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 "Logger.h" 30 #include "Preamble.h" 31 #include "Protocol.h" 32 #include "Quality.h" 33 #include "SourceCode.h" 34 #include "TUScheduler.h" 35 #include "Threading.h" 36 #include "Trace.h" 37 #include "URI.h" 38 #include "index/Index.h" 39 #include "index/Symbol.h" 40 #include "index/SymbolOrigin.h" 41 #include "clang/AST/Decl.h" 42 #include "clang/AST/DeclBase.h" 43 #include "clang/Basic/CharInfo.h" 44 #include "clang/Basic/LangOptions.h" 45 #include "clang/Basic/SourceLocation.h" 46 #include "clang/Format/Format.h" 47 #include "clang/Frontend/CompilerInstance.h" 48 #include "clang/Frontend/FrontendActions.h" 49 #include "clang/Lex/ExternalPreprocessorSource.h" 50 #include "clang/Lex/Preprocessor.h" 51 #include "clang/Lex/PreprocessorOptions.h" 52 #include "clang/Sema/CodeCompleteConsumer.h" 53 #include "clang/Sema/DeclSpec.h" 54 #include "clang/Sema/Sema.h" 55 #include "llvm/ADT/ArrayRef.h" 56 #include "llvm/ADT/None.h" 57 #include "llvm/ADT/Optional.h" 58 #include "llvm/ADT/SmallVector.h" 59 #include "llvm/ADT/StringExtras.h" 60 #include "llvm/ADT/StringRef.h" 61 #include "llvm/Support/Compiler.h" 62 #include "llvm/Support/Debug.h" 63 #include "llvm/Support/Error.h" 64 #include "llvm/Support/Format.h" 65 #include "llvm/Support/FormatVariadic.h" 66 #include "llvm/Support/ScopedPrinter.h" 67 #include <algorithm> 68 #include <iterator> 69 70 // We log detailed candidate here if you run with -debug-only=codecomplete. 71 #define DEBUG_TYPE "CodeComplete" 72 73 namespace clang { 74 namespace clangd { 75 namespace { 76 77 CompletionItemKind toCompletionItemKind(index::SymbolKind Kind) { 78 using SK = index::SymbolKind; 79 switch (Kind) { 80 case SK::Unknown: 81 return CompletionItemKind::Missing; 82 case SK::Module: 83 case SK::Namespace: 84 case SK::NamespaceAlias: 85 return CompletionItemKind::Module; 86 case SK::Macro: 87 return CompletionItemKind::Text; 88 case SK::Enum: 89 return CompletionItemKind::Enum; 90 case SK::Struct: 91 return CompletionItemKind::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 case SK::ConversionFunction: 105 return CompletionItemKind::Function; 106 case SK::Variable: 107 case SK::Parameter: 108 case SK::NonTypeTemplateParm: 109 return CompletionItemKind::Variable; 110 case SK::Field: 111 return CompletionItemKind::Field; 112 case SK::EnumConstant: 113 return CompletionItemKind::EnumMember; 114 case SK::InstanceMethod: 115 case SK::ClassMethod: 116 case SK::StaticMethod: 117 case SK::Destructor: 118 return CompletionItemKind::Method; 119 case SK::InstanceProperty: 120 case SK::ClassProperty: 121 case SK::StaticProperty: 122 return CompletionItemKind::Property; 123 case SK::Constructor: 124 return CompletionItemKind::Constructor; 125 case SK::TemplateTypeParm: 126 case SK::TemplateTemplateParm: 127 return CompletionItemKind::TypeParameter; 128 } 129 llvm_unreachable("Unhandled clang::index::SymbolKind."); 130 } 131 132 CompletionItemKind 133 toCompletionItemKind(CodeCompletionResult::ResultKind ResKind, 134 const NamedDecl *Decl, 135 CodeCompletionContext::Kind CtxKind) { 136 if (Decl) 137 return toCompletionItemKind(index::getSymbolInfo(Decl).Kind); 138 if (CtxKind == CodeCompletionContext::CCC_IncludedFile) 139 return CompletionItemKind::File; 140 switch (ResKind) { 141 case CodeCompletionResult::RK_Declaration: 142 llvm_unreachable("RK_Declaration without Decl"); 143 case CodeCompletionResult::RK_Keyword: 144 return CompletionItemKind::Keyword; 145 case CodeCompletionResult::RK_Macro: 146 return CompletionItemKind::Text; // unfortunately, there's no 'Macro' 147 // completion items in LSP. 148 case CodeCompletionResult::RK_Pattern: 149 return CompletionItemKind::Snippet; 150 } 151 llvm_unreachable("Unhandled CodeCompletionResult::ResultKind."); 152 } 153 154 // Identifier code completion result. 155 struct RawIdentifier { 156 llvm::StringRef Name; 157 unsigned References; // # of usages in file. 158 }; 159 160 /// A code completion result, in clang-native form. 161 /// It may be promoted to a CompletionItem if it's among the top-ranked results. 162 struct CompletionCandidate { 163 llvm::StringRef Name; // Used for filtering and sorting. 164 // We may have a result from Sema, from the index, or both. 165 const CodeCompletionResult *SemaResult = nullptr; 166 const Symbol *IndexResult = nullptr; 167 const RawIdentifier *IdentifierResult = nullptr; 168 llvm::SmallVector<llvm::StringRef, 1> RankedIncludeHeaders; 169 170 // Returns a token identifying the overload set this is part of. 171 // 0 indicates it's not part of any overload set. 172 size_t overloadSet(const CodeCompleteOptions &Opts) const { 173 if (!Opts.BundleOverloads.getValueOr(false)) 174 return 0; 175 llvm::SmallString<256> Scratch; 176 if (IndexResult) { 177 switch (IndexResult->SymInfo.Kind) { 178 case index::SymbolKind::ClassMethod: 179 case index::SymbolKind::InstanceMethod: 180 case index::SymbolKind::StaticMethod: 181 #ifndef NDEBUG 182 llvm_unreachable("Don't expect members from index in code completion"); 183 #else 184 LLVM_FALLTHROUGH; 185 #endif 186 case index::SymbolKind::Function: 187 // We can't group overloads together that need different #includes. 188 // This could break #include insertion. 189 return llvm::hash_combine( 190 (IndexResult->Scope + IndexResult->Name).toStringRef(Scratch), 191 headerToInsertIfAllowed(Opts).getValueOr("")); 192 default: 193 return 0; 194 } 195 } 196 if (SemaResult) { 197 // We need to make sure we're consistent with the IndexResult case! 198 const NamedDecl *D = SemaResult->Declaration; 199 if (!D || !D->isFunctionOrFunctionTemplate()) 200 return 0; 201 { 202 llvm::raw_svector_ostream OS(Scratch); 203 D->printQualifiedName(OS); 204 } 205 return llvm::hash_combine(Scratch, 206 headerToInsertIfAllowed(Opts).getValueOr("")); 207 } 208 assert(IdentifierResult); 209 return 0; 210 } 211 212 // The best header to include if include insertion is allowed. 213 llvm::Optional<llvm::StringRef> 214 headerToInsertIfAllowed(const CodeCompleteOptions &Opts) const { 215 if (Opts.InsertIncludes == CodeCompleteOptions::NeverInsert || 216 RankedIncludeHeaders.empty()) 217 return None; 218 if (SemaResult && SemaResult->Declaration) { 219 // Avoid inserting new #include if the declaration is found in the current 220 // file e.g. the symbol is forward declared. 221 auto &SM = SemaResult->Declaration->getASTContext().getSourceManager(); 222 for (const Decl *RD : SemaResult->Declaration->redecls()) 223 if (SM.isInMainFile(SM.getExpansionLoc(RD->getBeginLoc()))) 224 return None; 225 } 226 return RankedIncludeHeaders[0]; 227 } 228 229 using Bundle = llvm::SmallVector<CompletionCandidate, 4>; 230 }; 231 using ScoredBundle = 232 std::pair<CompletionCandidate::Bundle, CodeCompletion::Scores>; 233 struct ScoredBundleGreater { 234 bool operator()(const ScoredBundle &L, const ScoredBundle &R) { 235 if (L.second.Total != R.second.Total) 236 return L.second.Total > R.second.Total; 237 return L.first.front().Name < 238 R.first.front().Name; // Earlier name is better. 239 } 240 }; 241 242 // Assembles a code completion out of a bundle of >=1 completion candidates. 243 // Many of the expensive strings are only computed at this point, once we know 244 // the candidate bundle is going to be returned. 245 // 246 // Many fields are the same for all candidates in a bundle (e.g. name), and are 247 // computed from the first candidate, in the constructor. 248 // Others vary per candidate, so add() must be called for remaining candidates. 249 struct CodeCompletionBuilder { 250 CodeCompletionBuilder(ASTContext *ASTCtx, const CompletionCandidate &C, 251 CodeCompletionString *SemaCCS, 252 llvm::ArrayRef<std::string> QueryScopes, 253 const IncludeInserter &Includes, 254 llvm::StringRef FileName, 255 CodeCompletionContext::Kind ContextKind, 256 const CodeCompleteOptions &Opts, bool GenerateSnippets) 257 : ASTCtx(ASTCtx), ExtractDocumentation(Opts.IncludeComments), 258 EnableFunctionArgSnippets(Opts.EnableFunctionArgSnippets), 259 GenerateSnippets(GenerateSnippets) { 260 add(C, SemaCCS); 261 if (C.SemaResult) { 262 assert(ASTCtx); 263 Completion.Origin |= SymbolOrigin::AST; 264 Completion.Name = std::string(llvm::StringRef(SemaCCS->getTypedText())); 265 if (Completion.Scope.empty()) { 266 if ((C.SemaResult->Kind == CodeCompletionResult::RK_Declaration) || 267 (C.SemaResult->Kind == CodeCompletionResult::RK_Pattern)) 268 if (const auto *D = C.SemaResult->getDeclaration()) 269 if (const auto *ND = dyn_cast<NamedDecl>(D)) 270 Completion.Scope = std::string( 271 splitQualifiedName(printQualifiedName(*ND)).first); 272 } 273 Completion.Kind = toCompletionItemKind( 274 C.SemaResult->Kind, C.SemaResult->Declaration, ContextKind); 275 // Sema could provide more info on whether the completion was a file or 276 // folder. 277 if (Completion.Kind == CompletionItemKind::File && 278 Completion.Name.back() == '/') 279 Completion.Kind = CompletionItemKind::Folder; 280 for (const auto &FixIt : C.SemaResult->FixIts) { 281 Completion.FixIts.push_back(toTextEdit( 282 FixIt, ASTCtx->getSourceManager(), ASTCtx->getLangOpts())); 283 } 284 llvm::sort(Completion.FixIts, [](const TextEdit &X, const TextEdit &Y) { 285 return std::tie(X.range.start.line, X.range.start.character) < 286 std::tie(Y.range.start.line, Y.range.start.character); 287 }); 288 Completion.Deprecated |= 289 (C.SemaResult->Availability == CXAvailability_Deprecated); 290 } 291 if (C.IndexResult) { 292 Completion.Origin |= C.IndexResult->Origin; 293 if (Completion.Scope.empty()) 294 Completion.Scope = std::string(C.IndexResult->Scope); 295 if (Completion.Kind == CompletionItemKind::Missing) 296 Completion.Kind = toCompletionItemKind(C.IndexResult->SymInfo.Kind); 297 if (Completion.Name.empty()) 298 Completion.Name = std::string(C.IndexResult->Name); 299 // If the completion was visible to Sema, no qualifier is needed. This 300 // avoids unneeded qualifiers in cases like with `using ns::X`. 301 if (Completion.RequiredQualifier.empty() && !C.SemaResult) { 302 llvm::StringRef ShortestQualifier = C.IndexResult->Scope; 303 for (llvm::StringRef Scope : QueryScopes) { 304 llvm::StringRef Qualifier = C.IndexResult->Scope; 305 if (Qualifier.consume_front(Scope) && 306 Qualifier.size() < ShortestQualifier.size()) 307 ShortestQualifier = Qualifier; 308 } 309 Completion.RequiredQualifier = std::string(ShortestQualifier); 310 } 311 Completion.Deprecated |= (C.IndexResult->Flags & Symbol::Deprecated); 312 } 313 if (C.IdentifierResult) { 314 Completion.Origin |= SymbolOrigin::Identifier; 315 Completion.Kind = CompletionItemKind::Text; 316 Completion.Name = std::string(C.IdentifierResult->Name); 317 } 318 319 // Turn absolute path into a literal string that can be #included. 320 auto Inserted = [&](llvm::StringRef Header) 321 -> llvm::Expected<std::pair<std::string, bool>> { 322 auto ResolvedDeclaring = 323 URI::resolve(C.IndexResult->CanonicalDeclaration.FileURI, FileName); 324 if (!ResolvedDeclaring) 325 return ResolvedDeclaring.takeError(); 326 auto ResolvedInserted = toHeaderFile(Header, FileName); 327 if (!ResolvedInserted) 328 return ResolvedInserted.takeError(); 329 auto Spelled = Includes.calculateIncludePath(*ResolvedInserted, FileName); 330 if (!Spelled) 331 return llvm::createStringError(llvm::inconvertibleErrorCode(), 332 "Header not on include path"); 333 return std::make_pair( 334 std::move(*Spelled), 335 Includes.shouldInsertInclude(*ResolvedDeclaring, *ResolvedInserted)); 336 }; 337 bool ShouldInsert = C.headerToInsertIfAllowed(Opts).hasValue(); 338 // Calculate include paths and edits for all possible headers. 339 for (const auto &Inc : C.RankedIncludeHeaders) { 340 if (auto ToInclude = Inserted(Inc)) { 341 CodeCompletion::IncludeCandidate Include; 342 Include.Header = ToInclude->first; 343 if (ToInclude->second && ShouldInsert) 344 Include.Insertion = Includes.insert(ToInclude->first); 345 Completion.Includes.push_back(std::move(Include)); 346 } else 347 log("Failed to generate include insertion edits for adding header " 348 "(FileURI='{0}', IncludeHeader='{1}') into {2}: {3}", 349 C.IndexResult->CanonicalDeclaration.FileURI, Inc, FileName, 350 ToInclude.takeError()); 351 } 352 // Prefer includes that do not need edits (i.e. already exist). 353 std::stable_partition(Completion.Includes.begin(), 354 Completion.Includes.end(), 355 [](const CodeCompletion::IncludeCandidate &I) { 356 return !I.Insertion.hasValue(); 357 }); 358 } 359 360 void add(const CompletionCandidate &C, CodeCompletionString *SemaCCS) { 361 assert(bool(C.SemaResult) == bool(SemaCCS)); 362 Bundled.emplace_back(); 363 BundledEntry &S = Bundled.back(); 364 if (C.SemaResult) { 365 bool IsPattern = C.SemaResult->Kind == CodeCompletionResult::RK_Pattern; 366 getSignature(*SemaCCS, &S.Signature, &S.SnippetSuffix, 367 &Completion.RequiredQualifier, IsPattern); 368 S.ReturnType = getReturnType(*SemaCCS); 369 } else if (C.IndexResult) { 370 S.Signature = std::string(C.IndexResult->Signature); 371 S.SnippetSuffix = std::string(C.IndexResult->CompletionSnippetSuffix); 372 S.ReturnType = std::string(C.IndexResult->ReturnType); 373 } 374 if (ExtractDocumentation && Completion.Documentation.empty()) { 375 if (C.IndexResult) 376 Completion.Documentation = std::string(C.IndexResult->Documentation); 377 else if (C.SemaResult) 378 Completion.Documentation = getDocComment(*ASTCtx, *C.SemaResult, 379 /*CommentsFromHeader=*/false); 380 } 381 } 382 383 CodeCompletion build() { 384 Completion.ReturnType = summarizeReturnType(); 385 Completion.Signature = summarizeSignature(); 386 Completion.SnippetSuffix = summarizeSnippet(); 387 Completion.BundleSize = Bundled.size(); 388 return std::move(Completion); 389 } 390 391 private: 392 struct BundledEntry { 393 std::string SnippetSuffix; 394 std::string Signature; 395 std::string ReturnType; 396 }; 397 398 // If all BundledEntries have the same value for a property, return it. 399 template <std::string BundledEntry::*Member> 400 const std::string *onlyValue() const { 401 auto B = Bundled.begin(), E = Bundled.end(); 402 for (auto I = B + 1; I != E; ++I) 403 if (I->*Member != B->*Member) 404 return nullptr; 405 return &(B->*Member); 406 } 407 408 template <bool BundledEntry::*Member> const bool *onlyValue() const { 409 auto B = Bundled.begin(), E = Bundled.end(); 410 for (auto I = B + 1; I != E; ++I) 411 if (I->*Member != B->*Member) 412 return nullptr; 413 return &(B->*Member); 414 } 415 416 std::string summarizeReturnType() const { 417 if (auto *RT = onlyValue<&BundledEntry::ReturnType>()) 418 return *RT; 419 return ""; 420 } 421 422 std::string summarizeSnippet() const { 423 if (!GenerateSnippets) 424 return ""; 425 auto *Snippet = onlyValue<&BundledEntry::SnippetSuffix>(); 426 if (!Snippet) 427 // All bundles are function calls. 428 // FIXME(ibiryukov): sometimes add template arguments to a snippet, e.g. 429 // we need to complete 'forward<$1>($0)'. 430 return "($0)"; 431 if (EnableFunctionArgSnippets) 432 return *Snippet; 433 434 // Replace argument snippets with a simplified pattern. 435 if (Snippet->empty()) 436 return ""; 437 if (Completion.Kind == CompletionItemKind::Function || 438 Completion.Kind == CompletionItemKind::Method) { 439 // Functions snippets can be of 2 types: 440 // - containing only function arguments, e.g. 441 // foo(${1:int p1}, ${2:int p2}); 442 // We transform this pattern to '($0)' or '()'. 443 // - template arguments and function arguments, e.g. 444 // foo<${1:class}>(${2:int p1}). 445 // We transform this pattern to '<$1>()$0' or '<$0>()'. 446 447 bool EmptyArgs = llvm::StringRef(*Snippet).endswith("()"); 448 if (Snippet->front() == '<') 449 return EmptyArgs ? "<$1>()$0" : "<$1>($0)"; 450 if (Snippet->front() == '(') 451 return EmptyArgs ? "()" : "($0)"; 452 return *Snippet; // Not an arg snippet? 453 } 454 // 'CompletionItemKind::Interface' matches template type aliases. 455 if (Completion.Kind == CompletionItemKind::Interface || 456 Completion.Kind == CompletionItemKind::Class) { 457 if (Snippet->front() != '<') 458 return *Snippet; // Not an arg snippet? 459 460 // Classes and template using aliases can only have template arguments, 461 // e.g. Foo<${1:class}>. 462 if (llvm::StringRef(*Snippet).endswith("<>")) 463 return "<>"; // can happen with defaulted template arguments. 464 return "<$0>"; 465 } 466 return *Snippet; 467 } 468 469 std::string summarizeSignature() const { 470 if (auto *Signature = onlyValue<&BundledEntry::Signature>()) 471 return *Signature; 472 // All bundles are function calls. 473 return "(…)"; 474 } 475 476 // ASTCtx can be nullptr if not run with sema. 477 ASTContext *ASTCtx; 478 CodeCompletion Completion; 479 llvm::SmallVector<BundledEntry, 1> Bundled; 480 bool ExtractDocumentation; 481 bool EnableFunctionArgSnippets; 482 /// When false, no snippets are generated argument lists. 483 bool GenerateSnippets; 484 }; 485 486 // Determine the symbol ID for a Sema code completion result, if possible. 487 llvm::Optional<SymbolID> getSymbolID(const CodeCompletionResult &R, 488 const SourceManager &SM) { 489 switch (R.Kind) { 490 case CodeCompletionResult::RK_Declaration: 491 case CodeCompletionResult::RK_Pattern: { 492 // Computing USR caches linkage, which may change after code completion. 493 if (hasUnstableLinkage(R.Declaration)) 494 return llvm::None; 495 return clang::clangd::getSymbolID(R.Declaration); 496 } 497 case CodeCompletionResult::RK_Macro: 498 return clang::clangd::getSymbolID(R.Macro->getName(), R.MacroDefInfo, SM); 499 case CodeCompletionResult::RK_Keyword: 500 return None; 501 } 502 llvm_unreachable("unknown CodeCompletionResult kind"); 503 } 504 505 // Scopes of the partial identifier we're trying to complete. 506 // It is used when we query the index for more completion results. 507 struct SpecifiedScope { 508 // The scopes we should look in, determined by Sema. 509 // 510 // If the qualifier was fully resolved, we look for completions in these 511 // scopes; if there is an unresolved part of the qualifier, it should be 512 // resolved within these scopes. 513 // 514 // Examples of qualified completion: 515 // 516 // "::vec" => {""} 517 // "using namespace std; ::vec^" => {"", "std::"} 518 // "namespace ns {using namespace std;} ns::^" => {"ns::", "std::"} 519 // "std::vec^" => {""} // "std" unresolved 520 // 521 // Examples of unqualified completion: 522 // 523 // "vec^" => {""} 524 // "using namespace std; vec^" => {"", "std::"} 525 // "using namespace std; namespace ns { vec^ }" => {"ns::", "std::", ""} 526 // 527 // "" for global namespace, "ns::" for normal namespace. 528 std::vector<std::string> AccessibleScopes; 529 // The full scope qualifier as typed by the user (without the leading "::"). 530 // Set if the qualifier is not fully resolved by Sema. 531 llvm::Optional<std::string> UnresolvedQualifier; 532 533 // Construct scopes being queried in indexes. The results are deduplicated. 534 // This method format the scopes to match the index request representation. 535 std::vector<std::string> scopesForIndexQuery() { 536 std::set<std::string> Results; 537 for (llvm::StringRef AS : AccessibleScopes) 538 Results.insert( 539 (AS + (UnresolvedQualifier ? *UnresolvedQualifier : "")).str()); 540 return {Results.begin(), Results.end()}; 541 } 542 }; 543 544 // Get all scopes that will be queried in indexes and whether symbols from 545 // any scope is allowed. The first scope in the list is the preferred scope 546 // (e.g. enclosing namespace). 547 std::pair<std::vector<std::string>, bool> 548 getQueryScopes(CodeCompletionContext &CCContext, const Sema &CCSema, 549 const CompletionPrefix &HeuristicPrefix, 550 const CodeCompleteOptions &Opts) { 551 SpecifiedScope Scopes; 552 for (auto *Context : CCContext.getVisitedContexts()) { 553 if (isa<TranslationUnitDecl>(Context)) 554 Scopes.AccessibleScopes.push_back(""); // global namespace 555 else if (isa<NamespaceDecl>(Context)) 556 Scopes.AccessibleScopes.push_back(printNamespaceScope(*Context)); 557 } 558 559 const CXXScopeSpec *SemaSpecifier = 560 CCContext.getCXXScopeSpecifier().getValueOr(nullptr); 561 // Case 1: unqualified completion. 562 if (!SemaSpecifier) { 563 // Case 2 (exception): sema saw no qualifier, but there appears to be one! 564 // This can happen e.g. in incomplete macro expansions. Use heuristics. 565 if (!HeuristicPrefix.Qualifier.empty()) { 566 vlog("Sema said no scope specifier, but we saw {0} in the source code", 567 HeuristicPrefix.Qualifier); 568 StringRef SpelledSpecifier = HeuristicPrefix.Qualifier; 569 if (SpelledSpecifier.consume_front("::")) 570 Scopes.AccessibleScopes = {""}; 571 Scopes.UnresolvedQualifier = std::string(SpelledSpecifier); 572 return {Scopes.scopesForIndexQuery(), false}; 573 } 574 // The enclosing namespace must be first, it gets a quality boost. 575 std::vector<std::string> EnclosingAtFront; 576 std::string EnclosingScope = printNamespaceScope(*CCSema.CurContext); 577 EnclosingAtFront.push_back(EnclosingScope); 578 for (auto &S : Scopes.scopesForIndexQuery()) { 579 if (EnclosingScope != S) 580 EnclosingAtFront.push_back(std::move(S)); 581 } 582 // Allow AllScopes completion as there is no explicit scope qualifier. 583 return {EnclosingAtFront, Opts.AllScopes}; 584 } 585 // Case 3: sema saw and resolved a scope qualifier. 586 if (SemaSpecifier && SemaSpecifier->isValid()) 587 return {Scopes.scopesForIndexQuery(), false}; 588 589 // Case 4: There was a qualifier, and Sema didn't resolve it. 590 Scopes.AccessibleScopes.push_back(""); // Make sure global scope is included. 591 llvm::StringRef SpelledSpecifier = Lexer::getSourceText( 592 CharSourceRange::getCharRange(SemaSpecifier->getRange()), 593 CCSema.SourceMgr, clang::LangOptions()); 594 if (SpelledSpecifier.consume_front("::")) 595 Scopes.AccessibleScopes = {""}; 596 Scopes.UnresolvedQualifier = std::string(SpelledSpecifier); 597 // Sema excludes the trailing "::". 598 if (!Scopes.UnresolvedQualifier->empty()) 599 *Scopes.UnresolvedQualifier += "::"; 600 601 return {Scopes.scopesForIndexQuery(), false}; 602 } 603 604 // Should we perform index-based completion in a context of the specified kind? 605 // FIXME: consider allowing completion, but restricting the result types. 606 bool contextAllowsIndex(enum CodeCompletionContext::Kind K) { 607 switch (K) { 608 case CodeCompletionContext::CCC_TopLevel: 609 case CodeCompletionContext::CCC_ObjCInterface: 610 case CodeCompletionContext::CCC_ObjCImplementation: 611 case CodeCompletionContext::CCC_ObjCIvarList: 612 case CodeCompletionContext::CCC_ClassStructUnion: 613 case CodeCompletionContext::CCC_Statement: 614 case CodeCompletionContext::CCC_Expression: 615 case CodeCompletionContext::CCC_ObjCMessageReceiver: 616 case CodeCompletionContext::CCC_EnumTag: 617 case CodeCompletionContext::CCC_UnionTag: 618 case CodeCompletionContext::CCC_ClassOrStructTag: 619 case CodeCompletionContext::CCC_ObjCProtocolName: 620 case CodeCompletionContext::CCC_Namespace: 621 case CodeCompletionContext::CCC_Type: 622 case CodeCompletionContext::CCC_ParenthesizedExpression: 623 case CodeCompletionContext::CCC_ObjCInterfaceName: 624 case CodeCompletionContext::CCC_ObjCCategoryName: 625 case CodeCompletionContext::CCC_Symbol: 626 case CodeCompletionContext::CCC_SymbolOrNewName: 627 return true; 628 case CodeCompletionContext::CCC_OtherWithMacros: 629 case CodeCompletionContext::CCC_DotMemberAccess: 630 case CodeCompletionContext::CCC_ArrowMemberAccess: 631 case CodeCompletionContext::CCC_ObjCPropertyAccess: 632 case CodeCompletionContext::CCC_MacroName: 633 case CodeCompletionContext::CCC_MacroNameUse: 634 case CodeCompletionContext::CCC_PreprocessorExpression: 635 case CodeCompletionContext::CCC_PreprocessorDirective: 636 case CodeCompletionContext::CCC_SelectorName: 637 case CodeCompletionContext::CCC_TypeQualifiers: 638 case CodeCompletionContext::CCC_ObjCInstanceMessage: 639 case CodeCompletionContext::CCC_ObjCClassMessage: 640 case CodeCompletionContext::CCC_IncludedFile: 641 // FIXME: Provide identifier based completions for the following contexts: 642 case CodeCompletionContext::CCC_Other: // Be conservative. 643 case CodeCompletionContext::CCC_NaturalLanguage: 644 case CodeCompletionContext::CCC_Recovery: 645 case CodeCompletionContext::CCC_NewName: 646 return false; 647 } 648 llvm_unreachable("unknown code completion context"); 649 } 650 651 static bool isInjectedClass(const NamedDecl &D) { 652 if (auto *R = dyn_cast_or_null<RecordDecl>(&D)) 653 if (R->isInjectedClassName()) 654 return true; 655 return false; 656 } 657 658 // Some member calls are blacklisted because they're so rarely useful. 659 static bool isBlacklistedMember(const NamedDecl &D) { 660 // Destructor completion is rarely useful, and works inconsistently. 661 // (s.^ completes ~string, but s.~st^ is an error). 662 if (D.getKind() == Decl::CXXDestructor) 663 return true; 664 // Injected name may be useful for A::foo(), but who writes A::A::foo()? 665 if (isInjectedClass(D)) 666 return true; 667 // Explicit calls to operators are also rare. 668 auto NameKind = D.getDeclName().getNameKind(); 669 if (NameKind == DeclarationName::CXXOperatorName || 670 NameKind == DeclarationName::CXXLiteralOperatorName || 671 NameKind == DeclarationName::CXXConversionFunctionName) 672 return true; 673 return false; 674 } 675 676 // The CompletionRecorder captures Sema code-complete output, including context. 677 // It filters out ignored results (but doesn't apply fuzzy-filtering yet). 678 // It doesn't do scoring or conversion to CompletionItem yet, as we want to 679 // merge with index results first. 680 // Generally the fields and methods of this object should only be used from 681 // within the callback. 682 struct CompletionRecorder : public CodeCompleteConsumer { 683 CompletionRecorder(const CodeCompleteOptions &Opts, 684 llvm::unique_function<void()> ResultsCallback) 685 : CodeCompleteConsumer(Opts.getClangCompleteOpts()), 686 CCContext(CodeCompletionContext::CCC_Other), Opts(Opts), 687 CCAllocator(std::make_shared<GlobalCodeCompletionAllocator>()), 688 CCTUInfo(CCAllocator), ResultsCallback(std::move(ResultsCallback)) { 689 assert(this->ResultsCallback); 690 } 691 692 std::vector<CodeCompletionResult> Results; 693 CodeCompletionContext CCContext; 694 Sema *CCSema = nullptr; // Sema that created the results. 695 // FIXME: Sema is scary. Can we store ASTContext and Preprocessor, instead? 696 697 void ProcessCodeCompleteResults(class Sema &S, CodeCompletionContext Context, 698 CodeCompletionResult *InResults, 699 unsigned NumResults) override final { 700 // Results from recovery mode are generally useless, and the callback after 701 // recovery (if any) is usually more interesting. To make sure we handle the 702 // future callback from sema, we just ignore all callbacks in recovery mode, 703 // as taking only results from recovery mode results in poor completion 704 // results. 705 // FIXME: in case there is no future sema completion callback after the 706 // recovery mode, we might still want to provide some results (e.g. trivial 707 // identifier-based completion). 708 if (Context.getKind() == CodeCompletionContext::CCC_Recovery) { 709 log("Code complete: Ignoring sema code complete callback with Recovery " 710 "context."); 711 return; 712 } 713 // If a callback is called without any sema result and the context does not 714 // support index-based completion, we simply skip it to give way to 715 // potential future callbacks with results. 716 if (NumResults == 0 && !contextAllowsIndex(Context.getKind())) 717 return; 718 if (CCSema) { 719 log("Multiple code complete callbacks (parser backtracked?). " 720 "Dropping results from context {0}, keeping results from {1}.", 721 getCompletionKindString(Context.getKind()), 722 getCompletionKindString(this->CCContext.getKind())); 723 return; 724 } 725 // Record the completion context. 726 CCSema = &S; 727 CCContext = Context; 728 729 // Retain the results we might want. 730 for (unsigned I = 0; I < NumResults; ++I) { 731 auto &Result = InResults[I]; 732 // Class members that are shadowed by subclasses are usually noise. 733 if (Result.Hidden && Result.Declaration && 734 Result.Declaration->isCXXClassMember()) 735 continue; 736 if (!Opts.IncludeIneligibleResults && 737 (Result.Availability == CXAvailability_NotAvailable || 738 Result.Availability == CXAvailability_NotAccessible)) 739 continue; 740 if (Result.Declaration && 741 !Context.getBaseType().isNull() // is this a member-access context? 742 && isBlacklistedMember(*Result.Declaration)) 743 continue; 744 // Skip injected class name when no class scope is not explicitly set. 745 // E.g. show injected A::A in `using A::A^` but not in "A^". 746 if (Result.Declaration && !Context.getCXXScopeSpecifier().hasValue() && 747 isInjectedClass(*Result.Declaration)) 748 continue; 749 // We choose to never append '::' to completion results in clangd. 750 Result.StartsNestedNameSpecifier = false; 751 Results.push_back(Result); 752 } 753 ResultsCallback(); 754 } 755 756 CodeCompletionAllocator &getAllocator() override { return *CCAllocator; } 757 CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; } 758 759 // Returns the filtering/sorting name for Result, which must be from Results. 760 // Returned string is owned by this recorder (or the AST). 761 llvm::StringRef getName(const CodeCompletionResult &Result) { 762 switch (Result.Kind) { 763 case CodeCompletionResult::RK_Declaration: 764 if (auto *ID = Result.Declaration->getIdentifier()) 765 return ID->getName(); 766 break; 767 case CodeCompletionResult::RK_Keyword: 768 return Result.Keyword; 769 case CodeCompletionResult::RK_Macro: 770 return Result.Macro->getName(); 771 case CodeCompletionResult::RK_Pattern: 772 return Result.Pattern->getTypedText(); 773 } 774 auto *CCS = codeCompletionString(Result); 775 return CCS->getTypedText(); 776 } 777 778 // Build a CodeCompletion string for R, which must be from Results. 779 // The CCS will be owned by this recorder. 780 CodeCompletionString *codeCompletionString(const CodeCompletionResult &R) { 781 // CodeCompletionResult doesn't seem to be const-correct. We own it, anyway. 782 return const_cast<CodeCompletionResult &>(R).CreateCodeCompletionString( 783 *CCSema, CCContext, *CCAllocator, CCTUInfo, 784 /*IncludeBriefComments=*/false); 785 } 786 787 private: 788 CodeCompleteOptions Opts; 789 std::shared_ptr<GlobalCodeCompletionAllocator> CCAllocator; 790 CodeCompletionTUInfo CCTUInfo; 791 llvm::unique_function<void()> ResultsCallback; 792 }; 793 794 struct ScoredSignature { 795 // When set, requires documentation to be requested from the index with this 796 // ID. 797 llvm::Optional<SymbolID> IDForDoc; 798 SignatureInformation Signature; 799 SignatureQualitySignals Quality; 800 }; 801 802 class SignatureHelpCollector final : public CodeCompleteConsumer { 803 public: 804 SignatureHelpCollector(const clang::CodeCompleteOptions &CodeCompleteOpts, 805 const SymbolIndex *Index, SignatureHelp &SigHelp) 806 : CodeCompleteConsumer(CodeCompleteOpts), SigHelp(SigHelp), 807 Allocator(std::make_shared<clang::GlobalCodeCompletionAllocator>()), 808 CCTUInfo(Allocator), Index(Index) {} 809 810 void ProcessOverloadCandidates(Sema &S, unsigned CurrentArg, 811 OverloadCandidate *Candidates, 812 unsigned NumCandidates, 813 SourceLocation OpenParLoc) override { 814 assert(!OpenParLoc.isInvalid()); 815 SourceManager &SrcMgr = S.getSourceManager(); 816 OpenParLoc = SrcMgr.getFileLoc(OpenParLoc); 817 if (SrcMgr.isInMainFile(OpenParLoc)) 818 SigHelp.argListStart = sourceLocToPosition(SrcMgr, OpenParLoc); 819 else 820 elog("Location oustide main file in signature help: {0}", 821 OpenParLoc.printToString(SrcMgr)); 822 823 std::vector<ScoredSignature> ScoredSignatures; 824 SigHelp.signatures.reserve(NumCandidates); 825 ScoredSignatures.reserve(NumCandidates); 826 // FIXME(rwols): How can we determine the "active overload candidate"? 827 // Right now the overloaded candidates seem to be provided in a "best fit" 828 // order, so I'm not too worried about this. 829 SigHelp.activeSignature = 0; 830 assert(CurrentArg <= (unsigned)std::numeric_limits<int>::max() && 831 "too many arguments"); 832 SigHelp.activeParameter = static_cast<int>(CurrentArg); 833 for (unsigned I = 0; I < NumCandidates; ++I) { 834 OverloadCandidate Candidate = Candidates[I]; 835 // We want to avoid showing instantiated signatures, because they may be 836 // long in some cases (e.g. when 'T' is substituted with 'std::string', we 837 // would get 'std::basic_string<char>'). 838 if (auto *Func = Candidate.getFunction()) { 839 if (auto *Pattern = Func->getTemplateInstantiationPattern()) 840 Candidate = OverloadCandidate(Pattern); 841 } 842 843 const auto *CCS = Candidate.CreateSignatureString( 844 CurrentArg, S, *Allocator, CCTUInfo, true); 845 assert(CCS && "Expected the CodeCompletionString to be non-null"); 846 ScoredSignatures.push_back(processOverloadCandidate( 847 Candidate, *CCS, 848 Candidate.getFunction() 849 ? getDeclComment(S.getASTContext(), *Candidate.getFunction()) 850 : "")); 851 } 852 853 // Sema does not load the docs from the preamble, so we need to fetch extra 854 // docs from the index instead. 855 llvm::DenseMap<SymbolID, std::string> FetchedDocs; 856 if (Index) { 857 LookupRequest IndexRequest; 858 for (const auto &S : ScoredSignatures) { 859 if (!S.IDForDoc) 860 continue; 861 IndexRequest.IDs.insert(*S.IDForDoc); 862 } 863 Index->lookup(IndexRequest, [&](const Symbol &S) { 864 if (!S.Documentation.empty()) 865 FetchedDocs[S.ID] = std::string(S.Documentation); 866 }); 867 log("SigHelp: requested docs for {0} symbols from the index, got {1} " 868 "symbols with non-empty docs in the response", 869 IndexRequest.IDs.size(), FetchedDocs.size()); 870 } 871 872 llvm::sort(ScoredSignatures, [](const ScoredSignature &L, 873 const ScoredSignature &R) { 874 // Ordering follows: 875 // - Less number of parameters is better. 876 // - Function is better than FunctionType which is better than 877 // Function Template. 878 // - High score is better. 879 // - Shorter signature is better. 880 // - Alphabetically smaller is better. 881 if (L.Quality.NumberOfParameters != R.Quality.NumberOfParameters) 882 return L.Quality.NumberOfParameters < R.Quality.NumberOfParameters; 883 if (L.Quality.NumberOfOptionalParameters != 884 R.Quality.NumberOfOptionalParameters) 885 return L.Quality.NumberOfOptionalParameters < 886 R.Quality.NumberOfOptionalParameters; 887 if (L.Quality.Kind != R.Quality.Kind) { 888 using OC = CodeCompleteConsumer::OverloadCandidate; 889 switch (L.Quality.Kind) { 890 case OC::CK_Function: 891 return true; 892 case OC::CK_FunctionType: 893 return R.Quality.Kind != OC::CK_Function; 894 case OC::CK_FunctionTemplate: 895 return false; 896 } 897 llvm_unreachable("Unknown overload candidate type."); 898 } 899 if (L.Signature.label.size() != R.Signature.label.size()) 900 return L.Signature.label.size() < R.Signature.label.size(); 901 return L.Signature.label < R.Signature.label; 902 }); 903 904 for (auto &SS : ScoredSignatures) { 905 auto IndexDocIt = 906 SS.IDForDoc ? FetchedDocs.find(*SS.IDForDoc) : FetchedDocs.end(); 907 if (IndexDocIt != FetchedDocs.end()) 908 SS.Signature.documentation = IndexDocIt->second; 909 910 SigHelp.signatures.push_back(std::move(SS.Signature)); 911 } 912 } 913 914 GlobalCodeCompletionAllocator &getAllocator() override { return *Allocator; } 915 916 CodeCompletionTUInfo &getCodeCompletionTUInfo() override { return CCTUInfo; } 917 918 private: 919 void processParameterChunk(llvm::StringRef ChunkText, 920 SignatureInformation &Signature) const { 921 // (!) this is O(n), should still be fast compared to building ASTs. 922 unsigned ParamStartOffset = lspLength(Signature.label); 923 unsigned ParamEndOffset = ParamStartOffset + lspLength(ChunkText); 924 // A piece of text that describes the parameter that corresponds to 925 // the code-completion location within a function call, message send, 926 // macro invocation, etc. 927 Signature.label += ChunkText; 928 ParameterInformation Info; 929 Info.labelOffsets.emplace(ParamStartOffset, ParamEndOffset); 930 // FIXME: only set 'labelOffsets' when all clients migrate out of it. 931 Info.labelString = std::string(ChunkText); 932 933 Signature.parameters.push_back(std::move(Info)); 934 } 935 936 void processOptionalChunk(const CodeCompletionString &CCS, 937 SignatureInformation &Signature, 938 SignatureQualitySignals &Signal) const { 939 for (const auto &Chunk : CCS) { 940 switch (Chunk.Kind) { 941 case CodeCompletionString::CK_Optional: 942 assert(Chunk.Optional && 943 "Expected the optional code completion string to be non-null."); 944 processOptionalChunk(*Chunk.Optional, Signature, Signal); 945 break; 946 case CodeCompletionString::CK_VerticalSpace: 947 break; 948 case CodeCompletionString::CK_CurrentParameter: 949 case CodeCompletionString::CK_Placeholder: 950 processParameterChunk(Chunk.Text, Signature); 951 Signal.NumberOfOptionalParameters++; 952 break; 953 default: 954 Signature.label += Chunk.Text; 955 break; 956 } 957 } 958 } 959 960 // FIXME(ioeric): consider moving CodeCompletionString logic here to 961 // CompletionString.h. 962 ScoredSignature processOverloadCandidate(const OverloadCandidate &Candidate, 963 const CodeCompletionString &CCS, 964 llvm::StringRef DocComment) const { 965 SignatureInformation Signature; 966 SignatureQualitySignals Signal; 967 const char *ReturnType = nullptr; 968 969 Signature.documentation = formatDocumentation(CCS, DocComment); 970 Signal.Kind = Candidate.getKind(); 971 972 for (const auto &Chunk : CCS) { 973 switch (Chunk.Kind) { 974 case CodeCompletionString::CK_ResultType: 975 // A piece of text that describes the type of an entity or, 976 // for functions and methods, the return type. 977 assert(!ReturnType && "Unexpected CK_ResultType"); 978 ReturnType = Chunk.Text; 979 break; 980 case CodeCompletionString::CK_CurrentParameter: 981 case CodeCompletionString::CK_Placeholder: 982 processParameterChunk(Chunk.Text, Signature); 983 Signal.NumberOfParameters++; 984 break; 985 case CodeCompletionString::CK_Optional: { 986 // The rest of the parameters are defaulted/optional. 987 assert(Chunk.Optional && 988 "Expected the optional code completion string to be non-null."); 989 processOptionalChunk(*Chunk.Optional, Signature, Signal); 990 break; 991 } 992 case CodeCompletionString::CK_VerticalSpace: 993 break; 994 default: 995 Signature.label += Chunk.Text; 996 break; 997 } 998 } 999 if (ReturnType) { 1000 Signature.label += " -> "; 1001 Signature.label += ReturnType; 1002 } 1003 dlog("Signal for {0}: {1}", Signature, Signal); 1004 ScoredSignature Result; 1005 Result.Signature = std::move(Signature); 1006 Result.Quality = Signal; 1007 const FunctionDecl *Func = Candidate.getFunction(); 1008 if (Func && Result.Signature.documentation.empty()) { 1009 // Computing USR caches linkage, which may change after code completion. 1010 if (!hasUnstableLinkage(Func)) 1011 Result.IDForDoc = clangd::getSymbolID(Func); 1012 } 1013 return Result; 1014 } 1015 1016 SignatureHelp &SigHelp; 1017 std::shared_ptr<clang::GlobalCodeCompletionAllocator> Allocator; 1018 CodeCompletionTUInfo CCTUInfo; 1019 const SymbolIndex *Index; 1020 }; // SignatureHelpCollector 1021 1022 struct SemaCompleteInput { 1023 PathRef FileName; 1024 const tooling::CompileCommand &Command; 1025 const PreambleData *Preamble; 1026 llvm::StringRef Contents; 1027 size_t Offset; 1028 llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS; 1029 }; 1030 1031 void loadMainFilePreambleMacros(const Preprocessor &PP, 1032 const PreambleData &Preamble) { 1033 // The ExternalPreprocessorSource has our macros, if we know where to look. 1034 // We can read all the macros using PreambleMacros->ReadDefinedMacros(), 1035 // but this includes transitively included files, so may deserialize a lot. 1036 ExternalPreprocessorSource *PreambleMacros = PP.getExternalSource(); 1037 // As we have the names of the macros, we can look up their IdentifierInfo 1038 // and then use this to load just the macros we want. 1039 IdentifierInfoLookup *PreambleIdentifiers = 1040 PP.getIdentifierTable().getExternalIdentifierLookup(); 1041 if (!PreambleIdentifiers || !PreambleMacros) 1042 return; 1043 for (const auto &MacroName : Preamble.Macros.Names) 1044 if (auto *II = PreambleIdentifiers->get(MacroName.getKey())) 1045 if (II->isOutOfDate()) 1046 PreambleMacros->updateOutOfDateIdentifier(*II); 1047 } 1048 1049 // Invokes Sema code completion on a file. 1050 // If \p Includes is set, it will be updated based on the compiler invocation. 1051 bool semaCodeComplete(std::unique_ptr<CodeCompleteConsumer> Consumer, 1052 const clang::CodeCompleteOptions &Options, 1053 const SemaCompleteInput &Input, 1054 IncludeStructure *Includes = nullptr) { 1055 trace::Span Tracer("Sema completion"); 1056 llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS = Input.VFS; 1057 if (Input.Preamble && Input.Preamble->StatCache) 1058 VFS = Input.Preamble->StatCache->getConsumingFS(std::move(VFS)); 1059 ParseInputs ParseInput; 1060 ParseInput.CompileCommand = Input.Command; 1061 ParseInput.FS = VFS; 1062 ParseInput.Contents = std::string(Input.Contents); 1063 ParseInput.Opts = ParseOptions(); 1064 1065 IgnoreDiagnostics IgnoreDiags; 1066 auto CI = buildCompilerInvocation(ParseInput, IgnoreDiags); 1067 if (!CI) { 1068 elog("Couldn't create CompilerInvocation"); 1069 return false; 1070 } 1071 auto &FrontendOpts = CI->getFrontendOpts(); 1072 FrontendOpts.SkipFunctionBodies = true; 1073 // Disable typo correction in Sema. 1074 CI->getLangOpts()->SpellChecking = false; 1075 // Setup code completion. 1076 FrontendOpts.CodeCompleteOpts = Options; 1077 FrontendOpts.CodeCompletionAt.FileName = std::string(Input.FileName); 1078 std::tie(FrontendOpts.CodeCompletionAt.Line, 1079 FrontendOpts.CodeCompletionAt.Column) = 1080 offsetToClangLineColumn(Input.Contents, Input.Offset); 1081 1082 std::unique_ptr<llvm::MemoryBuffer> ContentsBuffer = 1083 llvm::MemoryBuffer::getMemBufferCopy(Input.Contents, Input.FileName); 1084 // The diagnostic options must be set before creating a CompilerInstance. 1085 CI->getDiagnosticOpts().IgnoreWarnings = true; 1086 // We reuse the preamble whether it's valid or not. This is a 1087 // correctness/performance tradeoff: building without a preamble is slow, and 1088 // completion is latency-sensitive. 1089 // However, if we're completing *inside* the preamble section of the draft, 1090 // overriding the preamble will break sema completion. Fortunately we can just 1091 // skip all includes in this case; these completions are really simple. 1092 PreambleBounds PreambleRegion = 1093 ComputePreambleBounds(*CI->getLangOpts(), ContentsBuffer.get(), 0); 1094 bool CompletingInPreamble = PreambleRegion.Size > Input.Offset; 1095 // NOTE: we must call BeginSourceFile after prepareCompilerInstance. Otherwise 1096 // the remapped buffers do not get freed. 1097 auto Clang = prepareCompilerInstance( 1098 std::move(CI), 1099 (Input.Preamble && !CompletingInPreamble) ? &Input.Preamble->Preamble 1100 : nullptr, 1101 std::move(ContentsBuffer), std::move(VFS), IgnoreDiags); 1102 Clang->getPreprocessorOpts().SingleFileParseMode = CompletingInPreamble; 1103 Clang->setCodeCompletionConsumer(Consumer.release()); 1104 1105 SyntaxOnlyAction Action; 1106 if (!Action.BeginSourceFile(*Clang, Clang->getFrontendOpts().Inputs[0])) { 1107 log("BeginSourceFile() failed when running codeComplete for {0}", 1108 Input.FileName); 1109 return false; 1110 } 1111 // Macros can be defined within the preamble region of the main file. 1112 // They don't fall nicely into our index/Sema dichotomy: 1113 // - they're not indexed for completion (they're not available across files) 1114 // - but Sema code complete won't see them: as part of the preamble, they're 1115 // deserialized only when mentioned. 1116 // Force them to be deserialized so SemaCodeComplete sees them. 1117 if (Input.Preamble) 1118 loadMainFilePreambleMacros(Clang->getPreprocessor(), *Input.Preamble); 1119 if (Includes) 1120 Clang->getPreprocessor().addPPCallbacks( 1121 collectIncludeStructureCallback(Clang->getSourceManager(), Includes)); 1122 if (llvm::Error Err = Action.Execute()) { 1123 log("Execute() failed when running codeComplete for {0}: {1}", 1124 Input.FileName, toString(std::move(Err))); 1125 return false; 1126 } 1127 Action.EndSourceFile(); 1128 1129 return true; 1130 } 1131 1132 // Should we allow index completions in the specified context? 1133 bool allowIndex(CodeCompletionContext &CC) { 1134 if (!contextAllowsIndex(CC.getKind())) 1135 return false; 1136 // We also avoid ClassName::bar (but allow namespace::bar). 1137 auto Scope = CC.getCXXScopeSpecifier(); 1138 if (!Scope) 1139 return true; 1140 NestedNameSpecifier *NameSpec = (*Scope)->getScopeRep(); 1141 if (!NameSpec) 1142 return true; 1143 // We only query the index when qualifier is a namespace. 1144 // If it's a class, we rely solely on sema completions. 1145 switch (NameSpec->getKind()) { 1146 case NestedNameSpecifier::Global: 1147 case NestedNameSpecifier::Namespace: 1148 case NestedNameSpecifier::NamespaceAlias: 1149 return true; 1150 case NestedNameSpecifier::Super: 1151 case NestedNameSpecifier::TypeSpec: 1152 case NestedNameSpecifier::TypeSpecWithTemplate: 1153 // Unresolved inside a template. 1154 case NestedNameSpecifier::Identifier: 1155 return false; 1156 } 1157 llvm_unreachable("invalid NestedNameSpecifier kind"); 1158 } 1159 1160 std::future<SymbolSlab> startAsyncFuzzyFind(const SymbolIndex &Index, 1161 const FuzzyFindRequest &Req) { 1162 return runAsync<SymbolSlab>([&Index, Req]() { 1163 trace::Span Tracer("Async fuzzyFind"); 1164 SymbolSlab::Builder Syms; 1165 Index.fuzzyFind(Req, [&Syms](const Symbol &Sym) { Syms.insert(Sym); }); 1166 return std::move(Syms).build(); 1167 }); 1168 } 1169 1170 // Creates a `FuzzyFindRequest` based on the cached index request from the 1171 // last completion, if any, and the speculated completion filter text in the 1172 // source code. 1173 FuzzyFindRequest speculativeFuzzyFindRequestForCompletion( 1174 FuzzyFindRequest CachedReq, const CompletionPrefix &HeuristicPrefix) { 1175 CachedReq.Query = std::string(HeuristicPrefix.Name); 1176 return CachedReq; 1177 } 1178 1179 // Runs Sema-based (AST) and Index-based completion, returns merged results. 1180 // 1181 // There are a few tricky considerations: 1182 // - the AST provides information needed for the index query (e.g. which 1183 // namespaces to search in). So Sema must start first. 1184 // - we only want to return the top results (Opts.Limit). 1185 // Building CompletionItems for everything else is wasteful, so we want to 1186 // preserve the "native" format until we're done with scoring. 1187 // - the data underlying Sema completion items is owned by the AST and various 1188 // other arenas, which must stay alive for us to build CompletionItems. 1189 // - we may get duplicate results from Sema and the Index, we need to merge. 1190 // 1191 // So we start Sema completion first, and do all our work in its callback. 1192 // We use the Sema context information to query the index. 1193 // Then we merge the two result sets, producing items that are Sema/Index/Both. 1194 // These items are scored, and the top N are synthesized into the LSP response. 1195 // Finally, we can clean up the data structures created by Sema completion. 1196 // 1197 // Main collaborators are: 1198 // - semaCodeComplete sets up the compiler machinery to run code completion. 1199 // - CompletionRecorder captures Sema completion results, including context. 1200 // - SymbolIndex (Opts.Index) provides index completion results as Symbols 1201 // - CompletionCandidates are the result of merging Sema and Index results. 1202 // Each candidate points to an underlying CodeCompletionResult (Sema), a 1203 // Symbol (Index), or both. It computes the result quality score. 1204 // CompletionCandidate also does conversion to CompletionItem (at the end). 1205 // - FuzzyMatcher scores how the candidate matches the partial identifier. 1206 // This score is combined with the result quality score for the final score. 1207 // - TopN determines the results with the best score. 1208 class CodeCompleteFlow { 1209 PathRef FileName; 1210 IncludeStructure Includes; // Complete once the compiler runs. 1211 SpeculativeFuzzyFind *SpecFuzzyFind; // Can be nullptr. 1212 const CodeCompleteOptions &Opts; 1213 1214 // Sema takes ownership of Recorder. Recorder is valid until Sema cleanup. 1215 CompletionRecorder *Recorder = nullptr; 1216 CodeCompletionContext::Kind CCContextKind = CodeCompletionContext::CCC_Other; 1217 bool IsUsingDeclaration = false; 1218 // Counters for logging. 1219 int NSema = 0, NIndex = 0, NSemaAndIndex = 0, NIdent = 0; 1220 bool Incomplete = false; // Would more be available with a higher limit? 1221 CompletionPrefix HeuristicPrefix; 1222 llvm::Optional<FuzzyMatcher> Filter; // Initialized once Sema runs. 1223 Range ReplacedRange; 1224 std::vector<std::string> QueryScopes; // Initialized once Sema runs. 1225 // Initialized once QueryScopes is initialized, if there are scopes. 1226 llvm::Optional<ScopeDistance> ScopeProximity; 1227 llvm::Optional<OpaqueType> PreferredType; // Initialized once Sema runs. 1228 // Whether to query symbols from any scope. Initialized once Sema runs. 1229 bool AllScopes = false; 1230 llvm::StringSet<> ContextWords; 1231 // Include-insertion and proximity scoring rely on the include structure. 1232 // This is available after Sema has run. 1233 llvm::Optional<IncludeInserter> Inserter; // Available during runWithSema. 1234 llvm::Optional<URIDistance> FileProximity; // Initialized once Sema runs. 1235 /// Speculative request based on the cached request and the filter text before 1236 /// the cursor. 1237 /// Initialized right before sema run. This is only set if `SpecFuzzyFind` is 1238 /// set and contains a cached request. 1239 llvm::Optional<FuzzyFindRequest> SpecReq; 1240 1241 public: 1242 // A CodeCompleteFlow object is only useful for calling run() exactly once. 1243 CodeCompleteFlow(PathRef FileName, const IncludeStructure &Includes, 1244 SpeculativeFuzzyFind *SpecFuzzyFind, 1245 const CodeCompleteOptions &Opts) 1246 : FileName(FileName), Includes(Includes), SpecFuzzyFind(SpecFuzzyFind), 1247 Opts(Opts) {} 1248 1249 CodeCompleteResult run(const SemaCompleteInput &SemaCCInput) && { 1250 trace::Span Tracer("CodeCompleteFlow"); 1251 HeuristicPrefix = 1252 guessCompletionPrefix(SemaCCInput.Contents, SemaCCInput.Offset); 1253 populateContextWords(SemaCCInput.Contents); 1254 if (Opts.Index && SpecFuzzyFind && SpecFuzzyFind->CachedReq.hasValue()) { 1255 assert(!SpecFuzzyFind->Result.valid()); 1256 SpecReq = speculativeFuzzyFindRequestForCompletion( 1257 *SpecFuzzyFind->CachedReq, HeuristicPrefix); 1258 SpecFuzzyFind->Result = startAsyncFuzzyFind(*Opts.Index, *SpecReq); 1259 } 1260 1261 // We run Sema code completion first. It builds an AST and calculates: 1262 // - completion results based on the AST. 1263 // - partial identifier and context. We need these for the index query. 1264 CodeCompleteResult Output; 1265 auto RecorderOwner = std::make_unique<CompletionRecorder>(Opts, [&]() { 1266 assert(Recorder && "Recorder is not set"); 1267 CCContextKind = Recorder->CCContext.getKind(); 1268 IsUsingDeclaration = Recorder->CCContext.isUsingDeclaration(); 1269 auto Style = getFormatStyleForFile( 1270 SemaCCInput.FileName, SemaCCInput.Contents, SemaCCInput.VFS.get()); 1271 // If preprocessor was run, inclusions from preprocessor callback should 1272 // already be added to Includes. 1273 Inserter.emplace( 1274 SemaCCInput.FileName, SemaCCInput.Contents, Style, 1275 SemaCCInput.Command.Directory, 1276 &Recorder->CCSema->getPreprocessor().getHeaderSearchInfo()); 1277 for (const auto &Inc : Includes.MainFileIncludes) 1278 Inserter->addExisting(Inc); 1279 1280 // Most of the cost of file proximity is in initializing the FileDistance 1281 // structures based on the observed includes, once per query. Conceptually 1282 // that happens here (though the per-URI-scheme initialization is lazy). 1283 // The per-result proximity scoring is (amortized) very cheap. 1284 FileDistanceOptions ProxOpts{}; // Use defaults. 1285 const auto &SM = Recorder->CCSema->getSourceManager(); 1286 llvm::StringMap<SourceParams> ProxSources; 1287 for (auto &Entry : Includes.includeDepth( 1288 SM.getFileEntryForID(SM.getMainFileID())->getName())) { 1289 auto &Source = ProxSources[Entry.getKey()]; 1290 Source.Cost = Entry.getValue() * ProxOpts.IncludeCost; 1291 // Symbols near our transitive includes are good, but only consider 1292 // things in the same directory or below it. Otherwise there can be 1293 // many false positives. 1294 if (Entry.getValue() > 0) 1295 Source.MaxUpTraversals = 1; 1296 } 1297 FileProximity.emplace(ProxSources, ProxOpts); 1298 1299 Output = runWithSema(); 1300 Inserter.reset(); // Make sure this doesn't out-live Clang. 1301 SPAN_ATTACH(Tracer, "sema_completion_kind", 1302 getCompletionKindString(CCContextKind)); 1303 log("Code complete: sema context {0}, query scopes [{1}] (AnyScope={2}), " 1304 "expected type {3}{4}", 1305 getCompletionKindString(CCContextKind), 1306 llvm::join(QueryScopes.begin(), QueryScopes.end(), ","), AllScopes, 1307 PreferredType ? Recorder->CCContext.getPreferredType().getAsString() 1308 : "<none>", 1309 IsUsingDeclaration ? ", inside using declaration" : ""); 1310 }); 1311 1312 Recorder = RecorderOwner.get(); 1313 1314 semaCodeComplete(std::move(RecorderOwner), Opts.getClangCompleteOpts(), 1315 SemaCCInput, &Includes); 1316 logResults(Output, Tracer); 1317 return Output; 1318 } 1319 1320 void logResults(const CodeCompleteResult &Output, const trace::Span &Tracer) { 1321 SPAN_ATTACH(Tracer, "sema_results", NSema); 1322 SPAN_ATTACH(Tracer, "index_results", NIndex); 1323 SPAN_ATTACH(Tracer, "merged_results", NSemaAndIndex); 1324 SPAN_ATTACH(Tracer, "identifier_results", NIdent); 1325 SPAN_ATTACH(Tracer, "returned_results", int64_t(Output.Completions.size())); 1326 SPAN_ATTACH(Tracer, "incomplete", Output.HasMore); 1327 log("Code complete: {0} results from Sema, {1} from Index, " 1328 "{2} matched, {3} from identifiers, {4} returned{5}.", 1329 NSema, NIndex, NSemaAndIndex, NIdent, Output.Completions.size(), 1330 Output.HasMore ? " (incomplete)" : ""); 1331 assert(!Opts.Limit || Output.Completions.size() <= Opts.Limit); 1332 // We don't assert that isIncomplete means we hit a limit. 1333 // Indexes may choose to impose their own limits even if we don't have one. 1334 } 1335 1336 CodeCompleteResult 1337 runWithoutSema(llvm::StringRef Content, size_t Offset, 1338 llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS) && { 1339 trace::Span Tracer("CodeCompleteWithoutSema"); 1340 // Fill in fields normally set by runWithSema() 1341 HeuristicPrefix = guessCompletionPrefix(Content, Offset); 1342 populateContextWords(Content); 1343 CCContextKind = CodeCompletionContext::CCC_Recovery; 1344 IsUsingDeclaration = false; 1345 Filter = FuzzyMatcher(HeuristicPrefix.Name); 1346 auto Pos = offsetToPosition(Content, Offset); 1347 ReplacedRange.start = ReplacedRange.end = Pos; 1348 ReplacedRange.start.character -= HeuristicPrefix.Name.size(); 1349 1350 llvm::StringMap<SourceParams> ProxSources; 1351 ProxSources[FileName].Cost = 0; 1352 FileProximity.emplace(ProxSources); 1353 1354 auto Style = getFormatStyleForFile(FileName, Content, VFS.get()); 1355 // This will only insert verbatim headers. 1356 Inserter.emplace(FileName, Content, Style, 1357 /*BuildDir=*/"", /*HeaderSearchInfo=*/nullptr); 1358 1359 auto Identifiers = collectIdentifiers(Content, Style); 1360 std::vector<RawIdentifier> IdentifierResults; 1361 for (const auto &IDAndCount : Identifiers) { 1362 RawIdentifier ID; 1363 ID.Name = IDAndCount.first(); 1364 ID.References = IDAndCount.second; 1365 // Avoid treating typed filter as an identifier. 1366 if (ID.Name == HeuristicPrefix.Name) 1367 --ID.References; 1368 if (ID.References > 0) 1369 IdentifierResults.push_back(std::move(ID)); 1370 } 1371 1372 // Simplified version of getQueryScopes(): 1373 // - accessible scopes are determined heuristically. 1374 // - all-scopes query if no qualifier was typed (and it's allowed). 1375 SpecifiedScope Scopes; 1376 Scopes.AccessibleScopes = visibleNamespaces( 1377 Content.take_front(Offset), format::getFormattingLangOpts(Style)); 1378 for (std::string &S : Scopes.AccessibleScopes) 1379 if (!S.empty()) 1380 S.append("::"); // visibleNamespaces doesn't include trailing ::. 1381 if (HeuristicPrefix.Qualifier.empty()) 1382 AllScopes = Opts.AllScopes; 1383 else if (HeuristicPrefix.Qualifier.startswith("::")) { 1384 Scopes.AccessibleScopes = {""}; 1385 Scopes.UnresolvedQualifier = 1386 std::string(HeuristicPrefix.Qualifier.drop_front(2)); 1387 } else 1388 Scopes.UnresolvedQualifier = std::string(HeuristicPrefix.Qualifier); 1389 // First scope is the (modified) enclosing scope. 1390 QueryScopes = Scopes.scopesForIndexQuery(); 1391 ScopeProximity.emplace(QueryScopes); 1392 1393 SymbolSlab IndexResults = Opts.Index ? queryIndex() : SymbolSlab(); 1394 1395 CodeCompleteResult Output = toCodeCompleteResult(mergeResults( 1396 /*SemaResults=*/{}, IndexResults, IdentifierResults)); 1397 Output.RanParser = false; 1398 logResults(Output, Tracer); 1399 return Output; 1400 } 1401 1402 private: 1403 void populateContextWords(llvm::StringRef Content) { 1404 // Take last 3 lines before the completion point. 1405 unsigned RangeEnd = HeuristicPrefix.Qualifier.begin() - Content.data(), 1406 RangeBegin = RangeEnd; 1407 for (size_t I = 0; I < 3 && RangeBegin > 0; ++I) { 1408 auto PrevNL = Content.rfind('\n', RangeBegin); 1409 if (PrevNL == StringRef::npos) { 1410 RangeBegin = 0; 1411 break; 1412 } 1413 RangeBegin = PrevNL; 1414 } 1415 1416 ContextWords = collectWords(Content.slice(RangeBegin, RangeEnd)); 1417 dlog("Completion context words: {0}", 1418 llvm::join(ContextWords.keys(), ", ")); 1419 } 1420 1421 // This is called by run() once Sema code completion is done, but before the 1422 // Sema data structures are torn down. It does all the real work. 1423 CodeCompleteResult runWithSema() { 1424 const auto &CodeCompletionRange = CharSourceRange::getCharRange( 1425 Recorder->CCSema->getPreprocessor().getCodeCompletionTokenRange()); 1426 // When we are getting completions with an empty identifier, for example 1427 // std::vector<int> asdf; 1428 // asdf.^; 1429 // Then the range will be invalid and we will be doing insertion, use 1430 // current cursor position in such cases as range. 1431 if (CodeCompletionRange.isValid()) { 1432 ReplacedRange = halfOpenToRange(Recorder->CCSema->getSourceManager(), 1433 CodeCompletionRange); 1434 } else { 1435 const auto &Pos = sourceLocToPosition( 1436 Recorder->CCSema->getSourceManager(), 1437 Recorder->CCSema->getPreprocessor().getCodeCompletionLoc()); 1438 ReplacedRange.start = ReplacedRange.end = Pos; 1439 } 1440 Filter = FuzzyMatcher( 1441 Recorder->CCSema->getPreprocessor().getCodeCompletionFilter()); 1442 std::tie(QueryScopes, AllScopes) = getQueryScopes( 1443 Recorder->CCContext, *Recorder->CCSema, HeuristicPrefix, Opts); 1444 if (!QueryScopes.empty()) 1445 ScopeProximity.emplace(QueryScopes); 1446 PreferredType = 1447 OpaqueType::fromType(Recorder->CCSema->getASTContext(), 1448 Recorder->CCContext.getPreferredType()); 1449 // Sema provides the needed context to query the index. 1450 // FIXME: in addition to querying for extra/overlapping symbols, we should 1451 // explicitly request symbols corresponding to Sema results. 1452 // We can use their signals even if the index can't suggest them. 1453 // We must copy index results to preserve them, but there are at most Limit. 1454 auto IndexResults = (Opts.Index && allowIndex(Recorder->CCContext)) 1455 ? queryIndex() 1456 : SymbolSlab(); 1457 trace::Span Tracer("Populate CodeCompleteResult"); 1458 // Merge Sema and Index results, score them, and pick the winners. 1459 auto Top = 1460 mergeResults(Recorder->Results, IndexResults, /*Identifiers*/ {}); 1461 return toCodeCompleteResult(Top); 1462 } 1463 1464 CodeCompleteResult 1465 toCodeCompleteResult(const std::vector<ScoredBundle> &Scored) { 1466 CodeCompleteResult Output; 1467 1468 // Convert the results to final form, assembling the expensive strings. 1469 for (auto &C : Scored) { 1470 Output.Completions.push_back(toCodeCompletion(C.first)); 1471 Output.Completions.back().Score = C.second; 1472 Output.Completions.back().CompletionTokenRange = ReplacedRange; 1473 } 1474 Output.HasMore = Incomplete; 1475 Output.Context = CCContextKind; 1476 Output.CompletionRange = ReplacedRange; 1477 return Output; 1478 } 1479 1480 SymbolSlab queryIndex() { 1481 trace::Span Tracer("Query index"); 1482 SPAN_ATTACH(Tracer, "limit", int64_t(Opts.Limit)); 1483 1484 // Build the query. 1485 FuzzyFindRequest Req; 1486 if (Opts.Limit) 1487 Req.Limit = Opts.Limit; 1488 Req.Query = std::string(Filter->pattern()); 1489 Req.RestrictForCodeCompletion = true; 1490 Req.Scopes = QueryScopes; 1491 Req.AnyScope = AllScopes; 1492 // FIXME: we should send multiple weighted paths here. 1493 Req.ProximityPaths.push_back(std::string(FileName)); 1494 if (PreferredType) 1495 Req.PreferredTypes.push_back(std::string(PreferredType->raw())); 1496 vlog("Code complete: fuzzyFind({0:2})", toJSON(Req)); 1497 1498 if (SpecFuzzyFind) 1499 SpecFuzzyFind->NewReq = Req; 1500 if (SpecFuzzyFind && SpecFuzzyFind->Result.valid() && (*SpecReq == Req)) { 1501 vlog("Code complete: speculative fuzzy request matches the actual index " 1502 "request. Waiting for the speculative index results."); 1503 SPAN_ATTACH(Tracer, "Speculative results", true); 1504 1505 trace::Span WaitSpec("Wait speculative results"); 1506 return SpecFuzzyFind->Result.get(); 1507 } 1508 1509 SPAN_ATTACH(Tracer, "Speculative results", false); 1510 1511 // Run the query against the index. 1512 SymbolSlab::Builder ResultsBuilder; 1513 if (Opts.Index->fuzzyFind( 1514 Req, [&](const Symbol &Sym) { ResultsBuilder.insert(Sym); })) 1515 Incomplete = true; 1516 return std::move(ResultsBuilder).build(); 1517 } 1518 1519 // Merges Sema and Index results where possible, to form CompletionCandidates. 1520 // \p Identifiers is raw identifiers that can also be completion candidates. 1521 // Identifiers are not merged with results from index or sema. 1522 // Groups overloads if desired, to form CompletionCandidate::Bundles. The 1523 // bundles are scored and top results are returned, best to worst. 1524 std::vector<ScoredBundle> 1525 mergeResults(const std::vector<CodeCompletionResult> &SemaResults, 1526 const SymbolSlab &IndexResults, 1527 const std::vector<RawIdentifier> &IdentifierResults) { 1528 trace::Span Tracer("Merge and score results"); 1529 std::vector<CompletionCandidate::Bundle> Bundles; 1530 llvm::DenseMap<size_t, size_t> BundleLookup; 1531 auto AddToBundles = [&](const CodeCompletionResult *SemaResult, 1532 const Symbol *IndexResult, 1533 const RawIdentifier *IdentifierResult) { 1534 CompletionCandidate C; 1535 C.SemaResult = SemaResult; 1536 C.IndexResult = IndexResult; 1537 C.IdentifierResult = IdentifierResult; 1538 if (C.IndexResult) { 1539 C.Name = IndexResult->Name; 1540 C.RankedIncludeHeaders = getRankedIncludes(*C.IndexResult); 1541 } else if (C.SemaResult) { 1542 C.Name = Recorder->getName(*SemaResult); 1543 } else { 1544 assert(IdentifierResult); 1545 C.Name = IdentifierResult->Name; 1546 } 1547 if (auto OverloadSet = C.overloadSet(Opts)) { 1548 auto Ret = BundleLookup.try_emplace(OverloadSet, Bundles.size()); 1549 if (Ret.second) 1550 Bundles.emplace_back(); 1551 Bundles[Ret.first->second].push_back(std::move(C)); 1552 } else { 1553 Bundles.emplace_back(); 1554 Bundles.back().push_back(std::move(C)); 1555 } 1556 }; 1557 llvm::DenseSet<const Symbol *> UsedIndexResults; 1558 auto CorrespondingIndexResult = 1559 [&](const CodeCompletionResult &SemaResult) -> const Symbol * { 1560 if (auto SymID = 1561 getSymbolID(SemaResult, Recorder->CCSema->getSourceManager())) { 1562 auto I = IndexResults.find(*SymID); 1563 if (I != IndexResults.end()) { 1564 UsedIndexResults.insert(&*I); 1565 return &*I; 1566 } 1567 } 1568 return nullptr; 1569 }; 1570 // Emit all Sema results, merging them with Index results if possible. 1571 for (auto &SemaResult : SemaResults) 1572 AddToBundles(&SemaResult, CorrespondingIndexResult(SemaResult), nullptr); 1573 // Now emit any Index-only results. 1574 for (const auto &IndexResult : IndexResults) { 1575 if (UsedIndexResults.count(&IndexResult)) 1576 continue; 1577 AddToBundles(/*SemaResult=*/nullptr, &IndexResult, nullptr); 1578 } 1579 // Emit identifier results. 1580 for (const auto &Ident : IdentifierResults) 1581 AddToBundles(/*SemaResult=*/nullptr, /*IndexResult=*/nullptr, &Ident); 1582 // We only keep the best N results at any time, in "native" format. 1583 TopN<ScoredBundle, ScoredBundleGreater> Top( 1584 Opts.Limit == 0 ? std::numeric_limits<size_t>::max() : Opts.Limit); 1585 for (auto &Bundle : Bundles) 1586 addCandidate(Top, std::move(Bundle)); 1587 return std::move(Top).items(); 1588 } 1589 1590 llvm::Optional<float> fuzzyScore(const CompletionCandidate &C) { 1591 // Macros can be very spammy, so we only support prefix completion. 1592 // We won't end up with underfull index results, as macros are sema-only. 1593 if (C.SemaResult && C.SemaResult->Kind == CodeCompletionResult::RK_Macro && 1594 !C.Name.startswith_lower(Filter->pattern())) 1595 return None; 1596 return Filter->match(C.Name); 1597 } 1598 1599 // Scores a candidate and adds it to the TopN structure. 1600 void addCandidate(TopN<ScoredBundle, ScoredBundleGreater> &Candidates, 1601 CompletionCandidate::Bundle Bundle) { 1602 SymbolQualitySignals Quality; 1603 SymbolRelevanceSignals Relevance; 1604 Relevance.Context = CCContextKind; 1605 Relevance.Name = Bundle.front().Name; 1606 Relevance.Query = SymbolRelevanceSignals::CodeComplete; 1607 Relevance.FileProximityMatch = FileProximity.getPointer(); 1608 if (ScopeProximity) 1609 Relevance.ScopeProximityMatch = ScopeProximity.getPointer(); 1610 if (PreferredType) 1611 Relevance.HadContextType = true; 1612 Relevance.ContextWords = &ContextWords; 1613 1614 auto &First = Bundle.front(); 1615 if (auto FuzzyScore = fuzzyScore(First)) 1616 Relevance.NameMatch = *FuzzyScore; 1617 else 1618 return; 1619 SymbolOrigin Origin = SymbolOrigin::Unknown; 1620 bool FromIndex = false; 1621 for (const auto &Candidate : Bundle) { 1622 if (Candidate.IndexResult) { 1623 Quality.merge(*Candidate.IndexResult); 1624 Relevance.merge(*Candidate.IndexResult); 1625 Origin |= Candidate.IndexResult->Origin; 1626 FromIndex = true; 1627 if (!Candidate.IndexResult->Type.empty()) 1628 Relevance.HadSymbolType |= true; 1629 if (PreferredType && 1630 PreferredType->raw() == Candidate.IndexResult->Type) { 1631 Relevance.TypeMatchesPreferred = true; 1632 } 1633 } 1634 if (Candidate.SemaResult) { 1635 Quality.merge(*Candidate.SemaResult); 1636 Relevance.merge(*Candidate.SemaResult); 1637 if (PreferredType) { 1638 if (auto CompletionType = OpaqueType::fromCompletionResult( 1639 Recorder->CCSema->getASTContext(), *Candidate.SemaResult)) { 1640 Relevance.HadSymbolType |= true; 1641 if (PreferredType == CompletionType) 1642 Relevance.TypeMatchesPreferred = true; 1643 } 1644 } 1645 Origin |= SymbolOrigin::AST; 1646 } 1647 if (Candidate.IdentifierResult) { 1648 Quality.References = Candidate.IdentifierResult->References; 1649 Relevance.Scope = SymbolRelevanceSignals::FileScope; 1650 Origin |= SymbolOrigin::Identifier; 1651 } 1652 } 1653 1654 CodeCompletion::Scores Scores; 1655 Scores.Quality = Quality.evaluate(); 1656 Scores.Relevance = Relevance.evaluate(); 1657 Scores.Total = evaluateSymbolAndRelevance(Scores.Quality, Scores.Relevance); 1658 // NameMatch is in fact a multiplier on total score, so rescoring is sound. 1659 Scores.ExcludingName = Relevance.NameMatch 1660 ? Scores.Total / Relevance.NameMatch 1661 : Scores.Quality; 1662 1663 if (Opts.RecordCCResult) 1664 Opts.RecordCCResult(toCodeCompletion(Bundle), Quality, Relevance, 1665 Scores.Total); 1666 1667 dlog("CodeComplete: {0} ({1}) = {2}\n{3}{4}\n", First.Name, 1668 llvm::to_string(Origin), Scores.Total, llvm::to_string(Quality), 1669 llvm::to_string(Relevance)); 1670 1671 NSema += bool(Origin & SymbolOrigin::AST); 1672 NIndex += FromIndex; 1673 NSemaAndIndex += bool(Origin & SymbolOrigin::AST) && FromIndex; 1674 NIdent += bool(Origin & SymbolOrigin::Identifier); 1675 if (Candidates.push({std::move(Bundle), Scores})) 1676 Incomplete = true; 1677 } 1678 1679 CodeCompletion toCodeCompletion(const CompletionCandidate::Bundle &Bundle) { 1680 llvm::Optional<CodeCompletionBuilder> Builder; 1681 for (const auto &Item : Bundle) { 1682 CodeCompletionString *SemaCCS = 1683 Item.SemaResult ? Recorder->codeCompletionString(*Item.SemaResult) 1684 : nullptr; 1685 if (!Builder) 1686 Builder.emplace(Recorder ? &Recorder->CCSema->getASTContext() : nullptr, 1687 Item, SemaCCS, QueryScopes, *Inserter, FileName, 1688 CCContextKind, Opts, 1689 /*GenerateSnippets=*/!IsUsingDeclaration); 1690 else 1691 Builder->add(Item, SemaCCS); 1692 } 1693 return Builder->build(); 1694 } 1695 }; 1696 1697 } // namespace 1698 1699 clang::CodeCompleteOptions CodeCompleteOptions::getClangCompleteOpts() const { 1700 clang::CodeCompleteOptions Result; 1701 Result.IncludeCodePatterns = EnableSnippets && IncludeCodePatterns; 1702 Result.IncludeMacros = IncludeMacros; 1703 Result.IncludeGlobals = true; 1704 // We choose to include full comments and not do doxygen parsing in 1705 // completion. 1706 // FIXME: ideally, we should support doxygen in some form, e.g. do markdown 1707 // formatting of the comments. 1708 Result.IncludeBriefComments = false; 1709 1710 // When an is used, Sema is responsible for completing the main file, 1711 // the index can provide results from the preamble. 1712 // Tell Sema not to deserialize the preamble to look for results. 1713 Result.LoadExternal = !Index; 1714 Result.IncludeFixIts = IncludeFixIts; 1715 1716 return Result; 1717 } 1718 1719 CompletionPrefix guessCompletionPrefix(llvm::StringRef Content, 1720 unsigned Offset) { 1721 assert(Offset <= Content.size()); 1722 StringRef Rest = Content.take_front(Offset); 1723 CompletionPrefix Result; 1724 1725 // Consume the unqualified name. We only handle ASCII characters. 1726 // isIdentifierBody will let us match "0invalid", but we don't mind. 1727 while (!Rest.empty() && isIdentifierBody(Rest.back())) 1728 Rest = Rest.drop_back(); 1729 Result.Name = Content.slice(Rest.size(), Offset); 1730 1731 // Consume qualifiers. 1732 while (Rest.consume_back("::") && !Rest.endswith(":")) // reject :::: 1733 while (!Rest.empty() && isIdentifierBody(Rest.back())) 1734 Rest = Rest.drop_back(); 1735 Result.Qualifier = 1736 Content.slice(Rest.size(), Result.Name.begin() - Content.begin()); 1737 1738 return Result; 1739 } 1740 1741 CodeCompleteResult 1742 codeComplete(PathRef FileName, const tooling::CompileCommand &Command, 1743 const PreambleData *Preamble, llvm::StringRef Contents, 1744 Position Pos, llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS, 1745 CodeCompleteOptions Opts, SpeculativeFuzzyFind *SpecFuzzyFind) { 1746 auto Offset = positionToOffset(Contents, Pos); 1747 if (!Offset) { 1748 elog("Code completion position was invalid {0}", Offset.takeError()); 1749 return CodeCompleteResult(); 1750 } 1751 auto Flow = CodeCompleteFlow( 1752 FileName, Preamble ? Preamble->Includes : IncludeStructure(), 1753 SpecFuzzyFind, Opts); 1754 return (!Preamble || Opts.RunParser == CodeCompleteOptions::NeverParse) 1755 ? std::move(Flow).runWithoutSema(Contents, *Offset, VFS) 1756 : std::move(Flow).run( 1757 {FileName, Command, Preamble, Contents, *Offset, VFS}); 1758 } 1759 1760 SignatureHelp signatureHelp(PathRef FileName, 1761 const tooling::CompileCommand &Command, 1762 const PreambleData *Preamble, 1763 llvm::StringRef Contents, Position Pos, 1764 llvm::IntrusiveRefCntPtr<llvm::vfs::FileSystem> VFS, 1765 const SymbolIndex *Index) { 1766 auto Offset = positionToOffset(Contents, Pos); 1767 if (!Offset) { 1768 elog("Code completion position was invalid {0}", Offset.takeError()); 1769 return SignatureHelp(); 1770 } 1771 SignatureHelp Result; 1772 clang::CodeCompleteOptions Options; 1773 Options.IncludeGlobals = false; 1774 Options.IncludeMacros = false; 1775 Options.IncludeCodePatterns = false; 1776 Options.IncludeBriefComments = false; 1777 IncludeStructure PreambleInclusions; // Unused for signatureHelp 1778 semaCodeComplete( 1779 std::make_unique<SignatureHelpCollector>(Options, Index, Result), Options, 1780 {FileName, Command, Preamble, Contents, *Offset, std::move(VFS)}); 1781 return Result; 1782 } 1783 1784 bool isIndexedForCodeCompletion(const NamedDecl &ND, ASTContext &ASTCtx) { 1785 auto InTopLevelScope = [](const NamedDecl &ND) { 1786 switch (ND.getDeclContext()->getDeclKind()) { 1787 case Decl::TranslationUnit: 1788 case Decl::Namespace: 1789 case Decl::LinkageSpec: 1790 return true; 1791 default: 1792 break; 1793 }; 1794 return false; 1795 }; 1796 // We only complete symbol's name, which is the same as the name of the 1797 // *primary* template in case of template specializations. 1798 if (isExplicitTemplateSpecialization(&ND)) 1799 return false; 1800 1801 if (InTopLevelScope(ND)) 1802 return true; 1803 1804 if (const auto *EnumDecl = dyn_cast<clang::EnumDecl>(ND.getDeclContext())) 1805 return InTopLevelScope(*EnumDecl) && !EnumDecl->isScoped(); 1806 1807 return false; 1808 } 1809 1810 CompletionItem CodeCompletion::render(const CodeCompleteOptions &Opts) const { 1811 CompletionItem LSP; 1812 const auto *InsertInclude = Includes.empty() ? nullptr : &Includes[0]; 1813 LSP.label = ((InsertInclude && InsertInclude->Insertion) 1814 ? Opts.IncludeIndicator.Insert 1815 : Opts.IncludeIndicator.NoInsert) + 1816 (Opts.ShowOrigins ? "[" + llvm::to_string(Origin) + "]" : "") + 1817 RequiredQualifier + Name + Signature; 1818 1819 LSP.kind = Kind; 1820 LSP.detail = BundleSize > 1 1821 ? std::string(llvm::formatv("[{0} overloads]", BundleSize)) 1822 : ReturnType; 1823 LSP.deprecated = Deprecated; 1824 if (InsertInclude) 1825 LSP.detail += "\n" + InsertInclude->Header; 1826 LSP.documentation = Documentation; 1827 LSP.sortText = sortText(Score.Total, Name); 1828 LSP.filterText = Name; 1829 LSP.textEdit = {CompletionTokenRange, RequiredQualifier + Name}; 1830 // Merge continuous additionalTextEdits into main edit. The main motivation 1831 // behind this is to help LSP clients, it seems most of them are confused when 1832 // they are provided with additionalTextEdits that are consecutive to main 1833 // edit. 1834 // Note that we store additional text edits from back to front in a line. That 1835 // is mainly to help LSP clients again, so that changes do not effect each 1836 // other. 1837 for (const auto &FixIt : FixIts) { 1838 if (isRangeConsecutive(FixIt.range, LSP.textEdit->range)) { 1839 LSP.textEdit->newText = FixIt.newText + LSP.textEdit->newText; 1840 LSP.textEdit->range.start = FixIt.range.start; 1841 } else { 1842 LSP.additionalTextEdits.push_back(FixIt); 1843 } 1844 } 1845 if (Opts.EnableSnippets) 1846 LSP.textEdit->newText += SnippetSuffix; 1847 1848 // FIXME(kadircet): Do not even fill insertText after making sure textEdit is 1849 // compatible with most of the editors. 1850 LSP.insertText = LSP.textEdit->newText; 1851 LSP.insertTextFormat = Opts.EnableSnippets ? InsertTextFormat::Snippet 1852 : InsertTextFormat::PlainText; 1853 if (InsertInclude && InsertInclude->Insertion) 1854 LSP.additionalTextEdits.push_back(*InsertInclude->Insertion); 1855 1856 LSP.score = Score.ExcludingName; 1857 1858 return LSP; 1859 } 1860 1861 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, const CodeCompletion &C) { 1862 // For now just lean on CompletionItem. 1863 return OS << C.render(CodeCompleteOptions()); 1864 } 1865 1866 llvm::raw_ostream &operator<<(llvm::raw_ostream &OS, 1867 const CodeCompleteResult &R) { 1868 OS << "CodeCompleteResult: " << R.Completions.size() << (R.HasMore ? "+" : "") 1869 << " (" << getCompletionKindString(R.Context) << ")" 1870 << " items:\n"; 1871 for (const auto &C : R.Completions) 1872 OS << C << "\n"; 1873 return OS; 1874 } 1875 1876 } // namespace clangd 1877 } // namespace clang 1878