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