[clang] [clang][CUDA/HIP] Fix parsing of `operator<<<...>` (PR #218384)

[Yaxun Liu via cfe-commits]

================
@@ -2485,6 +2485,30 @@ bool Parser::ParseUnqualifiedIdOperator(CXXScopeSpec &SS, bool EnteringContext,
       Actions.CodeCompletion().CodeCompleteOperatorName(getCurScope());
       return true;
     }
+    case tok::lesslessless: {
+      // For CUDA, the Lexer will greedily merge all three <<< in operator<<<
+      // which, in fact, can be a valid template specialization of operator<<,
+      // and will never be a valid kernel launch expression, so split.
+      bool CachingTokens = PP.IsPreviousCachedToken(Tok);
+      // If there was a cache, we should update it when doing token split.
+      // The code below never does.
+      assert(!CachingTokens && "No cache expected");
+
+      SourceLocation TokLoc = Tok.getLocation();
+      unsigned LessLessLength = Lexer::getTokenPrefixLength(
+          TokLoc, /*CharNo=*/2, PP.getSourceManager(), getLangOpts());
+
+      SourceLocation LessLoc = PP.SplitToken(TokLoc, LessLessLength);
+      unsigned OldLength = Tok.getLength();
+
+      Tok.setKind(tok::less);
+      Tok.setLength(OldLength - LessLessLength);
+      Tok.setLocation(LessLoc);
----------------
yxsamliu wrote:

`SplitToken` returns the location for the `<<` prefix, but it is assigned to the remaining `<` here. Should `SymbolLocations` use this location and `Tok` start after the first two characters? Please also add a source-location test that checks the template `<` points to the third character.

https://github.com/llvm/llvm-project/pull/218384