1 //===--- Selection.cpp ----------------------------------------------------===//
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 #include "Selection.h"
10 #include "SourceCode.h"
11 #include "support/Logger.h"
12 #include "support/Trace.h"
13 #include "clang/AST/ASTTypeTraits.h"
14 #include "clang/AST/Decl.h"
15 #include "clang/AST/DeclCXX.h"
16 #include "clang/AST/Expr.h"
17 #include "clang/AST/ExprCXX.h"
18 #include "clang/AST/PrettyPrinter.h"
19 #include "clang/AST/RecursiveASTVisitor.h"
20 #include "clang/AST/TypeLoc.h"
21 #include "clang/Basic/OperatorKinds.h"
22 #include "clang/Basic/SourceLocation.h"
23 #include "clang/Basic/SourceManager.h"
24 #include "clang/Basic/TokenKinds.h"
25 #include "clang/Lex/Lexer.h"
26 #include "clang/Tooling/Syntax/Tokens.h"
27 #include "llvm/ADT/STLExtras.h"
28 #include "llvm/ADT/StringExtras.h"
29 #include "llvm/Support/Casting.h"
30 #include "llvm/Support/raw_ostream.h"
31 #include <algorithm>
32 #include <string>
33
34 namespace clang {
35 namespace clangd {
36 namespace {
37 using Node = SelectionTree::Node;
38 using ast_type_traits::DynTypedNode;
39
40 // Measure the fraction of selections that were enabled by recovery AST.
recordMetrics(const SelectionTree & S)41 void recordMetrics(const SelectionTree &S) {
42 static constexpr trace::Metric SelectionUsedRecovery(
43 "selection_recovery", trace::Metric::Distribution);
44 static constexpr trace::Metric RecoveryType("selection_recovery_type",
45 trace::Metric::Distribution);
46 const auto *Common = S.commonAncestor();
47 for (const auto *N = Common; N; N = N->Parent) {
48 if (const auto *RE = N->ASTNode.get<RecoveryExpr>()) {
49 SelectionUsedRecovery.record(1); // used recovery ast.
50 RecoveryType.record(RE->isTypeDependent() ? 0 : 1);
51 return;
52 }
53 }
54 if (Common)
55 SelectionUsedRecovery.record(0); // unused.
56 }
57
58 // An IntervalSet maintains a set of disjoint subranges of an array.
59 //
60 // Initially, it contains the entire array.
61 // [-----------------------------------------------------------]
62 //
63 // When a range is erased(), it will typically split the array in two.
64 // Claim: [--------------------]
65 // after: [----------------] [-------------------]
66 //
67 // erase() returns the segments actually erased. Given the state above:
68 // Claim: [---------------------------------------]
69 // Out: [---------] [------]
70 // After: [-----] [-----------]
71 //
72 // It is used to track (expanded) tokens not yet associated with an AST node.
73 // On traversing an AST node, its token range is erased from the unclaimed set.
74 // The tokens actually removed are associated with that node, and hit-tested
75 // against the selection to determine whether the node is selected.
76 template <typename T> class IntervalSet {
77 public:
IntervalSet(llvm::ArrayRef<T> Range)78 IntervalSet(llvm::ArrayRef<T> Range) { UnclaimedRanges.insert(Range); }
79
80 // Removes the elements of Claim from the set, modifying or removing ranges
81 // that overlap it.
82 // Returns the continuous subranges of Claim that were actually removed.
erase(llvm::ArrayRef<T> Claim)83 llvm::SmallVector<llvm::ArrayRef<T>, 4> erase(llvm::ArrayRef<T> Claim) {
84 llvm::SmallVector<llvm::ArrayRef<T>, 4> Out;
85 if (Claim.empty())
86 return Out;
87
88 // General case:
89 // Claim: [-----------------]
90 // UnclaimedRanges: [-A-] [-B-] [-C-] [-D-] [-E-] [-F-] [-G-]
91 // Overlap: ^first ^second
92 // Ranges C and D are fully included. Ranges B and E must be trimmed.
93 auto Overlap = std::make_pair(
94 UnclaimedRanges.lower_bound({Claim.begin(), Claim.begin()}), // C
95 UnclaimedRanges.lower_bound({Claim.end(), Claim.end()})); // F
96 // Rewind to cover B.
97 if (Overlap.first != UnclaimedRanges.begin()) {
98 --Overlap.first;
99 // ...unless B isn't selected at all.
100 if (Overlap.first->end() <= Claim.begin())
101 ++Overlap.first;
102 }
103 if (Overlap.first == Overlap.second)
104 return Out;
105
106 // First, copy all overlapping ranges into the output.
107 auto OutFirst = Out.insert(Out.end(), Overlap.first, Overlap.second);
108 // If any of the overlapping ranges were sliced by the claim, split them:
109 // - restrict the returned range to the claimed part
110 // - save the unclaimed part so it can be reinserted
111 llvm::ArrayRef<T> RemainingHead, RemainingTail;
112 if (Claim.begin() > OutFirst->begin()) {
113 RemainingHead = {OutFirst->begin(), Claim.begin()};
114 *OutFirst = {Claim.begin(), OutFirst->end()};
115 }
116 if (Claim.end() < Out.back().end()) {
117 RemainingTail = {Claim.end(), Out.back().end()};
118 Out.back() = {Out.back().begin(), Claim.end()};
119 }
120
121 // Erase all the overlapping ranges (invalidating all iterators).
122 UnclaimedRanges.erase(Overlap.first, Overlap.second);
123 // Reinsert ranges that were merely trimmed.
124 if (!RemainingHead.empty())
125 UnclaimedRanges.insert(RemainingHead);
126 if (!RemainingTail.empty())
127 UnclaimedRanges.insert(RemainingTail);
128
129 return Out;
130 }
131
132 private:
133 using TokenRange = llvm::ArrayRef<T>;
134 struct RangeLess {
operator ()clang::clangd::__anon017f4d770111::IntervalSet::RangeLess135 bool operator()(llvm::ArrayRef<T> L, llvm::ArrayRef<T> R) const {
136 return L.begin() < R.begin();
137 }
138 };
139
140 // Disjoint sorted unclaimed ranges of expanded tokens.
141 std::set<llvm::ArrayRef<T>, RangeLess> UnclaimedRanges;
142 };
143
144 // Sentinel value for the selectedness of a node where we've seen no tokens yet.
145 // This resolves to Unselected if no tokens are ever seen.
146 // But Unselected + Complete -> Partial, while NoTokens + Complete --> Complete.
147 // This value is never exposed publicly.
148 constexpr SelectionTree::Selection NoTokens =
149 static_cast<SelectionTree::Selection>(
150 static_cast<unsigned char>(SelectionTree::Complete + 1));
151
152 // Nodes start with NoTokens, and then use this function to aggregate the
153 // selectedness as more tokens are found.
update(SelectionTree::Selection & Result,SelectionTree::Selection New)154 void update(SelectionTree::Selection &Result, SelectionTree::Selection New) {
155 if (New == NoTokens)
156 return;
157 if (Result == NoTokens)
158 Result = New;
159 else if (Result != New)
160 // Can only be completely selected (or unselected) if all tokens are.
161 Result = SelectionTree::Partial;
162 }
163
164 // As well as comments, don't count semicolons as real tokens.
165 // They're not properly claimed as expr-statement is missing from the AST.
shouldIgnore(const syntax::Token & Tok)166 bool shouldIgnore(const syntax::Token &Tok) {
167 return Tok.kind() == tok::comment || Tok.kind() == tok::semi;
168 }
169
170 // Determine whether 'Target' is the first expansion of the macro
171 // argument whose top-level spelling location is 'SpellingLoc'.
isFirstExpansion(FileID Target,SourceLocation SpellingLoc,const SourceManager & SM)172 bool isFirstExpansion(FileID Target, SourceLocation SpellingLoc,
173 const SourceManager &SM) {
174 SourceLocation Prev = SpellingLoc;
175 while (true) {
176 // If the arg is expanded multiple times, getMacroArgExpandedLocation()
177 // returns the first expansion.
178 SourceLocation Next = SM.getMacroArgExpandedLocation(Prev);
179 // So if we reach the target, target is the first-expansion of the
180 // first-expansion ...
181 if (SM.getFileID(Next) == Target)
182 return true;
183
184 // Otherwise, if the FileID stops changing, we've reached the innermost
185 // macro expansion, and Target was on a different branch.
186 if (SM.getFileID(Next) == SM.getFileID(Prev))
187 return false;
188
189 Prev = Next;
190 }
191 return false;
192 }
193
194 // SelectionTester can determine whether a range of tokens from the PP-expanded
195 // stream (corresponding to an AST node) is considered selected.
196 //
197 // When the tokens result from macro expansions, the appropriate tokens in the
198 // main file are examined (macro invocation or args). Similarly for #includes.
199 // However, only the first expansion of a given spelled token is considered
200 // selected.
201 //
202 // It tests each token in the range (not just the endpoints) as contiguous
203 // expanded tokens may not have contiguous spellings (with macros).
204 //
205 // Non-token text, and tokens not modeled in the AST (comments, semicolons)
206 // are ignored when determining selectedness.
207 class SelectionTester {
208 public:
209 // The selection is offsets [SelBegin, SelEnd) in SelFile.
SelectionTester(const syntax::TokenBuffer & Buf,FileID SelFile,unsigned SelBegin,unsigned SelEnd,const SourceManager & SM)210 SelectionTester(const syntax::TokenBuffer &Buf, FileID SelFile,
211 unsigned SelBegin, unsigned SelEnd, const SourceManager &SM)
212 : SelFile(SelFile), SM(SM) {
213 // Find all tokens (partially) selected in the file.
214 auto AllSpelledTokens = Buf.spelledTokens(SelFile);
215 const syntax::Token *SelFirst =
216 llvm::partition_point(AllSpelledTokens, [&](const syntax::Token &Tok) {
217 return SM.getFileOffset(Tok.endLocation()) <= SelBegin;
218 });
219 const syntax::Token *SelLimit = std::partition_point(
220 SelFirst, AllSpelledTokens.end(), [&](const syntax::Token &Tok) {
221 return SM.getFileOffset(Tok.location()) < SelEnd;
222 });
223 // Precompute selectedness and offset for selected spelled tokens.
224 for (const syntax::Token *T = SelFirst; T < SelLimit; ++T) {
225 if (shouldIgnore(*T))
226 continue;
227 SpelledTokens.emplace_back();
228 Tok &S = SpelledTokens.back();
229 S.Offset = SM.getFileOffset(T->location());
230 if (S.Offset >= SelBegin && S.Offset + T->length() <= SelEnd)
231 S.Selected = SelectionTree::Complete;
232 else
233 S.Selected = SelectionTree::Partial;
234 }
235 }
236
237 // Test whether a consecutive range of tokens is selected.
238 // The tokens are taken from the expanded token stream.
239 SelectionTree::Selection
test(llvm::ArrayRef<syntax::Token> ExpandedTokens) const240 test(llvm::ArrayRef<syntax::Token> ExpandedTokens) const {
241 if (SpelledTokens.empty())
242 return NoTokens;
243 SelectionTree::Selection Result = NoTokens;
244 while (!ExpandedTokens.empty()) {
245 // Take consecutive tokens from the same context together for efficiency.
246 FileID FID = SM.getFileID(ExpandedTokens.front().location());
247 auto Batch = ExpandedTokens.take_while([&](const syntax::Token &T) {
248 return SM.getFileID(T.location()) == FID;
249 });
250 assert(!Batch.empty());
251 ExpandedTokens = ExpandedTokens.drop_front(Batch.size());
252
253 update(Result, testChunk(FID, Batch));
254 }
255 return Result;
256 }
257
258 // Cheap check whether any of the tokens in R might be selected.
259 // If it returns false, test() will return NoTokens or Unselected.
260 // If it returns true, test() may return any value.
mayHit(SourceRange R) const261 bool mayHit(SourceRange R) const {
262 if (SpelledTokens.empty())
263 return false;
264 auto B = SM.getDecomposedLoc(R.getBegin());
265 auto E = SM.getDecomposedLoc(R.getEnd());
266 if (B.first == SelFile && E.first == SelFile)
267 if (E.second < SpelledTokens.front().Offset ||
268 B.second > SpelledTokens.back().Offset)
269 return false;
270 return true;
271 }
272
273 private:
274 // Hit-test a consecutive range of tokens from a single file ID.
275 SelectionTree::Selection
testChunk(FileID FID,llvm::ArrayRef<syntax::Token> Batch) const276 testChunk(FileID FID, llvm::ArrayRef<syntax::Token> Batch) const {
277 assert(!Batch.empty());
278 SourceLocation StartLoc = Batch.front().location();
279 // There are several possible categories of FileID depending on how the
280 // preprocessor was used to generate these tokens:
281 // main file, #included file, macro args, macro bodies.
282 // We need to identify the main-file tokens that represent Batch, and
283 // determine whether we want to exclusively claim them. Regular tokens
284 // represent one AST construct, but a macro invocation can represent many.
285
286 // Handle tokens written directly in the main file.
287 if (FID == SelFile) {
288 return testTokenRange(SM.getFileOffset(Batch.front().location()),
289 SM.getFileOffset(Batch.back().location()));
290 }
291
292 // Handle tokens in another file #included into the main file.
293 // Check if the #include is selected, but don't claim it exclusively.
294 if (StartLoc.isFileID()) {
295 for (SourceLocation Loc = Batch.front().location(); Loc.isValid();
296 Loc = SM.getIncludeLoc(SM.getFileID(Loc))) {
297 if (SM.getFileID(Loc) == SelFile)
298 // FIXME: use whole #include directive, not just the filename string.
299 return testToken(SM.getFileOffset(Loc));
300 }
301 return NoTokens;
302 }
303
304 assert(StartLoc.isMacroID());
305 // Handle tokens that were passed as a macro argument.
306 SourceLocation ArgStart = SM.getTopMacroCallerLoc(StartLoc);
307 if (SM.getFileID(ArgStart) == SelFile) {
308 if (isFirstExpansion(FID, ArgStart, SM)) {
309 SourceLocation ArgEnd =
310 SM.getTopMacroCallerLoc(Batch.back().location());
311 return testTokenRange(SM.getFileOffset(ArgStart),
312 SM.getFileOffset(ArgEnd));
313 } else {
314 /* fall through and treat as part of the macro body */
315 }
316 }
317
318 // Handle tokens produced by non-argument macro expansion.
319 // Check if the macro name is selected, don't claim it exclusively.
320 auto Expansion = SM.getDecomposedExpansionLoc(StartLoc);
321 if (Expansion.first == SelFile)
322 // FIXME: also check ( and ) for function-like macros?
323 return testToken(Expansion.second);
324 else
325 return NoTokens;
326 }
327
328 // Is the closed token range [Begin, End] selected?
testTokenRange(unsigned Begin,unsigned End) const329 SelectionTree::Selection testTokenRange(unsigned Begin, unsigned End) const {
330 assert(Begin <= End);
331 // Outside the selection entirely?
332 if (End < SpelledTokens.front().Offset ||
333 Begin > SpelledTokens.back().Offset)
334 return SelectionTree::Unselected;
335
336 // Compute range of tokens.
337 auto B = llvm::partition_point(
338 SpelledTokens, [&](const Tok &T) { return T.Offset < Begin; });
339 auto E = std::partition_point(
340 B, SpelledTokens.end(), [&](const Tok &T) { return T.Offset <= End; });
341
342 // Aggregate selectedness of tokens in range.
343 bool ExtendsOutsideSelection = Begin < SpelledTokens.front().Offset ||
344 End > SpelledTokens.back().Offset;
345 SelectionTree::Selection Result =
346 ExtendsOutsideSelection ? SelectionTree::Unselected : NoTokens;
347 for (auto It = B; It != E; ++It)
348 update(Result, It->Selected);
349 return Result;
350 }
351
352 // Is the token at `Offset` selected?
testToken(unsigned Offset) const353 SelectionTree::Selection testToken(unsigned Offset) const {
354 // Outside the selection entirely?
355 if (Offset < SpelledTokens.front().Offset ||
356 Offset > SpelledTokens.back().Offset)
357 return SelectionTree::Unselected;
358 // Find the token, if it exists.
359 auto It = llvm::partition_point(
360 SpelledTokens, [&](const Tok &T) { return T.Offset < Offset; });
361 if (It != SpelledTokens.end() && It->Offset == Offset)
362 return It->Selected;
363 return NoTokens;
364 }
365
366 struct Tok {
367 unsigned Offset;
368 SelectionTree::Selection Selected;
369 };
370 std::vector<Tok> SpelledTokens;
371 FileID SelFile;
372 const SourceManager &SM;
373 };
374
375 // Show the type of a node for debugging.
printNodeKind(llvm::raw_ostream & OS,const DynTypedNode & N)376 void printNodeKind(llvm::raw_ostream &OS, const DynTypedNode &N) {
377 if (const TypeLoc *TL = N.get<TypeLoc>()) {
378 // TypeLoc is a hierarchy, but has only a single ASTNodeKind.
379 // Synthesize the name from the Type subclass (except for QualifiedTypeLoc).
380 if (TL->getTypeLocClass() == TypeLoc::Qualified)
381 OS << "QualifiedTypeLoc";
382 else
383 OS << TL->getType()->getTypeClassName() << "TypeLoc";
384 } else {
385 OS << N.getNodeKind().asStringRef();
386 }
387 }
388
389 #ifndef NDEBUG
printNodeToString(const DynTypedNode & N,const PrintingPolicy & PP)390 std::string printNodeToString(const DynTypedNode &N, const PrintingPolicy &PP) {
391 std::string S;
392 llvm::raw_string_ostream OS(S);
393 printNodeKind(OS, N);
394 OS << " ";
395 return std::move(OS.str());
396 }
397 #endif
398
isImplicit(const Stmt * S)399 bool isImplicit(const Stmt *S) {
400 // Some Stmts are implicit and shouldn't be traversed, but there's no
401 // "implicit" attribute on Stmt/Expr.
402 // Unwrap implicit casts first if present (other nodes too?).
403 if (auto *ICE = llvm::dyn_cast<ImplicitCastExpr>(S))
404 S = ICE->getSubExprAsWritten();
405 // Implicit this in a MemberExpr is not filtered out by RecursiveASTVisitor.
406 // It would be nice if RAV handled this (!shouldTraverseImplicitCode()).
407 if (auto *CTI = llvm::dyn_cast<CXXThisExpr>(S))
408 if (CTI->isImplicit())
409 return true;
410 // Refs to operator() and [] are (almost?) always implicit as part of calls.
411 if (auto *DRE = llvm::dyn_cast<DeclRefExpr>(S)) {
412 if (auto *FD = llvm::dyn_cast<FunctionDecl>(DRE->getDecl())) {
413 switch (FD->getOverloadedOperator()) {
414 case OO_Call:
415 case OO_Subscript:
416 return true;
417 default:
418 break;
419 }
420 }
421 }
422 return false;
423 }
424
425 // We find the selection by visiting written nodes in the AST, looking for nodes
426 // that intersect with the selected character range.
427 //
428 // While traversing, we maintain a parent stack. As nodes pop off the stack,
429 // we decide whether to keep them or not. To be kept, they must either be
430 // selected or contain some nodes that are.
431 //
432 // For simple cases (not inside macros) we prune subtrees that don't intersect.
433 class SelectionVisitor : public RecursiveASTVisitor<SelectionVisitor> {
434 public:
435 // Runs the visitor to gather selected nodes and their ancestors.
436 // If there is any selection, the root (TUDecl) is the first node.
collect(ASTContext & AST,const syntax::TokenBuffer & Tokens,const PrintingPolicy & PP,unsigned Begin,unsigned End,FileID File)437 static std::deque<Node> collect(ASTContext &AST,
438 const syntax::TokenBuffer &Tokens,
439 const PrintingPolicy &PP, unsigned Begin,
440 unsigned End, FileID File) {
441 SelectionVisitor V(AST, Tokens, PP, Begin, End, File);
442 V.TraverseAST(AST);
443 assert(V.Stack.size() == 1 && "Unpaired push/pop?");
444 assert(V.Stack.top() == &V.Nodes.front());
445 return std::move(V.Nodes);
446 }
447
448 // We traverse all "well-behaved" nodes the same way:
449 // - push the node onto the stack
450 // - traverse its children recursively
451 // - pop it from the stack
452 // - hit testing: is intersection(node, selection) - union(children) empty?
453 // - attach it to the tree if it or any children hit the selection
454 //
455 // Two categories of nodes are not "well-behaved":
456 // - those without source range information, we don't record those
457 // - those that can't be stored in DynTypedNode.
458 // We're missing some interesting things like Attr due to the latter.
TraverseDecl(Decl * X)459 bool TraverseDecl(Decl *X) {
460 if (X && isa<TranslationUnitDecl>(X))
461 return Base::TraverseDecl(X); // Already pushed by constructor.
462 // Base::TraverseDecl will suppress children, but not this node itself.
463 if (X && X->isImplicit())
464 return true;
465 return traverseNode(X, [&] { return Base::TraverseDecl(X); });
466 }
TraverseTypeLoc(TypeLoc X)467 bool TraverseTypeLoc(TypeLoc X) {
468 return traverseNode(&X, [&] { return Base::TraverseTypeLoc(X); });
469 }
TraverseNestedNameSpecifierLoc(NestedNameSpecifierLoc X)470 bool TraverseNestedNameSpecifierLoc(NestedNameSpecifierLoc X) {
471 return traverseNode(
472 &X, [&] { return Base::TraverseNestedNameSpecifierLoc(X); });
473 }
TraverseConstructorInitializer(CXXCtorInitializer * X)474 bool TraverseConstructorInitializer(CXXCtorInitializer *X) {
475 return traverseNode(
476 X, [&] { return Base::TraverseConstructorInitializer(X); });
477 }
478 // Stmt is the same, but this form allows the data recursion optimization.
dataTraverseStmtPre(Stmt * X)479 bool dataTraverseStmtPre(Stmt *X) {
480 if (!X || isImplicit(X))
481 return false;
482 auto N = DynTypedNode::create(*X);
483 if (canSafelySkipNode(N))
484 return false;
485 push(std::move(N));
486 if (shouldSkipChildren(X)) {
487 pop();
488 return false;
489 }
490 return true;
491 }
dataTraverseStmtPost(Stmt * X)492 bool dataTraverseStmtPost(Stmt *X) {
493 pop();
494 return true;
495 }
496 // QualifiedTypeLoc is handled strangely in RecursiveASTVisitor: the derived
497 // TraverseTypeLoc is not called for the inner UnqualTypeLoc.
498 // This means we'd never see 'int' in 'const int'! Work around that here.
499 // (The reason for the behavior is to avoid traversing the nested Type twice,
500 // but we ignore TraverseType anyway).
TraverseQualifiedTypeLoc(QualifiedTypeLoc QX)501 bool TraverseQualifiedTypeLoc(QualifiedTypeLoc QX) {
502 return traverseNode<TypeLoc>(
503 &QX, [&] { return TraverseTypeLoc(QX.getUnqualifiedLoc()); });
504 }
505 // Uninteresting parts of the AST that don't have locations within them.
TraverseNestedNameSpecifier(NestedNameSpecifier *)506 bool TraverseNestedNameSpecifier(NestedNameSpecifier *) { return true; }
TraverseType(QualType)507 bool TraverseType(QualType) { return true; }
508
509 // The DeclStmt for the loop variable claims to cover the whole range
510 // inside the parens, this causes the range-init expression to not be hit.
511 // Traverse the loop VarDecl instead, which has the right source range.
TraverseCXXForRangeStmt(CXXForRangeStmt * S)512 bool TraverseCXXForRangeStmt(CXXForRangeStmt *S) {
513 return traverseNode(S, [&] {
514 return TraverseStmt(S->getInit()) && TraverseDecl(S->getLoopVariable()) &&
515 TraverseStmt(S->getRangeInit()) && TraverseStmt(S->getBody());
516 });
517 }
518 // OpaqueValueExpr blocks traversal, we must explicitly traverse it.
TraverseOpaqueValueExpr(OpaqueValueExpr * E)519 bool TraverseOpaqueValueExpr(OpaqueValueExpr *E) {
520 return traverseNode(E, [&] { return TraverseStmt(E->getSourceExpr()); });
521 }
522 // We only want to traverse the *syntactic form* to understand the selection.
TraversePseudoObjectExpr(PseudoObjectExpr * E)523 bool TraversePseudoObjectExpr(PseudoObjectExpr *E) {
524 return traverseNode(E, [&] { return TraverseStmt(E->getSyntacticForm()); });
525 }
526
527 private:
528 using Base = RecursiveASTVisitor<SelectionVisitor>;
529
SelectionVisitor(ASTContext & AST,const syntax::TokenBuffer & Tokens,const PrintingPolicy & PP,unsigned SelBegin,unsigned SelEnd,FileID SelFile)530 SelectionVisitor(ASTContext &AST, const syntax::TokenBuffer &Tokens,
531 const PrintingPolicy &PP, unsigned SelBegin, unsigned SelEnd,
532 FileID SelFile)
533 : SM(AST.getSourceManager()), LangOpts(AST.getLangOpts()),
534 #ifndef NDEBUG
535 PrintPolicy(PP),
536 #endif
537 TokenBuf(Tokens), SelChecker(Tokens, SelFile, SelBegin, SelEnd, SM),
538 UnclaimedExpandedTokens(Tokens.expandedTokens()) {
539 // Ensure we have a node for the TU decl, regardless of traversal scope.
540 Nodes.emplace_back();
541 Nodes.back().ASTNode = DynTypedNode::create(*AST.getTranslationUnitDecl());
542 Nodes.back().Parent = nullptr;
543 Nodes.back().Selected = SelectionTree::Unselected;
544 Stack.push(&Nodes.back());
545 }
546
547 // Generic case of TraverseFoo. Func should be the call to Base::TraverseFoo.
548 // Node is always a pointer so the generic code can handle any null checks.
549 template <typename T, typename Func>
traverseNode(T * Node,const Func & Body)550 bool traverseNode(T *Node, const Func &Body) {
551 if (Node == nullptr)
552 return true;
553 auto N = DynTypedNode::create(*Node);
554 if (canSafelySkipNode(N))
555 return true;
556 push(DynTypedNode::create(*Node));
557 bool Ret = Body();
558 pop();
559 return Ret;
560 }
561
562 // HIT TESTING
563 //
564 // We do rough hit testing on the way down the tree to avoid traversing
565 // subtrees that don't touch the selection (canSafelySkipNode), but
566 // fine-grained hit-testing is mostly done on the way back up (in pop()).
567 // This means children get to claim parts of the selection first, and parents
568 // are only selected if they own tokens that no child owned.
569 //
570 // Nodes *usually* nest nicely: a child's getSourceRange() lies within the
571 // parent's, and a node (transitively) owns all tokens in its range.
572 //
573 // Exception 1: child range claims tokens that should be owned by the parent.
574 // e.g. in `void foo(int);`, the FunctionTypeLoc should own
575 // `void (int)` but the parent FunctionDecl should own `foo`.
576 // To handle this case, certain nodes claim small token ranges *before*
577 // their children are traversed. (see earlySourceRange).
578 //
579 // Exception 2: siblings both claim the same node.
580 // e.g. `int x, y;` produces two sibling VarDecls.
581 // ~~~~~ x
582 // ~~~~~~~~ y
583 // Here the first ("leftmost") sibling claims the tokens it wants, and the
584 // other sibling gets what's left. So selecting "int" only includes the left
585 // VarDecl in the selection tree.
586
587 // An optimization for a common case: nodes outside macro expansions that
588 // don't intersect the selection may be recursively skipped.
canSafelySkipNode(const DynTypedNode & N)589 bool canSafelySkipNode(const DynTypedNode &N) {
590 SourceRange S = N.getSourceRange();
591 if (auto *TL = N.get<TypeLoc>()) {
592 // DeclTypeTypeLoc::getSourceRange() is incomplete, which would lead to
593 // failing
594 // to descend into the child expression.
595 // decltype(2+2);
596 // ~~~~~~~~~~~~~ <-- correct range
597 // ~~~~~~~~ <-- range reported by getSourceRange()
598 // ~~~~~~~~~~~~ <-- range with this hack(i.e, missing closing paren)
599 // FIXME: Alter DecltypeTypeLoc to contain parentheses locations and get
600 // rid of this patch.
601 if (auto DT = TL->getAs<DecltypeTypeLoc>())
602 S.setEnd(DT.getUnderlyingExpr()->getEndLoc());
603 }
604 if (!SelChecker.mayHit(S)) {
605 dlog("{1}skip: {0}", printNodeToString(N, PrintPolicy), indent());
606 dlog("{1}skipped range = {0}", S.printToString(SM), indent(1));
607 return true;
608 }
609 return false;
610 }
611
612 // There are certain nodes we want to treat as leaves in the SelectionTree,
613 // although they do have children.
shouldSkipChildren(const Stmt * X) const614 bool shouldSkipChildren(const Stmt *X) const {
615 // UserDefinedLiteral (e.g. 12_i) has two children (12 and _i).
616 // Unfortunately TokenBuffer sees 12_i as one token and can't split it.
617 // So we treat UserDefinedLiteral as a leaf node, owning the token.
618 return llvm::isa<UserDefinedLiteral>(X);
619 }
620
621 // Pushes a node onto the ancestor stack. Pairs with pop().
622 // Performs early hit detection for some nodes (on the earlySourceRange).
push(DynTypedNode Node)623 void push(DynTypedNode Node) {
624 SourceRange Early = earlySourceRange(Node);
625 dlog("{1}push: {0}", printNodeToString(Node, PrintPolicy), indent());
626 Nodes.emplace_back();
627 Nodes.back().ASTNode = std::move(Node);
628 Nodes.back().Parent = Stack.top();
629 Nodes.back().Selected = NoTokens;
630 Stack.push(&Nodes.back());
631 claimRange(Early, Nodes.back().Selected);
632 }
633
634 // Pops a node off the ancestor stack, and finalizes it. Pairs with push().
635 // Performs primary hit detection.
pop()636 void pop() {
637 Node &N = *Stack.top();
638 dlog("{1}pop: {0}", printNodeToString(N.ASTNode, PrintPolicy), indent(-1));
639 claimRange(N.ASTNode.getSourceRange(), N.Selected);
640 if (N.Selected == NoTokens)
641 N.Selected = SelectionTree::Unselected;
642 if (N.Selected || !N.Children.empty()) {
643 // Attach to the tree.
644 N.Parent->Children.push_back(&N);
645 } else {
646 // Neither N any children are selected, it doesn't belong in the tree.
647 assert(&N == &Nodes.back());
648 Nodes.pop_back();
649 }
650 Stack.pop();
651 }
652
653 // Returns the range of tokens that this node will claim directly, and
654 // is not available to the node's children.
655 // Usually empty, but sometimes children cover tokens but shouldn't own them.
earlySourceRange(const DynTypedNode & N)656 SourceRange earlySourceRange(const DynTypedNode &N) {
657 if (const Decl *D = N.get<Decl>()) {
658 // We want constructor name to be claimed by TypeLoc not the constructor
659 // itself. Similar for deduction guides, we rather want to select the
660 // underlying TypeLoc.
661 // FIXME: Unfortunately this doesn't work, even though RecursiveASTVisitor
662 // traverses the underlying TypeLoc inside DeclarationName, it is null for
663 // constructors.
664 if (isa<CXXConstructorDecl>(D) || isa<CXXDeductionGuideDecl>(D))
665 return SourceRange();
666 // This will capture Field, Function, MSProperty, NonTypeTemplateParm and
667 // VarDecls. We want the name in the declarator to be claimed by the decl
668 // and not by any children. For example:
669 // void [[foo]]();
670 // int (*[[s]])();
671 // struct X { int [[hash]] [32]; [[operator]] int();}
672 if (const auto *DD = llvm::dyn_cast<DeclaratorDecl>(D))
673 return DD->getLocation();
674 } else if (const auto *CCI = N.get<CXXCtorInitializer>()) {
675 // : [[b_]](42)
676 return CCI->getMemberLocation();
677 }
678 return SourceRange();
679 }
680
681 // Perform hit-testing of a complete Node against the selection.
682 // This runs for every node in the AST, and must be fast in common cases.
683 // This is usually called from pop(), so we can take children into account.
684 // The existing state of Result is relevant (early/late claims can interact).
claimRange(SourceRange S,SelectionTree::Selection & Result)685 void claimRange(SourceRange S, SelectionTree::Selection &Result) {
686 for (const auto &ClaimedRange :
687 UnclaimedExpandedTokens.erase(TokenBuf.expandedTokens(S)))
688 update(Result, SelChecker.test(ClaimedRange));
689
690 if (Result && Result != NoTokens)
691 dlog("{1}hit selection: {0}", S.printToString(SM), indent());
692 }
693
indent(int Offset=0)694 std::string indent(int Offset = 0) {
695 // Cast for signed arithmetic.
696 int Amount = int(Stack.size()) + Offset;
697 assert(Amount >= 0);
698 return std::string(Amount, ' ');
699 }
700
701 SourceManager &SM;
702 const LangOptions &LangOpts;
703 #ifndef NDEBUG
704 const PrintingPolicy &PrintPolicy;
705 #endif
706 const syntax::TokenBuffer &TokenBuf;
707 std::stack<Node *> Stack;
708 SelectionTester SelChecker;
709 IntervalSet<syntax::Token> UnclaimedExpandedTokens;
710 std::deque<Node> Nodes; // Stable pointers as we add more nodes.
711 };
712
713 } // namespace
714
abbreviatedString(DynTypedNode N,const PrintingPolicy & PP)715 llvm::SmallString<256> abbreviatedString(DynTypedNode N,
716 const PrintingPolicy &PP) {
717 llvm::SmallString<256> Result;
718 {
719 llvm::raw_svector_ostream OS(Result);
720 N.print(OS, PP);
721 }
722 auto Pos = Result.find('\n');
723 if (Pos != llvm::StringRef::npos) {
724 bool MoreText =
725 !llvm::all_of(llvm::StringRef(Result).drop_front(Pos), llvm::isSpace);
726 Result.resize(Pos);
727 if (MoreText)
728 Result.append(" …");
729 }
730 return Result;
731 }
732
print(llvm::raw_ostream & OS,const SelectionTree::Node & N,int Indent) const733 void SelectionTree::print(llvm::raw_ostream &OS, const SelectionTree::Node &N,
734 int Indent) const {
735 if (N.Selected)
736 OS.indent(Indent - 1) << (N.Selected == SelectionTree::Complete ? '*'
737 : '.');
738 else
739 OS.indent(Indent);
740 printNodeKind(OS, N.ASTNode);
741 OS << ' ' << abbreviatedString(N.ASTNode, PrintPolicy) << "\n";
742 for (const Node *Child : N.Children)
743 print(OS, *Child, Indent + 2);
744 }
745
kind() const746 std::string SelectionTree::Node::kind() const {
747 std::string S;
748 llvm::raw_string_ostream OS(S);
749 printNodeKind(OS, ASTNode);
750 return std::move(OS.str());
751 }
752
753 // Decide which selections emulate a "point" query in between characters.
754 // If it's ambiguous (the neighboring characters are selectable tokens), returns
755 // both possibilities in preference order.
756 // Always returns at least one range - if no tokens touched, and empty range.
757 static llvm::SmallVector<std::pair<unsigned, unsigned>, 2>
pointBounds(unsigned Offset,const syntax::TokenBuffer & Tokens)758 pointBounds(unsigned Offset, const syntax::TokenBuffer &Tokens) {
759 const auto &SM = Tokens.sourceManager();
760 SourceLocation Loc = SM.getComposedLoc(SM.getMainFileID(), Offset);
761 llvm::SmallVector<std::pair<unsigned, unsigned>, 2> Result;
762 // Prefer right token over left.
763 for (const syntax::Token &Tok :
764 llvm::reverse(spelledTokensTouching(Loc, Tokens))) {
765 if (shouldIgnore(Tok))
766 continue;
767 unsigned Offset = Tokens.sourceManager().getFileOffset(Tok.location());
768 Result.emplace_back(Offset, Offset + Tok.length());
769 }
770 if (Result.empty())
771 Result.emplace_back(Offset, Offset);
772 return Result;
773 }
774
createEach(ASTContext & AST,const syntax::TokenBuffer & Tokens,unsigned Begin,unsigned End,llvm::function_ref<bool (SelectionTree)> Func)775 bool SelectionTree::createEach(ASTContext &AST,
776 const syntax::TokenBuffer &Tokens,
777 unsigned Begin, unsigned End,
778 llvm::function_ref<bool(SelectionTree)> Func) {
779 if (Begin != End)
780 return Func(SelectionTree(AST, Tokens, Begin, End));
781 for (std::pair<unsigned, unsigned> Bounds : pointBounds(Begin, Tokens))
782 if (Func(SelectionTree(AST, Tokens, Bounds.first, Bounds.second)))
783 return true;
784 return false;
785 }
786
createRight(ASTContext & AST,const syntax::TokenBuffer & Tokens,unsigned int Begin,unsigned int End)787 SelectionTree SelectionTree::createRight(ASTContext &AST,
788 const syntax::TokenBuffer &Tokens,
789 unsigned int Begin, unsigned int End) {
790 llvm::Optional<SelectionTree> Result;
791 createEach(AST, Tokens, Begin, End, [&](SelectionTree T) {
792 Result = std::move(T);
793 return true;
794 });
795 return std::move(*Result);
796 }
797
SelectionTree(ASTContext & AST,const syntax::TokenBuffer & Tokens,unsigned Begin,unsigned End)798 SelectionTree::SelectionTree(ASTContext &AST, const syntax::TokenBuffer &Tokens,
799 unsigned Begin, unsigned End)
800 : PrintPolicy(AST.getLangOpts()) {
801 // No fundamental reason the selection needs to be in the main file,
802 // but that's all clangd has needed so far.
803 const SourceManager &SM = AST.getSourceManager();
804 FileID FID = SM.getMainFileID();
805 PrintPolicy.TerseOutput = true;
806 PrintPolicy.IncludeNewlines = false;
807
808 dlog("Computing selection for {0}",
809 SourceRange(SM.getComposedLoc(FID, Begin), SM.getComposedLoc(FID, End))
810 .printToString(SM));
811 Nodes = SelectionVisitor::collect(AST, Tokens, PrintPolicy, Begin, End, FID);
812 Root = Nodes.empty() ? nullptr : &Nodes.front();
813 recordMetrics(*this);
814 dlog("Built selection tree\n{0}", *this);
815 }
816
commonAncestor() const817 const Node *SelectionTree::commonAncestor() const {
818 const Node *Ancestor = Root;
819 while (Ancestor->Children.size() == 1 && !Ancestor->Selected)
820 Ancestor = Ancestor->Children.front();
821 // Returning nullptr here is a bit unprincipled, but it makes the API safer:
822 // the TranslationUnitDecl contains all of the preamble, so traversing it is a
823 // performance cliff. Callers can check for null and use root() if they want.
824 return Ancestor != Root ? Ancestor : nullptr;
825 }
826
getDeclContext() const827 const DeclContext &SelectionTree::Node::getDeclContext() const {
828 for (const Node *CurrentNode = this; CurrentNode != nullptr;
829 CurrentNode = CurrentNode->Parent) {
830 if (const Decl *Current = CurrentNode->ASTNode.get<Decl>()) {
831 if (CurrentNode != this)
832 if (auto *DC = dyn_cast<DeclContext>(Current))
833 return *DC;
834 return *Current->getDeclContext();
835 }
836 }
837 llvm_unreachable("A tree must always be rooted at TranslationUnitDecl.");
838 }
839
ignoreImplicit() const840 const SelectionTree::Node &SelectionTree::Node::ignoreImplicit() const {
841 if (Children.size() == 1 &&
842 Children.front()->ASTNode.getSourceRange() == ASTNode.getSourceRange())
843 return Children.front()->ignoreImplicit();
844 return *this;
845 }
846
outerImplicit() const847 const SelectionTree::Node &SelectionTree::Node::outerImplicit() const {
848 if (Parent && Parent->ASTNode.getSourceRange() == ASTNode.getSourceRange())
849 return Parent->outerImplicit();
850 return *this;
851 }
852
853 } // namespace clangd
854 } // namespace clang
855