[clang-tools-extra] Add support for renaming objc methods, even those with multiple selector pieces (PR #76466)

[David Goldman via cfe-commits]

================
@@ -538,11 +565,254 @@ std::optional<InvalidName> checkName(const NamedDecl &RenameDecl,
             Conflict->getLocation().printToString(ASTCtx.getSourceManager())};
     }
   }
-  if (Result)
+  if (Result) {
     InvalidNameMetric.record(1, toString(Result->K));
+    return makeError(*Result);
+  }
+  return std::nullopt;
+}
+
+bool isMatchingSelectorName(const syntax::Token &Cur, const syntax::Token &Next,
+                            const SourceManager &SM,
+                            llvm::StringRef SelectorName) {
+  if (SelectorName.empty())
+    return Cur.kind() == tok::colon;
+  return Cur.kind() == tok::identifier && Next.kind() == tok::colon &&
+         Cur.text(SM) == SelectorName &&
+         // We require the selector name and : to be contiguous to avoid
+         // potential conflicts with ternary expression.
+         //
+         // e.g. support `foo:` but not `foo :`.
+         Cur.endLocation() == Next.location();
+}
+
+bool isSelectorLike(const syntax::Token &Cur, const syntax::Token &Next) {
+  return Cur.kind() == tok::identifier && Next.kind() == tok::colon &&
+         // We require the selector name and : to be contiguous.
+         // e.g. support `foo:` but not `foo :`.
+         Cur.endLocation() == Next.location();
+}
+
+bool parseMessageExpression(llvm::ArrayRef<syntax::Token> Tokens,
+                            const SourceManager &SM, unsigned Index,
+                            unsigned Last, Selector Sel,
+                            std::vector<Range> &SelectorPieces) {
+
+  unsigned NumArgs = Sel.getNumArgs();
+  llvm::SmallVector<char, 8> Closes;
+  SelectorPieces.clear();
+  while (Index < Last) {
+    const auto &Tok = Tokens[Index];
+
+    if (Closes.empty()) {
+      auto PieceCount = SelectorPieces.size();
+      if (PieceCount < NumArgs &&
+          isMatchingSelectorName(Tok, Tokens[Index + 1], SM,
+                                 Sel.getNameForSlot(PieceCount))) {
+        // If 'foo:' instead of ':' (empty selector), we need to skip the ':'
+        // token after the name.
+        if (!Sel.getNameForSlot(PieceCount).empty()) {
+          ++Index;
+        }
+        SelectorPieces.push_back(
+            halfOpenToRange(SM, Tok.range(SM).toCharRange(SM)));
+        continue;
+      }
+      // If we've found all pieces but the current token looks like another
+      // selector piece, it means the method being renamed is a strict prefix of
+      // the selector we've found - should be skipped.
+      if (SelectorPieces.size() >= NumArgs &&
+          isSelectorLike(Tok, Tokens[Index + 1]))
+        return false;
+    }
+
+    switch (Tok.kind()) {
+    case tok::l_square:
+      Closes.push_back(']');
+      break;
+    case tok::l_paren:
+      Closes.push_back(')');
+      break;
+    case tok::l_brace:
+      Closes.push_back('}');
+      break;
+    case tok::r_square:
+      if (Closes.empty())
+        return SelectorPieces.size() == NumArgs;
+
+      if (Closes.back() != ']')
+        return false;
+      Closes.pop_back();
+      break;
+    case tok::r_paren:
+      if (Closes.empty() || Closes.back() != ')')
+        return false;
+      Closes.pop_back();
+      break;
+    case tok::r_brace:
+      if (Closes.empty() || Closes.back() != '}')
+        return false;
+      Closes.pop_back();
+      break;
+    case tok::semi:
+      // top level ; ends all statements.
+      if (Closes.empty())
+        return false;
+      break;
+    default:
+      break;
+    }
+
+    ++Index;
+  }
+  return false;
+}
+
+/// Collects all ranges of the given identifier/selector in the source code.
+///
+/// If a selector is given, this does a full lex of the given source code in
+/// order to identify all selector fragments (e.g. in method exprs/decls) since
+/// they are non-contiguous.
+std::vector<SymbolRange> collectRenameIdentifierRanges(
+    llvm::StringRef Identifier, llvm::StringRef Content,
+    const LangOptions &LangOpts, std::optional<Selector> Selector) {
+  std::vector<SymbolRange> Ranges;
+  if (!Selector) {
+    auto IdentifierRanges =
+        collectIdentifierRanges(Identifier, Content, LangOpts);
+    for (const auto &R : IdentifierRanges)
+      Ranges.emplace_back(R);
+    return Ranges;
+  }
+  // FIXME: InMemoryFileAdapter crashes unless the buffer is null terminated!
+  std::string NullTerminatedCode = Content.str();
+  SourceManagerForFile FileSM("mock_file_name.cpp", NullTerminatedCode);
+  auto &SM = FileSM.get();
+
+  // We track parens and braces to ensure that we don't accidentally try parsing
+  // a method declaration or definition which isn't at the top level or similar
+  // looking expressions (e.g. an @selector() expression).
+  unsigned ParenCount = 0;
+  unsigned BraceCount = 0;
+  unsigned NumArgs = Selector->getNumArgs();
+
+  std::vector<Range> SelectorPieces;
+  auto Tokens = syntax::tokenize(SM.getMainFileID(), SM, LangOpts);
+  unsigned Last = Tokens.size() - 1;
+  for (unsigned Index = 0; Index < Last; ++Index) {
+    const auto &Tok = Tokens[Index];
+
----------------
DavidGoldman wrote:

The main edge case I mentioned is having one selector being a prefix of another:

```
@interface Foo
- (void)foo:(id)foo bar:(id)bar;
- (void)foo:(id)foo bar:(id)bar another:(id)another;
@end
```

Renaming `foo:bar:` should not rename `foo:bar:another:` which means we need to track the termination token (`;` or `{` or method decl/defs respectively). For a method call (non top level) a `^{}` is actually a valid argument that we should skip but for a method decl/def a `{}` should terminate the method decl.

https://github.com/llvm/llvm-project/pull/76466