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