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 "Diagnostics.h"
25 #include "ExpectedTypes.h"
26 #include "FileDistance.h"
27 #include "FuzzyMatch.h"
28 #include "Headers.h"
29 #include "Hover.h"
30 #include "Preamble.h"
31 #include "Protocol.h"
32 #include "Quality.h"
33 #include "SourceCode.h"
34 #include "TUScheduler.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/Threading.h"
41 #include "support/ThreadsafeFS.h"
42 #include "support/Trace.h"
43 #include "clang/AST/Decl.h"
44 #include "clang/AST/DeclBase.h"
45 #include "clang/Basic/CharInfo.h"
46 #include "clang/Basic/LangOptions.h"
47 #include "clang/Basic/SourceLocation.h"
48 #include "clang/Basic/TokenKinds.h"
49 #include "clang/Format/Format.h"
50 #include "clang/Frontend/CompilerInstance.h"
51 #include "clang/Frontend/FrontendActions.h"
52 #include "clang/Lex/ExternalPreprocessorSource.h"
53 #include "clang/Lex/Lexer.h"
54 #include "clang/Lex/Preprocessor.h"
55 #include "clang/Lex/PreprocessorOptions.h"
56 #include "clang/Sema/CodeCompleteConsumer.h"
57 #include "clang/Sema/DeclSpec.h"
58 #include "clang/Sema/Sema.h"
59 #include "llvm/ADT/ArrayRef.h"
60 #include "llvm/ADT/None.h"
61 #include "llvm/ADT/Optional.h"
62 #include "llvm/ADT/SmallVector.h"
63 #include "llvm/ADT/StringExtras.h"
64 #include "llvm/ADT/StringRef.h"
65 #include "llvm/Support/Compiler.h"
66 #include "llvm/Support/Debug.h"
67 #include "llvm/Support/Error.h"
68 #include "llvm/Support/Format.h"
69 #include "llvm/Support/FormatVariadic.h"
70 #include "llvm/Support/ScopedPrinter.h"
71 #include <algorithm>
72 #include <iterator>
73 
74 // We log detailed candidate here if you run with -debug-only=codecomplete.
75 #define DEBUG_TYPE "CodeComplete"
76 
77 namespace clang {
78 namespace clangd {
79 namespace {
80 
81 CompletionItemKind toCompletionItemKind(index::SymbolKind Kind) {
82   using SK = index::SymbolKind;
83   switch (Kind) {
84   case SK::Unknown:
85     return CompletionItemKind::Missing;
86   case SK::Module:
87   case SK::Namespace:
88   case SK::NamespaceAlias:
89     return CompletionItemKind::Module;
90   case SK::Macro:
91     return CompletionItemKind::Text;
92   case SK::Enum:
93     return CompletionItemKind::Enum;
94   case SK::Struct:
95     return CompletionItemKind::Struct;
96   case SK::Class:
97   case SK::Protocol:
98   case SK::Extension:
99   case SK::Union:
100     return CompletionItemKind::Class;
101   case SK::TypeAlias:
102     // We use the same kind as the VSCode C++ extension.
103     // FIXME: pick a better option when we have one.
104     return CompletionItemKind::Interface;
105   case SK::Using:
106     return CompletionItemKind::Reference;
107   case SK::Function:
108   case SK::ConversionFunction:
109     return CompletionItemKind::Function;
110   case SK::Variable:
111   case SK::Parameter:
112   case SK::NonTypeTemplateParm:
113     return CompletionItemKind::Variable;
114   case SK::Field:
115     return CompletionItemKind::Field;
116   case SK::EnumConstant:
117     return CompletionItemKind::EnumMember;
118   case SK::InstanceMethod:
119   case SK::ClassMethod:
120   case SK::StaticMethod:
121   case SK::Destructor:
122     return CompletionItemKind::Method;
123   case SK::InstanceProperty:
124   case SK::ClassProperty:
125   case SK::StaticProperty:
126     return CompletionItemKind::Property;
127   case SK::Constructor:
128     return CompletionItemKind::Constructor;
129   case SK::TemplateTypeParm:
130   case SK::TemplateTemplateParm:
131     return CompletionItemKind::TypeParameter;
132   }
133   llvm_unreachable("Unhandled clang::index::SymbolKind.");
134 }
135 
136 CompletionItemKind
137 toCompletionItemKind(CodeCompletionResult::ResultKind ResKind,
138                      const NamedDecl *Decl,
139                      CodeCompletionContext::Kind CtxKind) {
140   if (Decl)
141     return toCompletionItemKind(index::getSymbolInfo(Decl).Kind);
142   if (CtxKind == CodeCompletionContext::CCC_IncludedFile)
143     return CompletionItemKind::File;
144   switch (ResKind) {
145   case CodeCompletionResult::RK_Declaration:
146     llvm_unreachable("RK_Declaration without Decl");
147   case CodeCompletionResult::RK_Keyword:
148     return CompletionItemKind::Keyword;
149   case CodeCompletionResult::RK_Macro:
150     return CompletionItemKind::Text; // unfortunately, there's no 'Macro'
151                                      // completion items in LSP.
152   case CodeCompletionResult::RK_Pattern:
153     return CompletionItemKind::Snippet;
154   }
155   llvm_unreachable("Unhandled CodeCompletionResult::ResultKind.");
156 }
157 
158 // Identifier code completion result.
159 struct RawIdentifier {
160   llvm::StringRef Name;
161   unsigned References; // # of usages in file.
162 };
163 
164 /// A code completion result, in clang-native form.
165 /// It may be promoted to a CompletionItem if it's among the top-ranked results.
166 struct CompletionCandidate {
167   llvm::StringRef Name; // Used for filtering and sorting.
168   // We may have a result from Sema, from the index, or both.
169   const CodeCompletionResult *SemaResult = nullptr;
170   const Symbol *IndexResult = nullptr;
171   const RawIdentifier *IdentifierResult = nullptr;
172   llvm::SmallVector<llvm::StringRef, 1> RankedIncludeHeaders;
173 
174   // Returns a token identifying the overload set this is part of.
175   // 0 indicates it's not part of any overload set.
176   size_t overloadSet(const CodeCompleteOptions &Opts) const {
177     if (!Opts.BundleOverloads.getValueOr(false))
178       return 0;
179     llvm::SmallString<256> Scratch;
180     if (IndexResult) {
181       switch (IndexResult->SymInfo.Kind) {
182       case index::SymbolKind::ClassMethod:
183       case index::SymbolKind::InstanceMethod:
184       case index::SymbolKind::StaticMethod:
185 #ifndef NDEBUG
186         llvm_unreachable("Don't expect members from index in code completion");
187 #else
188         LLVM_FALLTHROUGH;
189 #endif
190       case index::SymbolKind::Function:
191         // We can't group overloads together that need different #includes.
192         // This could break #include insertion.
193         return llvm::hash_combine(
194             (IndexResult->Scope + IndexResult->Name).toStringRef(Scratch),
195             headerToInsertIfAllowed(Opts).getValueOr(""));
196       default:
197         return 0;
198       }
199     }
200     if (SemaResult) {
201       // We need to make sure we're consistent with the IndexResult case!
202       const NamedDecl *D = SemaResult->Declaration;
203       if (!D || !D->isFunctionOrFunctionTemplate())
204         return 0;
205       {
206         llvm::raw_svector_ostream OS(Scratch);
207         D->printQualifiedName(OS);
208       }
209       return llvm::hash_combine(Scratch,
210                                 headerToInsertIfAllowed(Opts).getValueOr(""));
211     }
212     assert(IdentifierResult);
213     return 0;
214   }
215 
216   // The best header to include if include insertion is allowed.
217   llvm::Optional<llvm::StringRef>
218   headerToInsertIfAllowed(const CodeCompleteOptions &Opts) const {
219     if (Opts.InsertIncludes == CodeCompleteOptions::NeverInsert ||
220         RankedIncludeHeaders.empty())
221       return None;
222     if (SemaResult && SemaResult->Declaration) {
223       // Avoid inserting new #include if the declaration is found in the current
224       // file e.g. the symbol is forward declared.
225       auto &SM = SemaResult->Declaration->getASTContext().getSourceManager();
226       for (const Decl *RD : SemaResult->Declaration->redecls())
227         if (SM.isInMainFile(SM.getExpansionLoc(RD->getBeginLoc())))
228           return None;
229     }
230     return RankedIncludeHeaders[0];
231   }
232 
233   using Bundle = llvm::SmallVector<CompletionCandidate, 4>;
234 };
235 using ScoredBundle =
236     std::pair<CompletionCandidate::Bundle, CodeCompletion::Scores>;
237 struct ScoredBundleGreater {
238   bool operator()(const ScoredBundle &L, const ScoredBundle &R) {
239     if (L.second.Total != R.second.Total)
240       return L.second.Total > R.second.Total;
241     return L.first.front().Name <
242            R.first.front().Name; // Earlier name is better.
243   }
244 };
245 
246 // Assembles a code completion out of a bundle of >=1 completion candidates.
247 // Many of the expensive strings are only computed at this point, once we know
248 // the candidate bundle is going to be returned.
249 //
250 // Many fields are the same for all candidates in a bundle (e.g. name), and are
251 // computed from the first candidate, in the constructor.
252 // Others vary per candidate, so add() must be called for remaining candidates.
253 struct CodeCompletionBuilder {
254   CodeCompletionBuilder(ASTContext *ASTCtx, const CompletionCandidate &C,
255                         CodeCompletionString *SemaCCS,
256                         llvm::ArrayRef<std::string> QueryScopes,
257                         const IncludeInserter &Includes,
258                         llvm::StringRef FileName,
259                         CodeCompletionContext::Kind ContextKind,
260                         const CodeCompleteOptions &Opts,
261                         bool IsUsingDeclaration, tok::TokenKind NextTokenKind)
262       : ASTCtx(ASTCtx), ExtractDocumentation(Opts.IncludeComments),
263         EnableFunctionArgSnippets(Opts.EnableFunctionArgSnippets),
264         IsUsingDeclaration(IsUsingDeclaration), NextTokenKind(NextTokenKind) {
265     add(C, SemaCCS);
266     if (C.SemaResult) {
267       assert(ASTCtx);
268       Completion.Origin |= SymbolOrigin::AST;
269       Completion.Name = std::string(llvm::StringRef(SemaCCS->getTypedText()));
270       if (Completion.Scope.empty()) {
271         if ((C.SemaResult->Kind == CodeCompletionResult::RK_Declaration) ||
272             (C.SemaResult->Kind == CodeCompletionResult::RK_Pattern))
273           if (const auto *D = C.SemaResult->getDeclaration())
274             if (const auto *ND = dyn_cast<NamedDecl>(D))
275               Completion.Scope = std::string(
276                   splitQualifiedName(printQualifiedName(*ND)).first);
277       }
278       Completion.Kind = toCompletionItemKind(
279           C.SemaResult->Kind, C.SemaResult->Declaration, ContextKind);
280       // Sema could provide more info on whether the completion was a file or
281       // folder.
282       if (Completion.Kind == CompletionItemKind::File &&
283           Completion.Name.back() == '/')
284         Completion.Kind = CompletionItemKind::Folder;
285       for (const auto &FixIt : C.SemaResult->FixIts) {
286         Completion.FixIts.push_back(toTextEdit(
287             FixIt, ASTCtx->getSourceManager(), ASTCtx->getLangOpts()));
288       }
289       llvm::sort(Completion.FixIts, [](const TextEdit &X, const TextEdit &Y) {
290         return std::tie(X.range.start.line, X.range.start.character) <
291                std::tie(Y.range.start.line, Y.range.start.character);
292       });
293       Completion.Deprecated |=
294           (C.SemaResult->Availability == CXAvailability_Deprecated);
295     }
296     if (C.IndexResult) {
297       Completion.Origin |= C.IndexResult->Origin;
298       if (Completion.Scope.empty())
299         Completion.Scope = std::string(C.IndexResult->Scope);
300       if (Completion.Kind == CompletionItemKind::Missing)
301         Completion.Kind = toCompletionItemKind(C.IndexResult->SymInfo.Kind);
302       if (Completion.Name.empty())
303         Completion.Name = std::string(C.IndexResult->Name);
304       // If the completion was visible to Sema, no qualifier is needed. This
305       // avoids unneeded qualifiers in cases like with `using ns::X`.
306       if (Completion.RequiredQualifier.empty() && !C.SemaResult) {
307         llvm::StringRef ShortestQualifier = C.IndexResult->Scope;
308         for (llvm::StringRef Scope : QueryScopes) {
309           llvm::StringRef Qualifier = C.IndexResult->Scope;
310           if (Qualifier.consume_front(Scope) &&
311               Qualifier.size() < ShortestQualifier.size())
312             ShortestQualifier = Qualifier;
313         }
314         Completion.RequiredQualifier = std::string(ShortestQualifier);
315       }
316       Completion.Deprecated |= (C.IndexResult->Flags & Symbol::Deprecated);
317     }
318     if (C.IdentifierResult) {
319       Completion.Origin |= SymbolOrigin::Identifier;
320       Completion.Kind = CompletionItemKind::Text;
321       Completion.Name = std::string(C.IdentifierResult->Name);
322     }
323 
324     // Turn absolute path into a literal string that can be #included.
325     auto Inserted = [&](llvm::StringRef Header)
326         -> llvm::Expected<std::pair<std::string, bool>> {
327       auto ResolvedDeclaring =
328           URI::resolve(C.IndexResult->CanonicalDeclaration.FileURI, FileName);
329       if (!ResolvedDeclaring)
330         return ResolvedDeclaring.takeError();
331       auto ResolvedInserted = toHeaderFile(Header, FileName);
332       if (!ResolvedInserted)
333         return ResolvedInserted.takeError();
334       auto Spelled = Includes.calculateIncludePath(*ResolvedInserted, FileName);
335       if (!Spelled)
336         return error("Header not on include path");
337       return std::make_pair(
338           std::move(*Spelled),
339           Includes.shouldInsertInclude(*ResolvedDeclaring, *ResolvedInserted));
340     };
341     bool ShouldInsert = C.headerToInsertIfAllowed(Opts).hasValue();
342     // Calculate include paths and edits for all possible headers.
343     for (const auto &Inc : C.RankedIncludeHeaders) {
344       if (auto ToInclude = Inserted(Inc)) {
345         CodeCompletion::IncludeCandidate Include;
346         Include.Header = ToInclude->first;
347         if (ToInclude->second && ShouldInsert)
348           Include.Insertion = Includes.insert(ToInclude->first);
349         Completion.Includes.push_back(std::move(Include));
350       } else
351         log("Failed to generate include insertion edits for adding header "
352             "(FileURI='{0}', IncludeHeader='{1}') into {2}: {3}",
353             C.IndexResult->CanonicalDeclaration.FileURI, Inc, FileName,
354             ToInclude.takeError());
355     }
356     // Prefer includes that do not need edits (i.e. already exist).
357     std::stable_partition(Completion.Includes.begin(),
358                           Completion.Includes.end(),
359                           [](const CodeCompletion::IncludeCandidate &I) {
360                             return !I.Insertion.hasValue();
361                           });
362   }
363 
364   void add(const CompletionCandidate &C, CodeCompletionString *SemaCCS) {
365     assert(bool(C.SemaResult) == bool(SemaCCS));
366     Bundled.emplace_back();
367     BundledEntry &S = Bundled.back();
368     if (C.SemaResult) {
369       bool IsPattern = C.SemaResult->Kind == CodeCompletionResult::RK_Pattern;
370       getSignature(*SemaCCS, &S.Signature, &S.SnippetSuffix,
371                    &Completion.RequiredQualifier, IsPattern);
372       S.ReturnType = getReturnType(*SemaCCS);
373     } else if (C.IndexResult) {
374       S.Signature = std::string(C.IndexResult->Signature);
375       S.SnippetSuffix = std::string(C.IndexResult->CompletionSnippetSuffix);
376       S.ReturnType = std::string(C.IndexResult->ReturnType);
377     }
378     if (ExtractDocumentation && !Completion.Documentation) {
379       auto SetDoc = [&](llvm::StringRef Doc) {
380         if (!Doc.empty()) {
381           Completion.Documentation.emplace();
382           parseDocumentation(Doc, *Completion.Documentation);
383         }
384       };
385       if (C.IndexResult) {
386         SetDoc(C.IndexResult->Documentation);
387       } else if (C.SemaResult) {
388         SetDoc(getDocComment(*ASTCtx, *C.SemaResult,
389                              /*CommentsFromHeader=*/false));
390       }
391     }
392   }
393 
394   CodeCompletion build() {
395     Completion.ReturnType = summarizeReturnType();
396     Completion.Signature = summarizeSignature();
397     Completion.SnippetSuffix = summarizeSnippet();
398     Completion.BundleSize = Bundled.size();
399     return std::move(Completion);
400   }
401 
402 private:
403   struct BundledEntry {
404     std::string SnippetSuffix;
405     std::string Signature;
406     std::string ReturnType;
407   };
408 
409   // If all BundledEntries have the same value for a property, return it.
410   template <std::string BundledEntry::*Member>
411   const std::string *onlyValue() const {
412     auto B = Bundled.begin(), E = Bundled.end();
413     for (auto I = B + 1; I != E; ++I)
414       if (I->*Member != B->*Member)
415         return nullptr;
416     return &(B->*Member);
417   }
418 
419   template <bool BundledEntry::*Member> const bool *onlyValue() const {
420     auto B = Bundled.begin(), E = Bundled.end();
421     for (auto I = B + 1; I != E; ++I)
422       if (I->*Member != B->*Member)
423         return nullptr;
424     return &(B->*Member);
425   }
426 
427   std::string summarizeReturnType() const {
428     if (auto *RT = onlyValue<&BundledEntry::ReturnType>())
429       return *RT;
430     return "";
431   }
432 
433   std::string summarizeSnippet() const {
434     if (IsUsingDeclaration)
435       return "";
436     // Suppress function argument snippets if args are already present.
437     if ((Completion.Kind == CompletionItemKind::Function ||
438          Completion.Kind == CompletionItemKind::Method ||
439          Completion.Kind == CompletionItemKind::Constructor) &&
440         NextTokenKind == tok::l_paren)
441       return "";
442     auto *Snippet = onlyValue<&BundledEntry::SnippetSuffix>();
443     if (!Snippet)
444       // All bundles are function calls.
445       // FIXME(ibiryukov): sometimes add template arguments to a snippet, e.g.
446       // we need to complete 'forward<$1>($0)'.
447       return "($0)";
448     if (EnableFunctionArgSnippets)
449       return *Snippet;
450 
451     // Replace argument snippets with a simplified pattern.
452     if (Snippet->empty())
453       return "";
454     if (Completion.Kind == CompletionItemKind::Function ||
455         Completion.Kind == CompletionItemKind::Method) {
456       // Functions snippets can be of 2 types:
457       // - containing only function arguments, e.g.
458       //   foo(${1:int p1}, ${2:int p2});
459       //   We transform this pattern to '($0)' or '()'.
460       // - template arguments and function arguments, e.g.
461       //   foo<${1:class}>(${2:int p1}).
462       //   We transform this pattern to '<$1>()$0' or '<$0>()'.
463 
464       bool EmptyArgs = llvm::StringRef(*Snippet).endswith("()");
465       if (Snippet->front() == '<')
466         return EmptyArgs ? "<$1>()$0" : "<$1>($0)";
467       if (Snippet->front() == '(')
468         return EmptyArgs ? "()" : "($0)";
469       return *Snippet; // Not an arg snippet?
470     }
471     // 'CompletionItemKind::Interface' matches template type aliases.
472     if (Completion.Kind == CompletionItemKind::Interface ||
473         Completion.Kind == CompletionItemKind::Class) {
474       if (Snippet->front() != '<')
475         return *Snippet; // Not an arg snippet?
476 
477       // Classes and template using aliases can only have template arguments,
478       // e.g. Foo<${1:class}>.
479       if (llvm::StringRef(*Snippet).endswith("<>"))
480         return "<>"; // can happen with defaulted template arguments.
481       return "<$0>";
482     }
483     return *Snippet;
484   }
485 
486   std::string summarizeSignature() const {
487     if (auto *Signature = onlyValue<&BundledEntry::Signature>())
488       return *Signature;
489     // All bundles are function calls.
490     return "(…)";
491   }
492 
493   // ASTCtx can be nullptr if not run with sema.
494   ASTContext *ASTCtx;
495   CodeCompletion Completion;
496   llvm::SmallVector<BundledEntry, 1> Bundled;
497   bool ExtractDocumentation;
498   bool EnableFunctionArgSnippets;
499   // No snippets will be generated for using declarations and when the function
500   // arguments are already present.
501   bool IsUsingDeclaration;
502   tok::TokenKind NextTokenKind;
503 };
504 
505 // Determine the symbol ID for a Sema code completion result, if possible.
506 llvm::Optional<SymbolID> getSymbolID(const CodeCompletionResult &R,
507                                      const SourceManager &SM) {
508   switch (R.Kind) {
509   case CodeCompletionResult::RK_Declaration:
510   case CodeCompletionResult::RK_Pattern: {
511     // Computing USR caches linkage, which may change after code completion.
512     if (hasUnstableLinkage(R.Declaration))
513       return llvm::None;
514     return clang::clangd::getSymbolID(R.Declaration);
515   }
516   case CodeCompletionResult::RK_Macro:
517     return clang::clangd::getSymbolID(R.Macro->getName(), R.MacroDefInfo, SM);
518   case CodeCompletionResult::RK_Keyword:
519     return None;
520   }
521   llvm_unreachable("unknown CodeCompletionResult kind");
522 }
523 
524 // Scopes of the partial identifier we're trying to complete.
525 // It is used when we query the index for more completion results.
526 struct SpecifiedScope {
527   // The scopes we should look in, determined by Sema.
528   //
529   // If the qualifier was fully resolved, we look for completions in these
530   // scopes; if there is an unresolved part of the qualifier, it should be
531   // resolved within these scopes.
532   //
533   // Examples of qualified completion:
534   //
535   //   "::vec"                                      => {""}
536   //   "using namespace std; ::vec^"                => {"", "std::"}
537   //   "namespace ns {using namespace std;} ns::^"  => {"ns::", "std::"}
538   //   "std::vec^"                                  => {""}  // "std" unresolved
539   //
540   // Examples of unqualified completion:
541   //
542   //   "vec^"                                       => {""}
543   //   "using namespace std; vec^"                  => {"", "std::"}
544   //   "using namespace std; namespace ns { vec^ }" => {"ns::", "std::", ""}
545   //
546   // "" for global namespace, "ns::" for normal namespace.
547   std::vector<std::string> AccessibleScopes;
548   // The full scope qualifier as typed by the user (without the leading "::").
549   // Set if the qualifier is not fully resolved by Sema.
550   llvm::Optional<std::string> UnresolvedQualifier;
551 
552   // Construct scopes being queried in indexes. The results are deduplicated.
553   // This method format the scopes to match the index request representation.
554   std::vector<std::string> scopesForIndexQuery() {
555     std::set<std::string> Results;
556     for (llvm::StringRef AS : AccessibleScopes)
557       Results.insert(
558           (AS + (UnresolvedQualifier ? *UnresolvedQualifier : "")).str());
559     return {Results.begin(), Results.end()};
560   }
561 };
562 
563 // Get all scopes that will be queried in indexes and whether symbols from
564 // any scope is allowed. The first scope in the list is the preferred scope
565 // (e.g. enclosing namespace).
566 std::pair<std::vector<std::string>, bool>
567 getQueryScopes(CodeCompletionContext &CCContext, const Sema &CCSema,
568                const CompletionPrefix &HeuristicPrefix,
569                const CodeCompleteOptions &Opts) {
570   SpecifiedScope Scopes;
571   for (auto *Context : CCContext.getVisitedContexts()) {
572     if (isa<TranslationUnitDecl>(Context))
573       Scopes.AccessibleScopes.push_back(""); // global namespace
574     else if (isa<NamespaceDecl>(Context))
575       Scopes.AccessibleScopes.push_back(printNamespaceScope(*Context));
576   }
577 
578   const CXXScopeSpec *SemaSpecifier =
579       CCContext.getCXXScopeSpecifier().getValueOr(nullptr);
580   // Case 1: unqualified completion.
581   if (!SemaSpecifier) {
582     // Case 2 (exception): sema saw no qualifier, but there appears to be one!
583     // This can happen e.g. in incomplete macro expansions. Use heuristics.
584     if (!HeuristicPrefix.Qualifier.empty()) {
585       vlog("Sema said no scope specifier, but we saw {0} in the source code",
586            HeuristicPrefix.Qualifier);
587       StringRef SpelledSpecifier = HeuristicPrefix.Qualifier;
588       if (SpelledSpecifier.consume_front("::"))
589         Scopes.AccessibleScopes = {""};
590       Scopes.UnresolvedQualifier = std::string(SpelledSpecifier);
591       return {Scopes.scopesForIndexQuery(), false};
592     }
593     // The enclosing namespace must be first, it gets a quality boost.
594     std::vector<std::string> EnclosingAtFront;
595     std::string EnclosingScope = printNamespaceScope(*CCSema.CurContext);
596     EnclosingAtFront.push_back(EnclosingScope);
597     for (auto &S : Scopes.scopesForIndexQuery()) {
598       if (EnclosingScope != S)
599         EnclosingAtFront.push_back(std::move(S));
600     }
601     // Allow AllScopes completion as there is no explicit scope qualifier.
602     return {EnclosingAtFront, Opts.AllScopes};
603   }
604   // Case 3: sema saw and resolved a scope qualifier.
605   if (SemaSpecifier && SemaSpecifier->isValid())
606     return {Scopes.scopesForIndexQuery(), false};
607 
608   // Case 4: There was a qualifier, and Sema didn't resolve it.
609   Scopes.AccessibleScopes.push_back(""); // Make sure global scope is included.
610   llvm::StringRef SpelledSpecifier = Lexer::getSourceText(
611       CharSourceRange::getCharRange(SemaSpecifier->getRange()),
612       CCSema.SourceMgr, clang::LangOptions());
613   if (SpelledSpecifier.consume_front("::"))
614     Scopes.AccessibleScopes = {""};
615   Scopes.UnresolvedQualifier = std::string(SpelledSpecifier);
616   // Sema excludes the trailing "::".
617   if (!Scopes.UnresolvedQualifier->empty())
618     *Scopes.UnresolvedQualifier += "::";
619 
620   return {Scopes.scopesForIndexQuery(), false};
621 }
622 
623 // Should we perform index-based completion in a context of the specified kind?
624 // FIXME: consider allowing completion, but restricting the result types.
625 bool contextAllowsIndex(enum CodeCompletionContext::Kind K) {
626   switch (K) {
627   case CodeCompletionContext::CCC_TopLevel:
628   case CodeCompletionContext::CCC_ObjCInterface:
629   case CodeCompletionContext::CCC_ObjCImplementation:
630   case CodeCompletionContext::CCC_ObjCIvarList:
631   case CodeCompletionContext::CCC_ClassStructUnion:
632   case CodeCompletionContext::CCC_Statement:
633   case CodeCompletionContext::CCC_Expression:
634   case CodeCompletionContext::CCC_ObjCMessageReceiver:
635   case CodeCompletionContext::CCC_EnumTag:
636   case CodeCompletionContext::CCC_UnionTag:
637   case CodeCompletionContext::CCC_ClassOrStructTag:
638   case CodeCompletionContext::CCC_ObjCProtocolName:
639   case CodeCompletionContext::CCC_Namespace:
640   case CodeCompletionContext::CCC_Type:
641   case CodeCompletionContext::CCC_ParenthesizedExpression:
642   case CodeCompletionContext::CCC_ObjCInterfaceName:
643   case CodeCompletionContext::CCC_ObjCCategoryName:
644   case CodeCompletionContext::CCC_Symbol:
645   case CodeCompletionContext::CCC_SymbolOrNewName:
646     return true;
647   case CodeCompletionContext::CCC_OtherWithMacros:
648   case CodeCompletionContext::CCC_DotMemberAccess:
649   case CodeCompletionContext::CCC_ArrowMemberAccess:
650   case CodeCompletionContext::CCC_ObjCPropertyAccess:
651   case CodeCompletionContext::CCC_MacroName:
652   case CodeCompletionContext::CCC_MacroNameUse:
653   case CodeCompletionContext::CCC_PreprocessorExpression:
654   case CodeCompletionContext::CCC_PreprocessorDirective:
655   case CodeCompletionContext::CCC_SelectorName:
656   case CodeCompletionContext::CCC_TypeQualifiers:
657   case CodeCompletionContext::CCC_ObjCInstanceMessage:
658   case CodeCompletionContext::CCC_ObjCClassMessage:
659   case CodeCompletionContext::CCC_IncludedFile:
660   // FIXME: Provide identifier based completions for the following contexts:
661   case CodeCompletionContext::CCC_Other: // Be conservative.
662   case CodeCompletionContext::CCC_NaturalLanguage:
663   case CodeCompletionContext::CCC_Recovery:
664   case CodeCompletionContext::CCC_NewName:
665     return false;
666   }
667   llvm_unreachable("unknown code completion context");
668 }
669 
670 static bool isInjectedClass(const NamedDecl &D) {
671   if (auto *R = dyn_cast_or_null<RecordDecl>(&D))
672     if (R->isInjectedClassName())
673       return true;
674   return false;
675 }
676 
677 // Some member calls are excluded because they're so rarely useful.
678 static bool isExcludedMember(const NamedDecl &D) {
679   // Destructor completion is rarely useful, and works inconsistently.
680   // (s.^ completes ~string, but s.~st^ is an error).
681   if (D.getKind() == Decl::CXXDestructor)
682     return true;
683   // Injected name may be useful for A::foo(), but who writes A::A::foo()?
684   if (isInjectedClass(D))
685     return true;
686   // Explicit calls to operators are also rare.
687   auto NameKind = D.getDeclName().getNameKind();
688   if (NameKind == DeclarationName::CXXOperatorName ||
689       NameKind == DeclarationName::CXXLiteralOperatorName ||
690       NameKind == DeclarationName::CXXConversionFunctionName)
691     return true;
692   return false;
693 }
694 
695 // The CompletionRecorder captures Sema code-complete output, including context.
696 // It filters out ignored results (but doesn't apply fuzzy-filtering yet).
697 // It doesn't do scoring or conversion to CompletionItem yet, as we want to
698 // merge with index results first.
699 // Generally the fields and methods of this object should only be used from
700 // within the callback.
701 struct CompletionRecorder : public CodeCompleteConsumer {
702   CompletionRecorder(const CodeCompleteOptions &Opts,
703                      llvm::unique_function<void()> ResultsCallback)
704       : CodeCompleteConsumer(Opts.getClangCompleteOpts()),
705         CCContext(CodeCompletionContext::CCC_Other), Opts(Opts),
706         CCAllocator(std::make_shared<GlobalCodeCompletionAllocator>()),
707         CCTUInfo(CCAllocator), ResultsCallback(std::move(ResultsCallback)) {
708     assert(this->ResultsCallback);
709   }
710 
711   std::vector<CodeCompletionResult> Results;
712   CodeCompletionContext CCContext;
713   Sema *CCSema = nullptr; // Sema that created the results.
714   // FIXME: Sema is scary. Can we store ASTContext and Preprocessor, instead?
715 
716   void ProcessCodeCompleteResults(class Sema &S, CodeCompletionContext Context,
717                                   CodeCompletionResult *InResults,
718                                   unsigned NumResults) override final {
719     // Results from recovery mode are generally useless, and the callback after
720     // recovery (if any) is usually more interesting. To make sure we handle the
721     // future callback from sema, we just ignore all callbacks in recovery mode,
722     // as taking only results from recovery mode results in poor completion
723     // results.
724     // FIXME: in case there is no future sema completion callback after the
725     // recovery mode, we might still want to provide some results (e.g. trivial
726     // identifier-based completion).
727     if (Context.getKind() == CodeCompletionContext::CCC_Recovery) {
728       log("Code complete: Ignoring sema code complete callback with Recovery "
729           "context.");
730       return;
731     }
732     // If a callback is called without any sema result and the context does not
733     // support index-based completion, we simply skip it to give way to
734     // potential future callbacks with results.
735     if (NumResults == 0 && !contextAllowsIndex(Context.getKind()))
736       return;
737     if (CCSema) {
738       log("Multiple code complete callbacks (parser backtracked?). "
739           "Dropping results from context {0}, keeping results from {1}.",
740           getCompletionKindString(Context.getKind()),
741           getCompletionKindString(this->CCContext.getKind()));
742       return;
743     }
744     // Record the completion context.
745     CCSema = &S;
746     CCContext = Context;
747 
748     // Retain the results we might want.
749     for (unsigned I = 0; I < NumResults; ++I) {
750       auto &Result = InResults[I];
751       // Class members that are shadowed by subclasses are usually noise.
752       if (Result.Hidden && Result.Declaration &&
753           Result.Declaration->isCXXClassMember())
754         continue;
755       if (!Opts.IncludeIneligibleResults &&
756           (Result.Availability == CXAvailability_NotAvailable ||
757            Result.Availability == CXAvailability_NotAccessible))
758         continue;
759       if (Result.Declaration &&
760           !Context.getBaseType().isNull() // is this a member-access context?
761           && isExcludedMember(*Result.Declaration))
762         continue;
763       // Skip injected class name when no class scope is not explicitly set.
764       // E.g. show injected A::A in `using A::A^` but not in "A^".
765       if (Result.Declaration && !Context.getCXXScopeSpecifier().hasValue() &&
766           isInjectedClass(*Result.Declaration))
767         continue;
768       // We choose to never append '::' to completion results in clangd.
769       Result.StartsNestedNameSpecifier = false;
770       Results.push_back(Result);
771     }
772     ResultsCallback();
773   }
774 
775   CodeCompletionAllocator &getAllocator() override { return *CCAllocator; }
776   CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; }
777 
778   // Returns the filtering/sorting name for Result, which must be from Results.
779   // Returned string is owned by this recorder (or the AST).
780   llvm::StringRef getName(const CodeCompletionResult &Result) {
781     switch (Result.Kind) {
782     case CodeCompletionResult::RK_Declaration:
783       if (auto *ID = Result.Declaration->getIdentifier())
784         return ID->getName();
785       break;
786     case CodeCompletionResult::RK_Keyword:
787       return Result.Keyword;
788     case CodeCompletionResult::RK_Macro:
789       return Result.Macro->getName();
790     case CodeCompletionResult::RK_Pattern:
791       return Result.Pattern->getTypedText();
792     }
793     auto *CCS = codeCompletionString(Result);
794     return CCS->getTypedText();
795   }
796 
797   // Build a CodeCompletion string for R, which must be from Results.
798   // The CCS will be owned by this recorder.
799   CodeCompletionString *codeCompletionString(const CodeCompletionResult &R) {
800     // CodeCompletionResult doesn't seem to be const-correct. We own it, anyway.
801     return const_cast<CodeCompletionResult &>(R).CreateCodeCompletionString(
802         *CCSema, CCContext, *CCAllocator, CCTUInfo,
803         /*IncludeBriefComments=*/false);
804   }
805 
806 private:
807   CodeCompleteOptions Opts;
808   std::shared_ptr<GlobalCodeCompletionAllocator> CCAllocator;
809   CodeCompletionTUInfo CCTUInfo;
810   llvm::unique_function<void()> ResultsCallback;
811 };
812 
813 struct ScoredSignature {
814   // When set, requires documentation to be requested from the index with this
815   // ID.
816   llvm::Optional<SymbolID> IDForDoc;
817   SignatureInformation Signature;
818   SignatureQualitySignals Quality;
819 };
820 
821 class SignatureHelpCollector final : public CodeCompleteConsumer {
822 public:
823   SignatureHelpCollector(const clang::CodeCompleteOptions &CodeCompleteOpts,
824                          const SymbolIndex *Index, SignatureHelp &SigHelp)
825       : CodeCompleteConsumer(CodeCompleteOpts), SigHelp(SigHelp),
826         Allocator(std::make_shared<clang::GlobalCodeCompletionAllocator>()),
827         CCTUInfo(Allocator), Index(Index) {}
828 
829   void ProcessOverloadCandidates(Sema &S, unsigned CurrentArg,
830                                  OverloadCandidate *Candidates,
831                                  unsigned NumCandidates,
832                                  SourceLocation OpenParLoc) override {
833     assert(!OpenParLoc.isInvalid());
834     SourceManager &SrcMgr = S.getSourceManager();
835     OpenParLoc = SrcMgr.getFileLoc(OpenParLoc);
836     if (SrcMgr.isInMainFile(OpenParLoc))
837       SigHelp.argListStart = sourceLocToPosition(SrcMgr, OpenParLoc);
838     else
839       elog("Location oustide main file in signature help: {0}",
840            OpenParLoc.printToString(SrcMgr));
841 
842     std::vector<ScoredSignature> ScoredSignatures;
843     SigHelp.signatures.reserve(NumCandidates);
844     ScoredSignatures.reserve(NumCandidates);
845     // FIXME(rwols): How can we determine the "active overload candidate"?
846     // Right now the overloaded candidates seem to be provided in a "best fit"
847     // order, so I'm not too worried about this.
848     SigHelp.activeSignature = 0;
849     assert(CurrentArg <= (unsigned)std::numeric_limits<int>::max() &&
850            "too many arguments");
851     SigHelp.activeParameter = static_cast<int>(CurrentArg);
852     for (unsigned I = 0; I < NumCandidates; ++I) {
853       OverloadCandidate Candidate = Candidates[I];
854       // We want to avoid showing instantiated signatures, because they may be
855       // long in some cases (e.g. when 'T' is substituted with 'std::string', we
856       // would get 'std::basic_string<char>').
857       if (auto *Func = Candidate.getFunction()) {
858         if (auto *Pattern = Func->getTemplateInstantiationPattern())
859           Candidate = OverloadCandidate(Pattern);
860       }
861 
862       const auto *CCS = Candidate.CreateSignatureString(
863           CurrentArg, S, *Allocator, CCTUInfo, true);
864       assert(CCS && "Expected the CodeCompletionString to be non-null");
865       ScoredSignatures.push_back(processOverloadCandidate(
866           Candidate, *CCS,
867           Candidate.getFunction()
868               ? getDeclComment(S.getASTContext(), *Candidate.getFunction())
869               : ""));
870     }
871 
872     // Sema does not load the docs from the preamble, so we need to fetch extra
873     // docs from the index instead.
874     llvm::DenseMap<SymbolID, std::string> FetchedDocs;
875     if (Index) {
876       LookupRequest IndexRequest;
877       for (const auto &S : ScoredSignatures) {
878         if (!S.IDForDoc)
879           continue;
880         IndexRequest.IDs.insert(*S.IDForDoc);
881       }
882       Index->lookup(IndexRequest, [&](const Symbol &S) {
883         if (!S.Documentation.empty())
884           FetchedDocs[S.ID] = std::string(S.Documentation);
885       });
886       log("SigHelp: requested docs for {0} symbols from the index, got {1} "
887           "symbols with non-empty docs in the response",
888           IndexRequest.IDs.size(), FetchedDocs.size());
889     }
890 
891     llvm::sort(ScoredSignatures, [](const ScoredSignature &L,
892                                     const ScoredSignature &R) {
893       // Ordering follows:
894       // - Less number of parameters is better.
895       // - Function is better than FunctionType which is better than
896       // Function Template.
897       // - High score is better.
898       // - Shorter signature is better.
899       // - Alphabetically smaller is better.
900       if (L.Quality.NumberOfParameters != R.Quality.NumberOfParameters)
901         return L.Quality.NumberOfParameters < R.Quality.NumberOfParameters;
902       if (L.Quality.NumberOfOptionalParameters !=
903           R.Quality.NumberOfOptionalParameters)
904         return L.Quality.NumberOfOptionalParameters <
905                R.Quality.NumberOfOptionalParameters;
906       if (L.Quality.Kind != R.Quality.Kind) {
907         using OC = CodeCompleteConsumer::OverloadCandidate;
908         switch (L.Quality.Kind) {
909         case OC::CK_Function:
910           return true;
911         case OC::CK_FunctionType:
912           return R.Quality.Kind != OC::CK_Function;
913         case OC::CK_FunctionTemplate:
914           return false;
915         }
916         llvm_unreachable("Unknown overload candidate type.");
917       }
918       if (L.Signature.label.size() != R.Signature.label.size())
919         return L.Signature.label.size() < R.Signature.label.size();
920       return L.Signature.label < R.Signature.label;
921     });
922 
923     for (auto &SS : ScoredSignatures) {
924       auto IndexDocIt =
925           SS.IDForDoc ? FetchedDocs.find(*SS.IDForDoc) : FetchedDocs.end();
926       if (IndexDocIt != FetchedDocs.end())
927         SS.Signature.documentation = IndexDocIt->second;
928 
929       SigHelp.signatures.push_back(std::move(SS.Signature));
930     }
931   }
932 
933   GlobalCodeCompletionAllocator &getAllocator() override { return *Allocator; }
934 
935   CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; }
936 
937 private:
938   void processParameterChunk(llvm::StringRef ChunkText,
939                              SignatureInformation &Signature) const {
940     // (!) this is O(n), should still be fast compared to building ASTs.
941     unsigned ParamStartOffset = lspLength(Signature.label);
942     unsigned ParamEndOffset = ParamStartOffset + lspLength(ChunkText);
943     // A piece of text that describes the parameter that corresponds to
944     // the code-completion location within a function call, message send,
945     // macro invocation, etc.
946     Signature.label += ChunkText;
947     ParameterInformation Info;
948     Info.labelOffsets.emplace(ParamStartOffset, ParamEndOffset);
949     // FIXME: only set 'labelOffsets' when all clients migrate out of it.
950     Info.labelString = std::string(ChunkText);
951 
952     Signature.parameters.push_back(std::move(Info));
953   }
954 
955   void processOptionalChunk(const CodeCompletionString &CCS,
956                             SignatureInformation &Signature,
957                             SignatureQualitySignals &Signal) const {
958     for (const auto &Chunk : CCS) {
959       switch (Chunk.Kind) {
960       case CodeCompletionString::CK_Optional:
961         assert(Chunk.Optional &&
962                "Expected the optional code completion string to be non-null.");
963         processOptionalChunk(*Chunk.Optional, Signature, Signal);
964         break;
965       case CodeCompletionString::CK_VerticalSpace:
966         break;
967       case CodeCompletionString::CK_CurrentParameter:
968       case CodeCompletionString::CK_Placeholder:
969         processParameterChunk(Chunk.Text, Signature);
970         Signal.NumberOfOptionalParameters++;
971         break;
972       default:
973         Signature.label += Chunk.Text;
974         break;
975       }
976     }
977   }
978 
979   // FIXME(ioeric): consider moving CodeCompletionString logic here to
980   // CompletionString.h.
981   ScoredSignature processOverloadCandidate(const OverloadCandidate &Candidate,
982                                            const CodeCompletionString &CCS,
983                                            llvm::StringRef DocComment) const {
984     SignatureInformation Signature;
985     SignatureQualitySignals Signal;
986     const char *ReturnType = nullptr;
987 
988     Signature.documentation = formatDocumentation(CCS, DocComment);
989     Signal.Kind = Candidate.getKind();
990 
991     for (const auto &Chunk : CCS) {
992       switch (Chunk.Kind) {
993       case CodeCompletionString::CK_ResultType:
994         // A piece of text that describes the type of an entity or,
995         // for functions and methods, the return type.
996         assert(!ReturnType && "Unexpected CK_ResultType");
997         ReturnType = Chunk.Text;
998         break;
999       case CodeCompletionString::CK_CurrentParameter:
1000       case CodeCompletionString::CK_Placeholder:
1001         processParameterChunk(Chunk.Text, Signature);
1002         Signal.NumberOfParameters++;
1003         break;
1004       case CodeCompletionString::CK_Optional: {
1005         // The rest of the parameters are defaulted/optional.
1006         assert(Chunk.Optional &&
1007                "Expected the optional code completion string to be non-null.");
1008         processOptionalChunk(*Chunk.Optional, Signature, Signal);
1009         break;
1010       }
1011       case CodeCompletionString::CK_VerticalSpace:
1012         break;
1013       default:
1014         Signature.label += Chunk.Text;
1015         break;
1016       }
1017     }
1018     if (ReturnType) {
1019       Signature.label += " -> ";
1020       Signature.label += ReturnType;
1021     }
1022     dlog("Signal for {0}: {1}", Signature, Signal);
1023     ScoredSignature Result;
1024     Result.Signature = std::move(Signature);
1025     Result.Quality = Signal;
1026     const FunctionDecl *Func = Candidate.getFunction();
1027     if (Func && Result.Signature.documentation.empty()) {
1028       // Computing USR caches linkage, which may change after code completion.
1029       if (!hasUnstableLinkage(Func))
1030         Result.IDForDoc = clangd::getSymbolID(Func);
1031     }
1032     return Result;
1033   }
1034 
1035   SignatureHelp &SigHelp;
1036   std::shared_ptr<clang::GlobalCodeCompletionAllocator> Allocator;
1037   CodeCompletionTUInfo CCTUInfo;
1038   const SymbolIndex *Index;
1039 }; // SignatureHelpCollector
1040 
1041 struct SemaCompleteInput {
1042   PathRef FileName;
1043   size_t Offset;
1044   const PreambleData &Preamble;
1045   const llvm::Optional<PreamblePatch> Patch;
1046   const ParseInputs &ParseInput;
1047 };
1048 
1049 void loadMainFilePreambleMacros(const Preprocessor &PP,
1050                                 const PreambleData &Preamble) {
1051   // The ExternalPreprocessorSource has our macros, if we know where to look.
1052   // We can read all the macros using PreambleMacros->ReadDefinedMacros(),
1053   // but this includes transitively included files, so may deserialize a lot.
1054   ExternalPreprocessorSource *PreambleMacros = PP.getExternalSource();
1055   // As we have the names of the macros, we can look up their IdentifierInfo
1056   // and then use this to load just the macros we want.
1057   IdentifierInfoLookup *PreambleIdentifiers =
1058       PP.getIdentifierTable().getExternalIdentifierLookup();
1059   if (!PreambleIdentifiers || !PreambleMacros)
1060     return;
1061   for (const auto &MacroName : Preamble.Macros.Names)
1062     if (auto *II = PreambleIdentifiers->get(MacroName.getKey()))
1063       if (II->isOutOfDate())
1064         PreambleMacros->updateOutOfDateIdentifier(*II);
1065 }
1066 
1067 // Invokes Sema code completion on a file.
1068 // If \p Includes is set, it will be updated based on the compiler invocation.
1069 bool semaCodeComplete(std::unique_ptr<CodeCompleteConsumer> Consumer,
1070                       const clang::CodeCompleteOptions &Options,
1071                       const SemaCompleteInput &Input,
1072                       IncludeStructure *Includes = nullptr) {
1073   trace::Span Tracer("Sema completion");
1074 
1075   IgnoreDiagnostics IgnoreDiags;
1076   auto CI = buildCompilerInvocation(Input.ParseInput, IgnoreDiags);
1077   if (!CI) {
1078     elog("Couldn't create CompilerInvocation");
1079     return false;
1080   }
1081   auto &FrontendOpts = CI->getFrontendOpts();
1082   FrontendOpts.SkipFunctionBodies = true;
1083   // Disable typo correction in Sema.
1084   CI->getLangOpts()->SpellChecking = false;
1085   // Code completion won't trigger in delayed template bodies.
1086   // This is on-by-default in windows to allow parsing SDK headers; we're only
1087   // disabling it for the main-file (not preamble).
1088   CI->getLangOpts()->DelayedTemplateParsing = false;
1089   // Setup code completion.
1090   FrontendOpts.CodeCompleteOpts = Options;
1091   FrontendOpts.CodeCompletionAt.FileName = std::string(Input.FileName);
1092   std::tie(FrontendOpts.CodeCompletionAt.Line,
1093            FrontendOpts.CodeCompletionAt.Column) =
1094       offsetToClangLineColumn(Input.ParseInput.Contents, Input.Offset);
1095 
1096   std::unique_ptr<llvm::MemoryBuffer> ContentsBuffer =
1097       llvm::MemoryBuffer::getMemBufferCopy(Input.ParseInput.Contents,
1098                                            Input.FileName);
1099   // The diagnostic options must be set before creating a CompilerInstance.
1100   CI->getDiagnosticOpts().IgnoreWarnings = true;
1101   // We reuse the preamble whether it's valid or not. This is a
1102   // correctness/performance tradeoff: building without a preamble is slow, and
1103   // completion is latency-sensitive.
1104   // However, if we're completing *inside* the preamble section of the draft,
1105   // overriding the preamble will break sema completion. Fortunately we can just
1106   // skip all includes in this case; these completions are really simple.
1107   PreambleBounds PreambleRegion =
1108       ComputePreambleBounds(*CI->getLangOpts(), ContentsBuffer.get(), 0);
1109   bool CompletingInPreamble = PreambleRegion.Size > Input.Offset;
1110   if (Input.Patch)
1111     Input.Patch->apply(*CI);
1112   // NOTE: we must call BeginSourceFile after prepareCompilerInstance. Otherwise
1113   // the remapped buffers do not get freed.
1114   llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS =
1115       Input.ParseInput.TFS->view(Input.ParseInput.CompileCommand.Directory);
1116   if (Input.Preamble.StatCache)
1117     VFS = Input.Preamble.StatCache->getConsumingFS(std::move(VFS));
1118   auto Clang = prepareCompilerInstance(
1119       std::move(CI), !CompletingInPreamble ? &Input.Preamble.Preamble : nullptr,
1120       std::move(ContentsBuffer), std::move(VFS), IgnoreDiags);
1121   Clang->getPreprocessorOpts().SingleFileParseMode = CompletingInPreamble;
1122   Clang->setCodeCompletionConsumer(Consumer.release());
1123 
1124   SyntaxOnlyAction Action;
1125   if (!Action.BeginSourceFile(*Clang, Clang->getFrontendOpts().Inputs[0])) {
1126     log("BeginSourceFile() failed when running codeComplete for {0}",
1127         Input.FileName);
1128     return false;
1129   }
1130   // Macros can be defined within the preamble region of the main file.
1131   // They don't fall nicely into our index/Sema dichotomy:
1132   //  - they're not indexed for completion (they're not available across files)
1133   //  - but Sema code complete won't see them: as part of the preamble, they're
1134   //    deserialized only when mentioned.
1135   // Force them to be deserialized so SemaCodeComplete sees them.
1136   loadMainFilePreambleMacros(Clang->getPreprocessor(), Input.Preamble);
1137   if (Includes)
1138     Clang->getPreprocessor().addPPCallbacks(
1139         collectIncludeStructureCallback(Clang->getSourceManager(), Includes));
1140   if (llvm::Error Err = Action.Execute()) {
1141     log("Execute() failed when running codeComplete for {0}: {1}",
1142         Input.FileName, toString(std::move(Err)));
1143     return false;
1144   }
1145   Action.EndSourceFile();
1146 
1147   return true;
1148 }
1149 
1150 // Should we allow index completions in the specified context?
1151 bool allowIndex(CodeCompletionContext &CC) {
1152   if (!contextAllowsIndex(CC.getKind()))
1153     return false;
1154   // We also avoid ClassName::bar (but allow namespace::bar).
1155   auto Scope = CC.getCXXScopeSpecifier();
1156   if (!Scope)
1157     return true;
1158   NestedNameSpecifier *NameSpec = (*Scope)->getScopeRep();
1159   if (!NameSpec)
1160     return true;
1161   // We only query the index when qualifier is a namespace.
1162   // If it's a class, we rely solely on sema completions.
1163   switch (NameSpec->getKind()) {
1164   case NestedNameSpecifier::Global:
1165   case NestedNameSpecifier::Namespace:
1166   case NestedNameSpecifier::NamespaceAlias:
1167     return true;
1168   case NestedNameSpecifier::Super:
1169   case NestedNameSpecifier::TypeSpec:
1170   case NestedNameSpecifier::TypeSpecWithTemplate:
1171   // Unresolved inside a template.
1172   case NestedNameSpecifier::Identifier:
1173     return false;
1174   }
1175   llvm_unreachable("invalid NestedNameSpecifier kind");
1176 }
1177 
1178 std::future<SymbolSlab> startAsyncFuzzyFind(const SymbolIndex &Index,
1179                                             const FuzzyFindRequest &Req) {
1180   return runAsync<SymbolSlab>([&Index, Req]() {
1181     trace::Span Tracer("Async fuzzyFind");
1182     SymbolSlab::Builder Syms;
1183     Index.fuzzyFind(Req, [&Syms](const Symbol &Sym) { Syms.insert(Sym); });
1184     return std::move(Syms).build();
1185   });
1186 }
1187 
1188 // Creates a `FuzzyFindRequest` based on the cached index request from the
1189 // last completion, if any, and the speculated completion filter text in the
1190 // source code.
1191 FuzzyFindRequest speculativeFuzzyFindRequestForCompletion(
1192     FuzzyFindRequest CachedReq, const CompletionPrefix &HeuristicPrefix) {
1193   CachedReq.Query = std::string(HeuristicPrefix.Name);
1194   return CachedReq;
1195 }
1196 
1197 // Runs Sema-based (AST) and Index-based completion, returns merged results.
1198 //
1199 // There are a few tricky considerations:
1200 //   - the AST provides information needed for the index query (e.g. which
1201 //     namespaces to search in). So Sema must start first.
1202 //   - we only want to return the top results (Opts.Limit).
1203 //     Building CompletionItems for everything else is wasteful, so we want to
1204 //     preserve the "native" format until we're done with scoring.
1205 //   - the data underlying Sema completion items is owned by the AST and various
1206 //     other arenas, which must stay alive for us to build CompletionItems.
1207 //   - we may get duplicate results from Sema and the Index, we need to merge.
1208 //
1209 // So we start Sema completion first, and do all our work in its callback.
1210 // We use the Sema context information to query the index.
1211 // Then we merge the two result sets, producing items that are Sema/Index/Both.
1212 // These items are scored, and the top N are synthesized into the LSP response.
1213 // Finally, we can clean up the data structures created by Sema completion.
1214 //
1215 // Main collaborators are:
1216 //   - semaCodeComplete sets up the compiler machinery to run code completion.
1217 //   - CompletionRecorder captures Sema completion results, including context.
1218 //   - SymbolIndex (Opts.Index) provides index completion results as Symbols
1219 //   - CompletionCandidates are the result of merging Sema and Index results.
1220 //     Each candidate points to an underlying CodeCompletionResult (Sema), a
1221 //     Symbol (Index), or both. It computes the result quality score.
1222 //     CompletionCandidate also does conversion to CompletionItem (at the end).
1223 //   - FuzzyMatcher scores how the candidate matches the partial identifier.
1224 //     This score is combined with the result quality score for the final score.
1225 //   - TopN determines the results with the best score.
1226 class CodeCompleteFlow {
1227   PathRef FileName;
1228   IncludeStructure Includes;           // Complete once the compiler runs.
1229   SpeculativeFuzzyFind *SpecFuzzyFind; // Can be nullptr.
1230   const CodeCompleteOptions &Opts;
1231 
1232   // Sema takes ownership of Recorder. Recorder is valid until Sema cleanup.
1233   CompletionRecorder *Recorder = nullptr;
1234   CodeCompletionContext::Kind CCContextKind = CodeCompletionContext::CCC_Other;
1235   bool IsUsingDeclaration = false;
1236   // The snippets will not be generated if the token following completion
1237   // location is an opening parenthesis (tok::l_paren) because this would add
1238   // extra parenthesis.
1239   tok::TokenKind NextTokenKind = tok::eof;
1240   // Counters for logging.
1241   int NSema = 0, NIndex = 0, NSemaAndIndex = 0, NIdent = 0;
1242   bool Incomplete = false; // Would more be available with a higher limit?
1243   CompletionPrefix HeuristicPrefix;
1244   llvm::Optional<FuzzyMatcher> Filter; // Initialized once Sema runs.
1245   Range ReplacedRange;
1246   std::vector<std::string> QueryScopes; // Initialized once Sema runs.
1247   // Initialized once QueryScopes is initialized, if there are scopes.
1248   llvm::Optional<ScopeDistance> ScopeProximity;
1249   llvm::Optional<OpaqueType> PreferredType; // Initialized once Sema runs.
1250   // Whether to query symbols from any scope. Initialized once Sema runs.
1251   bool AllScopes = false;
1252   llvm::StringSet<> ContextWords;
1253   // Include-insertion and proximity scoring rely on the include structure.
1254   // This is available after Sema has run.
1255   llvm::Optional<IncludeInserter> Inserter;  // Available during runWithSema.
1256   llvm::Optional<URIDistance> FileProximity; // Initialized once Sema runs.
1257   /// Speculative request based on the cached request and the filter text before
1258   /// the cursor.
1259   /// Initialized right before sema run. This is only set if `SpecFuzzyFind` is
1260   /// set and contains a cached request.
1261   llvm::Optional<FuzzyFindRequest> SpecReq;
1262 
1263 public:
1264   // A CodeCompleteFlow object is only useful for calling run() exactly once.
1265   CodeCompleteFlow(PathRef FileName, const IncludeStructure &Includes,
1266                    SpeculativeFuzzyFind *SpecFuzzyFind,
1267                    const CodeCompleteOptions &Opts)
1268       : FileName(FileName), Includes(Includes), SpecFuzzyFind(SpecFuzzyFind),
1269         Opts(Opts) {}
1270 
1271   CodeCompleteResult run(const SemaCompleteInput &SemaCCInput) && {
1272     trace::Span Tracer("CodeCompleteFlow");
1273     HeuristicPrefix = guessCompletionPrefix(SemaCCInput.ParseInput.Contents,
1274                                             SemaCCInput.Offset);
1275     populateContextWords(SemaCCInput.ParseInput.Contents);
1276     if (Opts.Index && SpecFuzzyFind && SpecFuzzyFind->CachedReq.hasValue()) {
1277       assert(!SpecFuzzyFind->Result.valid());
1278       SpecReq = speculativeFuzzyFindRequestForCompletion(
1279           *SpecFuzzyFind->CachedReq, HeuristicPrefix);
1280       SpecFuzzyFind->Result = startAsyncFuzzyFind(*Opts.Index, *SpecReq);
1281     }
1282 
1283     // We run Sema code completion first. It builds an AST and calculates:
1284     //   - completion results based on the AST.
1285     //   - partial identifier and context. We need these for the index query.
1286     CodeCompleteResult Output;
1287     auto RecorderOwner = std::make_unique<CompletionRecorder>(Opts, [&]() {
1288       assert(Recorder && "Recorder is not set");
1289       CCContextKind = Recorder->CCContext.getKind();
1290       IsUsingDeclaration = Recorder->CCContext.isUsingDeclaration();
1291       auto Style = getFormatStyleForFile(SemaCCInput.FileName,
1292                                          SemaCCInput.ParseInput.Contents,
1293                                          *SemaCCInput.ParseInput.TFS);
1294       const auto NextToken = Lexer::findNextToken(
1295           Recorder->CCSema->getPreprocessor().getCodeCompletionLoc(),
1296           Recorder->CCSema->getSourceManager(), Recorder->CCSema->LangOpts);
1297       if (NextToken)
1298         NextTokenKind = NextToken->getKind();
1299       // If preprocessor was run, inclusions from preprocessor callback should
1300       // already be added to Includes.
1301       Inserter.emplace(
1302           SemaCCInput.FileName, SemaCCInput.ParseInput.Contents, Style,
1303           SemaCCInput.ParseInput.CompileCommand.Directory,
1304           &Recorder->CCSema->getPreprocessor().getHeaderSearchInfo());
1305       for (const auto &Inc : Includes.MainFileIncludes)
1306         Inserter->addExisting(Inc);
1307 
1308       // Most of the cost of file proximity is in initializing the FileDistance
1309       // structures based on the observed includes, once per query. Conceptually
1310       // that happens here (though the per-URI-scheme initialization is lazy).
1311       // The per-result proximity scoring is (amortized) very cheap.
1312       FileDistanceOptions ProxOpts{}; // Use defaults.
1313       const auto &SM = Recorder->CCSema->getSourceManager();
1314       llvm::StringMap<SourceParams> ProxSources;
1315       for (auto &Entry : Includes.includeDepth(
1316                SM.getFileEntryForID(SM.getMainFileID())->getName())) {
1317         auto &Source = ProxSources[Entry.getKey()];
1318         Source.Cost = Entry.getValue() * ProxOpts.IncludeCost;
1319         // Symbols near our transitive includes are good, but only consider
1320         // things in the same directory or below it. Otherwise there can be
1321         // many false positives.
1322         if (Entry.getValue() > 0)
1323           Source.MaxUpTraversals = 1;
1324       }
1325       FileProximity.emplace(ProxSources, ProxOpts);
1326 
1327       Output = runWithSema();
1328       Inserter.reset(); // Make sure this doesn't out-live Clang.
1329       SPAN_ATTACH(Tracer, "sema_completion_kind",
1330                   getCompletionKindString(CCContextKind));
1331       log("Code complete: sema context {0}, query scopes [{1}] (AnyScope={2}), "
1332           "expected type {3}{4}",
1333           getCompletionKindString(CCContextKind),
1334           llvm::join(QueryScopes.begin(), QueryScopes.end(), ","), AllScopes,
1335           PreferredType ? Recorder->CCContext.getPreferredType().getAsString()
1336                         : "<none>",
1337           IsUsingDeclaration ? ", inside using declaration" : "");
1338     });
1339 
1340     Recorder = RecorderOwner.get();
1341 
1342     semaCodeComplete(std::move(RecorderOwner), Opts.getClangCompleteOpts(),
1343                      SemaCCInput, &Includes);
1344     logResults(Output, Tracer);
1345     return Output;
1346   }
1347 
1348   void logResults(const CodeCompleteResult &Output, const trace::Span &Tracer) {
1349     SPAN_ATTACH(Tracer, "sema_results", NSema);
1350     SPAN_ATTACH(Tracer, "index_results", NIndex);
1351     SPAN_ATTACH(Tracer, "merged_results", NSemaAndIndex);
1352     SPAN_ATTACH(Tracer, "identifier_results", NIdent);
1353     SPAN_ATTACH(Tracer, "returned_results", int64_t(Output.Completions.size()));
1354     SPAN_ATTACH(Tracer, "incomplete", Output.HasMore);
1355     log("Code complete: {0} results from Sema, {1} from Index, "
1356         "{2} matched, {3} from identifiers, {4} returned{5}.",
1357         NSema, NIndex, NSemaAndIndex, NIdent, Output.Completions.size(),
1358         Output.HasMore ? " (incomplete)" : "");
1359     assert(!Opts.Limit || Output.Completions.size() <= Opts.Limit);
1360     // We don't assert that isIncomplete means we hit a limit.
1361     // Indexes may choose to impose their own limits even if we don't have one.
1362   }
1363 
1364   CodeCompleteResult runWithoutSema(llvm::StringRef Content, size_t Offset,
1365                                     const ThreadsafeFS &TFS) && {
1366     trace::Span Tracer("CodeCompleteWithoutSema");
1367     // Fill in fields normally set by runWithSema()
1368     HeuristicPrefix = guessCompletionPrefix(Content, Offset);
1369     populateContextWords(Content);
1370     CCContextKind = CodeCompletionContext::CCC_Recovery;
1371     IsUsingDeclaration = false;
1372     Filter = FuzzyMatcher(HeuristicPrefix.Name);
1373     auto Pos = offsetToPosition(Content, Offset);
1374     ReplacedRange.start = ReplacedRange.end = Pos;
1375     ReplacedRange.start.character -= HeuristicPrefix.Name.size();
1376 
1377     llvm::StringMap<SourceParams> ProxSources;
1378     ProxSources[FileName].Cost = 0;
1379     FileProximity.emplace(ProxSources);
1380 
1381     auto Style = getFormatStyleForFile(FileName, Content, TFS);
1382     // This will only insert verbatim headers.
1383     Inserter.emplace(FileName, Content, Style,
1384                      /*BuildDir=*/"", /*HeaderSearchInfo=*/nullptr);
1385 
1386     auto Identifiers = collectIdentifiers(Content, Style);
1387     std::vector<RawIdentifier> IdentifierResults;
1388     for (const auto &IDAndCount : Identifiers) {
1389       RawIdentifier ID;
1390       ID.Name = IDAndCount.first();
1391       ID.References = IDAndCount.second;
1392       // Avoid treating typed filter as an identifier.
1393       if (ID.Name == HeuristicPrefix.Name)
1394         --ID.References;
1395       if (ID.References > 0)
1396         IdentifierResults.push_back(std::move(ID));
1397     }
1398 
1399     // Simplified version of getQueryScopes():
1400     //  - accessible scopes are determined heuristically.
1401     //  - all-scopes query if no qualifier was typed (and it's allowed).
1402     SpecifiedScope Scopes;
1403     Scopes.AccessibleScopes = visibleNamespaces(
1404         Content.take_front(Offset), format::getFormattingLangOpts(Style));
1405     for (std::string &S : Scopes.AccessibleScopes)
1406       if (!S.empty())
1407         S.append("::"); // visibleNamespaces doesn't include trailing ::.
1408     if (HeuristicPrefix.Qualifier.empty())
1409       AllScopes = Opts.AllScopes;
1410     else if (HeuristicPrefix.Qualifier.startswith("::")) {
1411       Scopes.AccessibleScopes = {""};
1412       Scopes.UnresolvedQualifier =
1413           std::string(HeuristicPrefix.Qualifier.drop_front(2));
1414     } else
1415       Scopes.UnresolvedQualifier = std::string(HeuristicPrefix.Qualifier);
1416     // First scope is the (modified) enclosing scope.
1417     QueryScopes = Scopes.scopesForIndexQuery();
1418     ScopeProximity.emplace(QueryScopes);
1419 
1420     SymbolSlab IndexResults = Opts.Index ? queryIndex() : SymbolSlab();
1421 
1422     CodeCompleteResult Output = toCodeCompleteResult(mergeResults(
1423         /*SemaResults=*/{}, IndexResults, IdentifierResults));
1424     Output.RanParser = false;
1425     logResults(Output, Tracer);
1426     return Output;
1427   }
1428 
1429 private:
1430   void populateContextWords(llvm::StringRef Content) {
1431     // Take last 3 lines before the completion point.
1432     unsigned RangeEnd = HeuristicPrefix.Qualifier.begin() - Content.data(),
1433              RangeBegin = RangeEnd;
1434     for (size_t I = 0; I < 3 && RangeBegin > 0; ++I) {
1435       auto PrevNL = Content.rfind('\n', RangeBegin);
1436       if (PrevNL == StringRef::npos) {
1437         RangeBegin = 0;
1438         break;
1439       }
1440       RangeBegin = PrevNL;
1441     }
1442 
1443     ContextWords = collectWords(Content.slice(RangeBegin, RangeEnd));
1444     dlog("Completion context words: {0}",
1445          llvm::join(ContextWords.keys(), ", "));
1446   }
1447 
1448   // This is called by run() once Sema code completion is done, but before the
1449   // Sema data structures are torn down. It does all the real work.
1450   CodeCompleteResult runWithSema() {
1451     const auto &CodeCompletionRange = CharSourceRange::getCharRange(
1452         Recorder->CCSema->getPreprocessor().getCodeCompletionTokenRange());
1453     // When we are getting completions with an empty identifier, for example
1454     //    std::vector<int> asdf;
1455     //    asdf.^;
1456     // Then the range will be invalid and we will be doing insertion, use
1457     // current cursor position in such cases as range.
1458     if (CodeCompletionRange.isValid()) {
1459       ReplacedRange = halfOpenToRange(Recorder->CCSema->getSourceManager(),
1460                                       CodeCompletionRange);
1461     } else {
1462       const auto &Pos = sourceLocToPosition(
1463           Recorder->CCSema->getSourceManager(),
1464           Recorder->CCSema->getPreprocessor().getCodeCompletionLoc());
1465       ReplacedRange.start = ReplacedRange.end = Pos;
1466     }
1467     Filter = FuzzyMatcher(
1468         Recorder->CCSema->getPreprocessor().getCodeCompletionFilter());
1469     std::tie(QueryScopes, AllScopes) = getQueryScopes(
1470         Recorder->CCContext, *Recorder->CCSema, HeuristicPrefix, Opts);
1471     if (!QueryScopes.empty())
1472       ScopeProximity.emplace(QueryScopes);
1473     PreferredType =
1474         OpaqueType::fromType(Recorder->CCSema->getASTContext(),
1475                              Recorder->CCContext.getPreferredType());
1476     // Sema provides the needed context to query the index.
1477     // FIXME: in addition to querying for extra/overlapping symbols, we should
1478     //        explicitly request symbols corresponding to Sema results.
1479     //        We can use their signals even if the index can't suggest them.
1480     // We must copy index results to preserve them, but there are at most Limit.
1481     auto IndexResults = (Opts.Index && allowIndex(Recorder->CCContext))
1482                             ? queryIndex()
1483                             : SymbolSlab();
1484     trace::Span Tracer("Populate CodeCompleteResult");
1485     // Merge Sema and Index results, score them, and pick the winners.
1486     auto Top =
1487         mergeResults(Recorder->Results, IndexResults, /*Identifiers*/ {});
1488     return toCodeCompleteResult(Top);
1489   }
1490 
1491   CodeCompleteResult
1492   toCodeCompleteResult(const std::vector<ScoredBundle> &Scored) {
1493     CodeCompleteResult Output;
1494 
1495     // Convert the results to final form, assembling the expensive strings.
1496     for (auto &C : Scored) {
1497       Output.Completions.push_back(toCodeCompletion(C.first));
1498       Output.Completions.back().Score = C.second;
1499       Output.Completions.back().CompletionTokenRange = ReplacedRange;
1500     }
1501     Output.HasMore = Incomplete;
1502     Output.Context = CCContextKind;
1503     Output.CompletionRange = ReplacedRange;
1504     return Output;
1505   }
1506 
1507   SymbolSlab queryIndex() {
1508     trace::Span Tracer("Query index");
1509     SPAN_ATTACH(Tracer, "limit", int64_t(Opts.Limit));
1510 
1511     // Build the query.
1512     FuzzyFindRequest Req;
1513     if (Opts.Limit)
1514       Req.Limit = Opts.Limit;
1515     Req.Query = std::string(Filter->pattern());
1516     Req.RestrictForCodeCompletion = true;
1517     Req.Scopes = QueryScopes;
1518     Req.AnyScope = AllScopes;
1519     // FIXME: we should send multiple weighted paths here.
1520     Req.ProximityPaths.push_back(std::string(FileName));
1521     if (PreferredType)
1522       Req.PreferredTypes.push_back(std::string(PreferredType->raw()));
1523     vlog("Code complete: fuzzyFind({0:2})", toJSON(Req));
1524 
1525     if (SpecFuzzyFind)
1526       SpecFuzzyFind->NewReq = Req;
1527     if (SpecFuzzyFind && SpecFuzzyFind->Result.valid() && (*SpecReq == Req)) {
1528       vlog("Code complete: speculative fuzzy request matches the actual index "
1529            "request. Waiting for the speculative index results.");
1530       SPAN_ATTACH(Tracer, "Speculative results", true);
1531 
1532       trace::Span WaitSpec("Wait speculative results");
1533       return SpecFuzzyFind->Result.get();
1534     }
1535 
1536     SPAN_ATTACH(Tracer, "Speculative results", false);
1537 
1538     // Run the query against the index.
1539     SymbolSlab::Builder ResultsBuilder;
1540     if (Opts.Index->fuzzyFind(
1541             Req, [&](const Symbol &Sym) { ResultsBuilder.insert(Sym); }))
1542       Incomplete = true;
1543     return std::move(ResultsBuilder).build();
1544   }
1545 
1546   // Merges Sema and Index results where possible, to form CompletionCandidates.
1547   // \p Identifiers is raw identifiers that can also be completion candidates.
1548   // Identifiers are not merged with results from index or sema.
1549   // Groups overloads if desired, to form CompletionCandidate::Bundles. The
1550   // bundles are scored and top results are returned, best to worst.
1551   std::vector<ScoredBundle>
1552   mergeResults(const std::vector<CodeCompletionResult> &SemaResults,
1553                const SymbolSlab &IndexResults,
1554                const std::vector<RawIdentifier> &IdentifierResults) {
1555     trace::Span Tracer("Merge and score results");
1556     std::vector<CompletionCandidate::Bundle> Bundles;
1557     llvm::DenseMap<size_t, size_t> BundleLookup;
1558     auto AddToBundles = [&](const CodeCompletionResult *SemaResult,
1559                             const Symbol *IndexResult,
1560                             const RawIdentifier *IdentifierResult) {
1561       CompletionCandidate C;
1562       C.SemaResult = SemaResult;
1563       C.IndexResult = IndexResult;
1564       C.IdentifierResult = IdentifierResult;
1565       if (C.IndexResult) {
1566         C.Name = IndexResult->Name;
1567         C.RankedIncludeHeaders = getRankedIncludes(*C.IndexResult);
1568       } else if (C.SemaResult) {
1569         C.Name = Recorder->getName(*SemaResult);
1570       } else {
1571         assert(IdentifierResult);
1572         C.Name = IdentifierResult->Name;
1573       }
1574       if (auto OverloadSet = C.overloadSet(Opts)) {
1575         auto Ret = BundleLookup.try_emplace(OverloadSet, Bundles.size());
1576         if (Ret.second)
1577           Bundles.emplace_back();
1578         Bundles[Ret.first->second].push_back(std::move(C));
1579       } else {
1580         Bundles.emplace_back();
1581         Bundles.back().push_back(std::move(C));
1582       }
1583     };
1584     llvm::DenseSet<const Symbol *> UsedIndexResults;
1585     auto CorrespondingIndexResult =
1586         [&](const CodeCompletionResult &SemaResult) -> const Symbol * {
1587       if (auto SymID =
1588               getSymbolID(SemaResult, Recorder->CCSema->getSourceManager())) {
1589         auto I = IndexResults.find(*SymID);
1590         if (I != IndexResults.end()) {
1591           UsedIndexResults.insert(&*I);
1592           return &*I;
1593         }
1594       }
1595       return nullptr;
1596     };
1597     // Emit all Sema results, merging them with Index results if possible.
1598     for (auto &SemaResult : SemaResults)
1599       AddToBundles(&SemaResult, CorrespondingIndexResult(SemaResult), nullptr);
1600     // Now emit any Index-only results.
1601     for (const auto &IndexResult : IndexResults) {
1602       if (UsedIndexResults.count(&IndexResult))
1603         continue;
1604       AddToBundles(/*SemaResult=*/nullptr, &IndexResult, nullptr);
1605     }
1606     // Emit identifier results.
1607     for (const auto &Ident : IdentifierResults)
1608       AddToBundles(/*SemaResult=*/nullptr, /*IndexResult=*/nullptr, &Ident);
1609     // We only keep the best N results at any time, in "native" format.
1610     TopN<ScoredBundle, ScoredBundleGreater> Top(
1611         Opts.Limit == 0 ? std::numeric_limits<size_t>::max() : Opts.Limit);
1612     for (auto &Bundle : Bundles)
1613       addCandidate(Top, std::move(Bundle));
1614     return std::move(Top).items();
1615   }
1616 
1617   llvm::Optional<float> fuzzyScore(const CompletionCandidate &C) {
1618     // Macros can be very spammy, so we only support prefix completion.
1619     if (((C.SemaResult &&
1620           C.SemaResult->Kind == CodeCompletionResult::RK_Macro) ||
1621          (C.IndexResult &&
1622           C.IndexResult->SymInfo.Kind == index::SymbolKind::Macro)) &&
1623         !C.Name.startswith_lower(Filter->pattern()))
1624       return None;
1625     return Filter->match(C.Name);
1626   }
1627 
1628   CodeCompletion::Scores
1629   evaluateCompletion(const SymbolQualitySignals &Quality,
1630                      const SymbolRelevanceSignals &Relevance) {
1631     using RM = CodeCompleteOptions::CodeCompletionRankingModel;
1632     CodeCompletion::Scores Scores;
1633     switch (Opts.RankingModel) {
1634     case RM::Heuristics:
1635       Scores.Quality = Quality.evaluateHeuristics();
1636       Scores.Relevance = Relevance.evaluateHeuristics();
1637       Scores.Total =
1638           evaluateSymbolAndRelevance(Scores.Quality, Scores.Relevance);
1639       // NameMatch is in fact a multiplier on total score, so rescoring is
1640       // sound.
1641       Scores.ExcludingName = Relevance.NameMatch
1642                                  ? Scores.Total / Relevance.NameMatch
1643                                  : Scores.Quality;
1644       return Scores;
1645 
1646     case RM::DecisionForest:
1647       Scores.Quality = 0;
1648       Scores.Relevance = 0;
1649       // Exponentiating DecisionForest prediction makes the score of each tree a
1650       // multiplciative boost (like NameMatch). This allows us to weigh the
1651       // prediciton score and NameMatch appropriately.
1652       Scores.ExcludingName = pow(Opts.DecisionForestBase,
1653                                  evaluateDecisionForest(Quality, Relevance));
1654       // NeedsFixIts is not part of the DecisionForest as generating training
1655       // data that needs fixits is not-feasible.
1656       if (Relevance.NeedsFixIts)
1657         Scores.ExcludingName *= 0.5;
1658       // NameMatch should be a multiplier on total score to support rescoring.
1659       Scores.Total = Relevance.NameMatch * Scores.ExcludingName;
1660       return Scores;
1661     }
1662     llvm_unreachable("Unhandled CodeCompletion ranking model.");
1663   }
1664 
1665   // Scores a candidate and adds it to the TopN structure.
1666   void addCandidate(TopN<ScoredBundle, ScoredBundleGreater> &Candidates,
1667                     CompletionCandidate::Bundle Bundle) {
1668     SymbolQualitySignals Quality;
1669     SymbolRelevanceSignals Relevance;
1670     Relevance.Context = CCContextKind;
1671     Relevance.Name = Bundle.front().Name;
1672     Relevance.FilterLength = HeuristicPrefix.Name.size();
1673     Relevance.Query = SymbolRelevanceSignals::CodeComplete;
1674     Relevance.FileProximityMatch = FileProximity.getPointer();
1675     if (ScopeProximity)
1676       Relevance.ScopeProximityMatch = ScopeProximity.getPointer();
1677     if (PreferredType)
1678       Relevance.HadContextType = true;
1679     Relevance.ContextWords = &ContextWords;
1680 
1681     auto &First = Bundle.front();
1682     if (auto FuzzyScore = fuzzyScore(First))
1683       Relevance.NameMatch = *FuzzyScore;
1684     else
1685       return;
1686     SymbolOrigin Origin = SymbolOrigin::Unknown;
1687     bool FromIndex = false;
1688     for (const auto &Candidate : Bundle) {
1689       if (Candidate.IndexResult) {
1690         Quality.merge(*Candidate.IndexResult);
1691         Relevance.merge(*Candidate.IndexResult);
1692         Origin |= Candidate.IndexResult->Origin;
1693         FromIndex = true;
1694         if (!Candidate.IndexResult->Type.empty())
1695           Relevance.HadSymbolType |= true;
1696         if (PreferredType &&
1697             PreferredType->raw() == Candidate.IndexResult->Type) {
1698           Relevance.TypeMatchesPreferred = true;
1699         }
1700       }
1701       if (Candidate.SemaResult) {
1702         Quality.merge(*Candidate.SemaResult);
1703         Relevance.merge(*Candidate.SemaResult);
1704         if (PreferredType) {
1705           if (auto CompletionType = OpaqueType::fromCompletionResult(
1706                   Recorder->CCSema->getASTContext(), *Candidate.SemaResult)) {
1707             Relevance.HadSymbolType |= true;
1708             if (PreferredType == CompletionType)
1709               Relevance.TypeMatchesPreferred = true;
1710           }
1711         }
1712         Origin |= SymbolOrigin::AST;
1713       }
1714       if (Candidate.IdentifierResult) {
1715         Quality.References = Candidate.IdentifierResult->References;
1716         Relevance.Scope = SymbolRelevanceSignals::FileScope;
1717         Origin |= SymbolOrigin::Identifier;
1718       }
1719     }
1720 
1721     CodeCompletion::Scores Scores = evaluateCompletion(Quality, Relevance);
1722     if (Opts.RecordCCResult)
1723       Opts.RecordCCResult(toCodeCompletion(Bundle), Quality, Relevance,
1724                           Scores.Total);
1725 
1726     dlog("CodeComplete: {0} ({1}) = {2}\n{3}{4}\n", First.Name,
1727          llvm::to_string(Origin), Scores.Total, llvm::to_string(Quality),
1728          llvm::to_string(Relevance));
1729 
1730     NSema += bool(Origin & SymbolOrigin::AST);
1731     NIndex += FromIndex;
1732     NSemaAndIndex += bool(Origin & SymbolOrigin::AST) && FromIndex;
1733     NIdent += bool(Origin & SymbolOrigin::Identifier);
1734     if (Candidates.push({std::move(Bundle), Scores}))
1735       Incomplete = true;
1736   }
1737 
1738   CodeCompletion toCodeCompletion(const CompletionCandidate::Bundle &Bundle) {
1739     llvm::Optional<CodeCompletionBuilder> Builder;
1740     for (const auto &Item : Bundle) {
1741       CodeCompletionString *SemaCCS =
1742           Item.SemaResult ? Recorder->codeCompletionString(*Item.SemaResult)
1743                           : nullptr;
1744       if (!Builder)
1745         Builder.emplace(Recorder ? &Recorder->CCSema->getASTContext() : nullptr,
1746                         Item, SemaCCS, QueryScopes, *Inserter, FileName,
1747                         CCContextKind, Opts, IsUsingDeclaration, NextTokenKind);
1748       else
1749         Builder->add(Item, SemaCCS);
1750     }
1751     return Builder->build();
1752   }
1753 };
1754 
1755 } // namespace
1756 
1757 clang::CodeCompleteOptions CodeCompleteOptions::getClangCompleteOpts() const {
1758   clang::CodeCompleteOptions Result;
1759   Result.IncludeCodePatterns = EnableSnippets && IncludeCodePatterns;
1760   Result.IncludeMacros = IncludeMacros;
1761   Result.IncludeGlobals = true;
1762   // We choose to include full comments and not do doxygen parsing in
1763   // completion.
1764   // FIXME: ideally, we should support doxygen in some form, e.g. do markdown
1765   // formatting of the comments.
1766   Result.IncludeBriefComments = false;
1767 
1768   // When an is used, Sema is responsible for completing the main file,
1769   // the index can provide results from the preamble.
1770   // Tell Sema not to deserialize the preamble to look for results.
1771   Result.LoadExternal = !Index;
1772   Result.IncludeFixIts = IncludeFixIts;
1773 
1774   return Result;
1775 }
1776 
1777 CompletionPrefix guessCompletionPrefix(llvm::StringRef Content,
1778                                        unsigned Offset) {
1779   assert(Offset <= Content.size());
1780   StringRef Rest = Content.take_front(Offset);
1781   CompletionPrefix Result;
1782 
1783   // Consume the unqualified name. We only handle ASCII characters.
1784   // isIdentifierBody will let us match "0invalid", but we don't mind.
1785   while (!Rest.empty() && isIdentifierBody(Rest.back()))
1786     Rest = Rest.drop_back();
1787   Result.Name = Content.slice(Rest.size(), Offset);
1788 
1789   // Consume qualifiers.
1790   while (Rest.consume_back("::") && !Rest.endswith(":")) // reject ::::
1791     while (!Rest.empty() && isIdentifierBody(Rest.back()))
1792       Rest = Rest.drop_back();
1793   Result.Qualifier =
1794       Content.slice(Rest.size(), Result.Name.begin() - Content.begin());
1795 
1796   return Result;
1797 }
1798 
1799 CodeCompleteResult codeComplete(PathRef FileName, Position Pos,
1800                                 const PreambleData *Preamble,
1801                                 const ParseInputs &ParseInput,
1802                                 CodeCompleteOptions Opts,
1803                                 SpeculativeFuzzyFind *SpecFuzzyFind) {
1804   auto Offset = positionToOffset(ParseInput.Contents, Pos);
1805   if (!Offset) {
1806     elog("Code completion position was invalid {0}", Offset.takeError());
1807     return CodeCompleteResult();
1808   }
1809   auto Flow = CodeCompleteFlow(
1810       FileName, Preamble ? Preamble->Includes : IncludeStructure(),
1811       SpecFuzzyFind, Opts);
1812   return (!Preamble || Opts.RunParser == CodeCompleteOptions::NeverParse)
1813              ? std::move(Flow).runWithoutSema(ParseInput.Contents, *Offset,
1814                                               *ParseInput.TFS)
1815              : std::move(Flow).run({FileName, *Offset, *Preamble,
1816                                     // We want to serve code completions with
1817                                     // low latency, so don't bother patching.
1818                                     /*PreamblePatch=*/llvm::None, ParseInput});
1819 }
1820 
1821 SignatureHelp signatureHelp(PathRef FileName, Position Pos,
1822                             const PreambleData &Preamble,
1823                             const ParseInputs &ParseInput) {
1824   auto Offset = positionToOffset(ParseInput.Contents, Pos);
1825   if (!Offset) {
1826     elog("Signature help position was invalid {0}", Offset.takeError());
1827     return SignatureHelp();
1828   }
1829   SignatureHelp Result;
1830   clang::CodeCompleteOptions Options;
1831   Options.IncludeGlobals = false;
1832   Options.IncludeMacros = false;
1833   Options.IncludeCodePatterns = false;
1834   Options.IncludeBriefComments = false;
1835   semaCodeComplete(
1836       std::make_unique<SignatureHelpCollector>(Options, ParseInput.Index,
1837                                                Result),
1838       Options,
1839       {FileName, *Offset, Preamble,
1840        PreamblePatch::create(FileName, ParseInput, Preamble), ParseInput});
1841   return Result;
1842 }
1843 
1844 bool isIndexedForCodeCompletion(const NamedDecl &ND, ASTContext &ASTCtx) {
1845   auto InTopLevelScope = [](const NamedDecl &ND) {
1846     switch (ND.getDeclContext()->getDeclKind()) {
1847     case Decl::TranslationUnit:
1848     case Decl::Namespace:
1849     case Decl::LinkageSpec:
1850       return true;
1851     default:
1852       break;
1853     };
1854     return false;
1855   };
1856   // We only complete symbol's name, which is the same as the name of the
1857   // *primary* template in case of template specializations.
1858   if (isExplicitTemplateSpecialization(&ND))
1859     return false;
1860 
1861   if (InTopLevelScope(ND))
1862     return true;
1863 
1864   if (const auto *EnumDecl = dyn_cast<clang::EnumDecl>(ND.getDeclContext()))
1865     return InTopLevelScope(*EnumDecl) && !EnumDecl->isScoped();
1866 
1867   return false;
1868 }
1869 
1870 // FIXME: find a home for this (that can depend on both markup and Protocol).
1871 static MarkupContent renderDoc(const markup::Document &Doc, MarkupKind Kind) {
1872   MarkupContent Result;
1873   Result.kind = Kind;
1874   switch (Kind) {
1875   case MarkupKind::PlainText:
1876     Result.value.append(Doc.asPlainText());
1877     break;
1878   case MarkupKind::Markdown:
1879     Result.value.append(Doc.asMarkdown());
1880     break;
1881   }
1882   return Result;
1883 }
1884 
1885 CompletionItem CodeCompletion::render(const CodeCompleteOptions &Opts) const {
1886   CompletionItem LSP;
1887   const auto *InsertInclude = Includes.empty() ? nullptr : &Includes[0];
1888   LSP.label = ((InsertInclude && InsertInclude->Insertion)
1889                    ? Opts.IncludeIndicator.Insert
1890                    : Opts.IncludeIndicator.NoInsert) +
1891               (Opts.ShowOrigins ? "[" + llvm::to_string(Origin) + "]" : "") +
1892               RequiredQualifier + Name + Signature;
1893 
1894   LSP.kind = Kind;
1895   LSP.detail = BundleSize > 1
1896                    ? std::string(llvm::formatv("[{0} overloads]", BundleSize))
1897                    : ReturnType;
1898   LSP.deprecated = Deprecated;
1899   // Combine header information and documentation in LSP `documentation` field.
1900   // This is not quite right semantically, but tends to display well in editors.
1901   if (InsertInclude || Documentation) {
1902     markup::Document Doc;
1903     if (InsertInclude)
1904       Doc.addParagraph().appendText("From ").appendCode(InsertInclude->Header);
1905     if (Documentation)
1906       Doc.append(*Documentation);
1907     LSP.documentation = renderDoc(Doc, Opts.DocumentationFormat);
1908   }
1909   LSP.sortText = sortText(Score.Total, Name);
1910   LSP.filterText = Name;
1911   LSP.textEdit = {CompletionTokenRange, RequiredQualifier + Name};
1912   // Merge continuous additionalTextEdits into main edit. The main motivation
1913   // behind this is to help LSP clients, it seems most of them are confused when
1914   // they are provided with additionalTextEdits that are consecutive to main
1915   // edit.
1916   // Note that we store additional text edits from back to front in a line. That
1917   // is mainly to help LSP clients again, so that changes do not effect each
1918   // other.
1919   for (const auto &FixIt : FixIts) {
1920     if (FixIt.range.end == LSP.textEdit->range.start) {
1921       LSP.textEdit->newText = FixIt.newText + LSP.textEdit->newText;
1922       LSP.textEdit->range.start = FixIt.range.start;
1923     } else {
1924       LSP.additionalTextEdits.push_back(FixIt);
1925     }
1926   }
1927   if (Opts.EnableSnippets)
1928     LSP.textEdit->newText += SnippetSuffix;
1929 
1930   // FIXME(kadircet): Do not even fill insertText after making sure textEdit is
1931   // compatible with most of the editors.
1932   LSP.insertText = LSP.textEdit->newText;
1933   LSP.insertTextFormat = Opts.EnableSnippets ? InsertTextFormat::Snippet
1934                                              : InsertTextFormat::PlainText;
1935   if (InsertInclude && InsertInclude->Insertion)
1936     LSP.additionalTextEdits.push_back(*InsertInclude->Insertion);
1937 
1938   LSP.score = Score.ExcludingName;
1939 
1940   return LSP;
1941 }
1942 
1943 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, const CodeCompletion &C) {
1944   // For now just lean on CompletionItem.
1945   return OS << C.render(CodeCompleteOptions());
1946 }
1947 
1948 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS,
1949                               const CodeCompleteResult &R) {
1950   OS << "CodeCompleteResult: " << R.Completions.size() << (R.HasMore ? "+" : "")
1951      << " (" << getCompletionKindString(R.Context) << ")"
1952      << " items:\n";
1953   for (const auto &C : R.Completions)
1954     OS << C << "\n";
1955   return OS;
1956 }
1957 
1958 // Heuristically detect whether the `Line` is an unterminated include filename.
1959 bool isIncludeFile(llvm::StringRef Line) {
1960   Line = Line.ltrim();
1961   if (!Line.consume_front("#"))
1962     return false;
1963   Line = Line.ltrim();
1964   if (!(Line.consume_front("include_next") || Line.consume_front("include") ||
1965         Line.consume_front("import")))
1966     return false;
1967   Line = Line.ltrim();
1968   if (Line.consume_front("<"))
1969     return Line.count('>') == 0;
1970   if (Line.consume_front("\""))
1971     return Line.count('"') == 0;
1972   return false;
1973 }
1974 
1975 bool allowImplicitCompletion(llvm::StringRef Content, unsigned Offset) {
1976   // Look at last line before completion point only.
1977   Content = Content.take_front(Offset);
1978   auto Pos = Content.rfind('\n');
1979   if (Pos != llvm::StringRef::npos)
1980     Content = Content.substr(Pos + 1);
1981 
1982   // Complete after scope operators.
1983   if (Content.endswith(".") || Content.endswith("->") || Content.endswith("::"))
1984     return true;
1985   // Complete after `#include <` and #include `<foo/`.
1986   if ((Content.endswith("<") || Content.endswith("\"") ||
1987        Content.endswith("/")) &&
1988       isIncludeFile(Content))
1989     return true;
1990 
1991   // Complete words. Give non-ascii characters the benefit of the doubt.
1992   return !Content.empty() &&
1993          (isIdentifierBody(Content.back()) || !llvm::isASCII(Content.back()));
1994 }
1995 
1996 } // namespace clangd
1997 } // namespace clang
1998