();
2191   Result.Qualifier =
2192       Content.slice(Rest.size(), Result.Name.begin() - Content.begin());
2193 
2194   return Result;
2195 }
2196 
2197 // Code complete the argument name on "/*" inside function call.
2198 // Offset should be pointing to the start of the comment, i.e.:
2199 // foo(^/*, rather than foo(/*^) where the cursor probably is.
2200 CodeCompleteResult codeCompleteComment(PathRef FileName, unsigned Offset,
2201                                        llvm::StringRef Prefix,
2202                                        const PreambleData *Preamble,
2203                                        const ParseInputs &ParseInput) {
2204   if (Preamble == nullptr) // Can't run without Sema.
2205     return CodeCompleteResult();
2206 
2207   clang::CodeCompleteOptions Options;
2208   Options.IncludeGlobals = false;
2209   Options.IncludeMacros = false;
2210   Options.IncludeCodePatterns = false;
2211   Options.IncludeBriefComments = false;
2212   std::set<std::string> ParamNames;
2213   // We want to see signatures coming from newly introduced includes, hence a
2214   // full patch.
2215   semaCodeComplete(
2216       std::make_unique<ParamNameCollector>(Options, ParamNames), Options,
2217       {FileName, Offset, *Preamble,
2218        PreamblePatch::createFullPatch(FileName, ParseInput, *Preamble),
2219        ParseInput});
2220   if (ParamNames.empty())
2221     return CodeCompleteResult();
2222 
2223   CodeCompleteResult Result;
2224   Range CompletionRange;
2225   // Skip /*
2226   Offset += 2;
2227   CompletionRange.start = offsetToPosition(ParseInput.Contents, Offset);
2228   CompletionRange.end =
2229       offsetToPosition(ParseInput.Contents, Offset + Prefix.size());
2230   Result.CompletionRange = CompletionRange;
2231   Result.Context = CodeCompletionContext::CCC_NaturalLanguage;
2232   for (llvm::StringRef Name : ParamNames) {
2233     if (!Name.starts_with(Prefix))
2234       continue;
2235     CodeCompletion Item;
2236     Item.Name = Name.str() + "=*/";
2237     Item.FilterText = Item.Name;
2238     Item.Kind = CompletionItemKind::Text;
2239     Item.CompletionTokenRange = CompletionRange;
2240     Item.Origin = SymbolOrigin::AST;
2241     Result.Completions.push_back(Item);
2242   }
2243 
2244   return Result;
2245 }
2246 
2247 // If Offset is inside what looks like argument comment (e.g.
2248 // "/*^" or "/* foo^"), returns new offset pointing to the start of the /*
2249 // (place where semaCodeComplete should run).
2250 std::optional<unsigned>
2251 maybeFunctionArgumentCommentStart(llvm::StringRef Content) {
2252   while (!Content.empty() && isAsciiIdentifierContinue(Content.back()))
2253     Content = Content.drop_back();
2254   Content = Content.rtrim();
2255   if (Content.ends_with("/*"))
2256     return Content.size() - 2;
2257   return std::nullopt;
2258 }
2259 
2260 CodeCompleteResult codeComplete(PathRef FileName, Position Pos,
2261                                 const PreambleData *Preamble,
2262                                 const ParseInputs &ParseInput,
2263                                 CodeCompleteOptions Opts,
2264                                 SpeculativeFuzzyFind *SpecFuzzyFind) {
2265   auto Offset = positionToOffset(ParseInput.Contents, Pos);
2266   if (!Offset) {
2267     elog("Code completion position was invalid {0}", Offset.takeError());
2268     return CodeCompleteResult();
2269   }
2270 
2271   auto Content = llvm::StringRef(ParseInput.Contents).take_front(*Offset);
2272   if (auto OffsetBeforeComment = maybeFunctionArgumentCommentStart(Content)) {
2273     // We are doing code completion of a comment, where we currently only
2274     // support completing param names in function calls. To do this, we
2275     // require information from Sema, but Sema's comment completion stops at
2276     // parsing, so we must move back the position before running it, extract
2277     // information we need and construct completion items ourselves.
2278     auto CommentPrefix = Content.substr(*OffsetBeforeComment + 2).trim();
2279     return codeCompleteComment(FileName, *OffsetBeforeComment, CommentPrefix,
2280                                Preamble, ParseInput);
2281   }
2282 
2283   auto Flow = CodeCompleteFlow(
2284       FileName, Preamble ? Preamble->Includes : IncludeStructure(),
2285       SpecFuzzyFind, Opts);
2286   return (!Preamble || Opts.RunParser == CodeCompleteOptions::NeverParse)
2287              ? std::move(Flow).runWithoutSema(ParseInput.Contents, *Offset,
2288                                               *ParseInput.TFS)
2289              : std::move(Flow).run({FileName, *Offset, *Preamble,
2290                                     /*PreamblePatch=*/
2291                                     PreamblePatch::createMacroPatch(
2292                                         FileName, ParseInput, *Preamble),
2293                                     ParseInput});
2294 }
2295 
2296 SignatureHelp signatureHelp(PathRef FileName, Position Pos,
2297                             const PreambleData &Preamble,
2298                             const ParseInputs &ParseInput,
2299                             MarkupKind DocumentationFormat) {
2300   auto Offset = positionToOffset(ParseInput.Contents, Pos);
2301   if (!Offset) {
2302     elog("Signature help position was invalid {0}", Offset.takeError());
2303     return SignatureHelp();
2304   }
2305   SignatureHelp Result;
2306   clang::CodeCompleteOptions Options;
2307   Options.IncludeGlobals = false;
2308   Options.IncludeMacros = false;
2309   Options.IncludeCodePatterns = false;
2310   Options.IncludeBriefComments = false;
2311   semaCodeComplete(
2312       std::make_unique<SignatureHelpCollector>(Options, DocumentationFormat,
2313                                                ParseInput.Index, Result),
2314       Options,
2315       {FileName, *Offset, Preamble,
2316        PreamblePatch::createFullPatch(FileName, ParseInput, Preamble),
2317        ParseInput});
2318   return Result;
2319 }
2320 
2321 bool isIndexedForCodeCompletion(const NamedDecl &ND, ASTContext &ASTCtx) {
2322   auto InTopLevelScope = [](const NamedDecl &ND) {
2323     switch (ND.getDeclContext()->getDeclKind()) {
2324     case Decl::TranslationUnit:
2325     case Decl::Namespace:
2326     case Decl::LinkageSpec:
2327       return true;
2328     default:
2329       break;
2330     };
2331     return false;
2332   };
2333   auto InClassScope = [](const NamedDecl &ND) {
2334     return ND.getDeclContext()->getDeclKind() == Decl::CXXRecord;
2335   };
2336   // We only complete symbol's name, which is the same as the name of the
2337   // *primary* template in case of template specializations.
2338   if (isExplicitTemplateSpecialization(&ND))
2339     return false;
2340 
2341   // Category decls are not useful on their own outside the interface or
2342   // implementation blocks. Moreover, sema already provides completion for
2343   // these, even if it requires preamble deserialization. So by excluding them
2344   // from the index, we reduce the noise in all the other completion scopes.
2345   if (llvm::isa<ObjCCategoryDecl>(&ND) || llvm::isa<ObjCCategoryImplDecl>(&ND))
2346     return false;
2347 
2348   if (InTopLevelScope(ND))
2349     return true;
2350 
2351   // Always index enum constants, even if they're not in the top level scope:
2352   // when
2353   // --all-scopes-completion is set, we'll want to complete those as well.
2354   if (const auto *EnumDecl = dyn_cast<clang::EnumDecl>(ND.getDeclContext()))
2355     return (InTopLevelScope(*EnumDecl) || InClassScope(*EnumDecl));
2356 
2357   return false;
2358 }
2359 
2360 CompletionItem CodeCompletion::render(const CodeCompleteOptions &Opts) const {
2361   CompletionItem LSP;
2362   const auto *InsertInclude = Includes.empty() ? nullptr : &Includes[0];
2363   // We could move our indicators from label into labelDetails->description.
2364   // In VSCode there are rendering issues that prevent these being aligned.
2365   LSP.label = ((InsertInclude && InsertInclude->Insertion)
2366                    ? Opts.IncludeIndicator.Insert
2367                    : Opts.IncludeIndicator.NoInsert) +
2368               (Opts.ShowOrigins ? "[" + llvm::to_string(Origin) + "]" : "") +
2369               RequiredQualifier + Name;
2370   LSP.labelDetails.emplace();
2371   LSP.labelDetails->detail = Signature;
2372 
2373   LSP.kind = Kind;
2374   LSP.detail = BundleSize > 1
2375                    ? std::string(llvm::formatv("[{0} overloads]", BundleSize))
2376                    : ReturnType;
2377   LSP.deprecated = Deprecated;
2378   // Combine header information and documentation in LSP `documentation` field.
2379   // This is not quite right semantically, but tends to display well in editors.
2380   if (InsertInclude || Documentation) {
2381     markup::Document Doc;
2382     if (InsertInclude)
2383       Doc.addParagraph().appendText("From ").appendCode(InsertInclude->Header);
2384     if (Documentation)
2385       Doc.append(*Documentation);
2386     LSP.documentation = renderDoc(Doc, Opts.DocumentationFormat);
2387   }
2388   LSP.sortText = sortText(Score.Total, FilterText);
2389   LSP.filterText = FilterText;
2390   LSP.textEdit = {CompletionTokenRange, RequiredQualifier + Name, ""};
2391   // Merge continuous additionalTextEdits into main edit. The main motivation
2392   // behind this is to help LSP clients, it seems most of them are confused when
2393   // they are provided with additionalTextEdits that are consecutive to main
2394   // edit.
2395   // Note that we store additional text edits from back to front in a line. That
2396   // is mainly to help LSP clients again, so that changes do not effect each
2397   // other.
2398   for (const auto &FixIt : FixIts) {
2399     if (FixIt.range.end == LSP.textEdit->range.start) {
2400       LSP.textEdit->newText = FixIt.newText + LSP.textEdit->newText;
2401       LSP.textEdit->range.start = FixIt.range.start;
2402     } else {
2403       LSP.additionalTextEdits.push_back(FixIt);
2404     }
2405   }
2406   if (Opts.EnableSnippets)
2407     LSP.textEdit->newText += SnippetSuffix;
2408 
2409   // FIXME(kadircet): Do not even fill insertText after making sure textEdit is
2410   // compatible with most of the editors.
2411   LSP.insertText = LSP.textEdit->newText;
2412   // Some clients support snippets but work better with plaintext.
2413   // So if the snippet is trivial, let the client know.
2414   // https://github.com/clangd/clangd/issues/922
2415   LSP.insertTextFormat = (Opts.EnableSnippets && !SnippetSuffix.empty())
2416                              ? InsertTextFormat::Snippet
2417                              : InsertTextFormat::PlainText;
2418   if (InsertInclude && InsertInclude->Insertion)
2419     LSP.additionalTextEdits.push_back(*InsertInclude->Insertion);
2420 
2421   LSP.score = Score.ExcludingName;
2422 
2423   return LSP;
2424 }
2425 
2426 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, const CodeCompletion &C) {
2427   // For now just lean on CompletionItem.
2428   return OS << C.render(CodeCompleteOptions());
2429 }
2430 
2431 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
2432                               const CodeCompleteResult &R) {
2433   OS << "CodeCompleteResult: " << R.Completions.size() << (R.HasMore ? "+" : "")
2434      << " (" << getCompletionKindString(R.Context) << ")"
2435      << " items:\n";
2436   for (const auto &C : R.Completions)
2437     OS << C << "\n";
2438   return OS;
2439 }
2440 
2441 // Heuristically detect whether the `Line` is an unterminated include filename.
2442 bool isIncludeFile(llvm::StringRef Line) {
2443   Line = Line.ltrim();
2444   if (!Line.consume_front("#"))
2445     return false;
2446   Line = Line.ltrim();
2447   if (!(Line.consume_front("include_next") || Line.consume_front("include") ||
2448         Line.consume_front("import")))
2449     return false;
2450   Line = Line.ltrim();
2451   if (Line.consume_front("<"))
2452     return Line.count('>') == 0;
2453   if (Line.consume_front("\""))
2454     return Line.count('"') == 0;
2455   return false;
2456 }
2457 
2458 bool allowImplicitCompletion(llvm::StringRef Content, unsigned Offset) {
2459   // Look at last line before completion point only.
2460   Content = Content.take_front(Offset);
2461   auto Pos = Content.rfind('\n');
2462   if (Pos != llvm::StringRef::npos)
2463     Content = Content.substr(Pos + 1);
2464 
2465   // Complete after scope operators.
2466   if (Content.ends_with(".") || Content.ends_with("->") ||
2467       Content.ends_with("::") || Content.ends_with("/*"))
2468     return true;
2469   // Complete after `#include <` and #include `<foo/`.
2470   if ((Content.ends_with("<") || Content.ends_with("\"") ||
2471        Content.ends_with("/")) &&
2472       isIncludeFile(Content))
2473     return true;
2474 
2475   // Complete words. Give non-ascii characters the benefit of the doubt.
2476   return !Content.empty() && (isAsciiIdentifierContinue(Content.back()) ||
2477                               !llvm::isASCII(Content.back()));
2478 }
2479 
2480 } // namespace clangd
2481 } // namespace clang
2482